Daily Papers

Large Language Models (LLMs) are increasingly deployed in scenarios demanding ultra-long context reasoning, such as agentic workflows and deep research understanding. However, long-context inference is constrained by the KV cache, a transient memory structure that grows linearly with sequence length and batch size, quickly dominating GPU memory usage. Existing memory reduction techniques, including eviction and quantization, often rely on static heuristics and suffer from degraded quality under tight budgets. In this paper, we propose ARKV, a lightweight and adaptive framework that dynamically allocates precision levels to cached tokens based on per-layer attention dynamics and token-level importance. During a short prefill phase, ARKV estimates the original quantization (OQ) ratio of each layer by computing statistical scores such as attention entropy, variance and kurtosis. During decoding, tokens are assigned to one of three states, Original (full precision), Quantization (low precision), or Eviction, according to a fast heavy-hitter scoring strategy. Our experiments on LLaMA3 and Qwen3 models across diverse long- and short-context tasks demonstrate that ARKV preserves ~97% of baseline accuracy on long-context benchmarks while reducing KV memory usage by 4x, with minimal throughput loss. On short-context tasks, ARKV matches full-precision baselines; on GSM8K math reasoning, it significantly outperforms uniform quantization. These results highlight the practical viability of ARKV for scalable LLM deployment, offering fine-grained, data-driven memory control without retraining or architectural modifications. The source code and artifacts can be found in: https://github.com/Large-scale-Sustainable-Computing-LSC/ARKV

We propose combining memory saving techniques with traditional U-Net architectures to increase the complexity of the models on the Brain Tumor Segmentation (BraTS) challenge. The BraTS challenge consists of a 3D segmentation of a 240x240x155x4 input image into a set of tumor classes. Because of the large volume and need for 3D convolutional layers, this task is very memory intensive. To address this, prior approaches use smaller cropped images while constraining the model's depth and width. Our 3D U-Net uses a reversible version of the mobile inverted bottleneck block defined in MobileNetV2, MnasNet and the more recent EfficientNet architectures to save activation memory during training. Using reversible layers enables the model to recompute input activations given the outputs of that layer, saving memory by eliminating the need to store activations during the forward pass. The inverted residual bottleneck block uses lightweight depthwise separable convolutions to reduce computation by decomposing convolutions into a pointwise convolution and a depthwise convolution. Further, this block inverts traditional bottleneck blocks by placing an intermediate expansion layer between the input and output linear 1x1 convolution, reducing the total number of channels. Given a fixed memory budget, with these memory saving techniques, we are able to train image volumes up to 3x larger, models with 25% more depth, or models with up to 2x the number of channels than a corresponding non-reversible network.

Streaming visual transformers like StreamVGGT achieve strong 3D perception but suffer from unbounded growth of key value (KV) memory, which limits scalability. We propose a training-free, inference-time token eviction policy that bounds memory by discarding redundant tokens while keeping the most informative ones. Our method uses significantly less memory with little to no drop in accuracy: on 7-Scenes with long sequences it reduces peak memory from 18.63 GB to 9.39 GB while accuracy and completeness drop by only 0.003. Under strict memory budgets, eviction enables denser frame sampling, which improves reconstruction accuracy compared to the baseline. Experiments across video depth estimation (Sintel, KITTI), 3D reconstruction (7-Scenes, NRGBD), and camera pose estimation (Sintel, TUM-dynamics) show that our approach closely matches StreamVGGT at a fraction of the memory and makes long-horizon streaming inference more practical.

Long-horizon agentic reasoning necessitates effectively compressing growing interaction histories into a limited context window. Most existing memory systems serialize history as text, where token-level cost is uniform and scales linearly with length, often spending scarce budget on low-value details. To this end, we introduce MemOCR, a multimodal memory agent that improves long-horizon reasoning under tight context budgets by allocating memory space with adaptive information density through visual layout. Concretely, MemOCR maintains a structured rich-text memory (e.g., headings, highlights) and renders it into an image that the agent consults for memory access, visually prioritizing crucial evidence while aggressively compressing auxiliary details. To ensure robustness across varying memory budgets, we train MemOCR with reinforcement learning under budget-aware objectives that expose the agent to diverse compression levels. Across long-context multi-hop and single-hop question-answering benchmarks, MemOCR outperforms strong text-based baselines and achieves more effective context utilization under extreme budgets.

Transformers have emerged as the backbone of large language models (LLMs). However, generation remains inefficient due to the need to store in memory a cache of key-value representations for past tokens, whose size scales linearly with the input sequence length and batch size. As a solution, we propose Dynamic Memory Compression (DMC), a method for on-line key-value cache compression at inference time. Most importantly, the model learns to apply different compression rates in different heads and layers. We retrofit pre-trained LLMs such as Llama 2 (7B, 13B and 70B) into DMC Transformers, achieving up to ~3.7x throughput increase in auto-regressive inference on a NVIDIA H100 GPU. DMC is applied via continued pre-training on a negligible percentage of the original data without adding any extra parameters. We find that DMC preserves the original downstream performance with up to 4x cache compression, outperforming up-trained grouped-query attention (GQA). GQA and DMC can be even combined to obtain compounded gains. As a result DMC fits longer contexts and larger batches within any given memory budget.

Recent advances in Multimodal Large Language Models (MLLMs) have improved image recognition and reasoning, but video-related tasks remain challenging due to memory constraints from dense frame processing. Existing Video Moment Retrieval (VMR) methodologies rely on sparse frame sampling, risking potential information loss, especially in lengthy videos. We propose SMORE (See MORE, store less), a framework that enhances memory efficiency while maintaining high information resolution. SMORE (1) uses query-guided captions to encode semantics aligned with user intent, (2) applies query-aware importance modulation to highlight relevant segments, and (3) adaptively compresses frames to preserve key content while reducing redundancy. This enables efficient video understanding without exceeding memory budgets. Experimental validation reveals that SMORE achieves state-of-the-art performance on QVHighlights, Charades-STA, and ActivityNet-Captions benchmarks.

Memory and computation remain core bottlenecks in long-horizon LLM inference due to the quadratic cost of self-attention and the ever-growing key-value (KV) cache. Existing strategies for memory-bounded inference, such as quantization, offloading, or heuristic KV eviction, either incur high orchestration costs or rely on unreliable attention-based proxies of importance. We propose TRIM-KV, a novel approach that learns each token's intrinsic importance at creation time via a lightweight retention gate. Each gate predicts a scalar retention score that decays over time, reflecting the long-term utility of the token for a specific layer and head. Tokens with low scores are evicted when the memory budget is exceeded, ensuring that the cache always contains the most critical tokens. TRIM-KV is trained efficiently through distillation from a frozen LLM combined with a capacity loss, requiring only gate fine-tuning and adding negligible inference overhead. Across mathematical reasoning (GSM8K, MATH-500, AIME24), procedural generation (LongProc), conversational long-memory benchmarks (LongMemEval), and long-context understanding (LongBench and SCBench), TRIM-KV consistently outperforms strong eviction and learnable retrieval baselines, especially in low-memory regimes. Remarkably, it even surpasses full-cache models in some settings, showing that selective retention can serve as a form of regularization, suppressing noise from uninformative tokens. Qualitative analyses further reveal that learned retention scores align with human intuition, naturally recovering heuristics such as sink tokens, sliding windows, and gist compression without explicit design. Beyond efficiency, retention scores provide insights into layer- and head-specific roles, suggesting a new path toward LLM interpretability.

Real-world applications require the classification model to adapt to new classes without forgetting old ones. Correspondingly, Class-Incremental Learning (CIL) aims to train a model with limited memory size to meet this requirement. Typical CIL methods tend to save representative exemplars from former classes to resist forgetting, while recent works find that storing models from history can substantially boost the performance. However, the stored models are not counted into the memory budget, which implicitly results in unfair comparisons. We find that when counting the model size into the total budget and comparing methods with aligned memory size, saving models do not consistently work, especially for the case with limited memory budgets. As a result, we need to holistically evaluate different CIL methods at different memory scales and simultaneously consider accuracy and memory size for measurement. On the other hand, we dive deeply into the construction of the memory buffer for memory efficiency. By analyzing the effect of different layers in the network, we find that shallow and deep layers have different characteristics in CIL. Motivated by this, we propose a simple yet effective baseline, denoted as MEMO for Memory-efficient Expandable MOdel. MEMO extends specialized layers based on the shared generalized representations, efficiently extracting diverse representations with modest cost and maintaining representative exemplars. Extensive experiments on benchmark datasets validate MEMO's competitive performance. Code is available at: https://github.com/wangkiw/ICLR23-MEMO

With the increase in the number of parameters in large language models, the process of pre-training and fine-tuning increasingly demands larger volumes of GPU memory. A significant portion of this memory is typically consumed by the optimizer state. To overcome this challenge, recent approaches such as low-rank adaptation (LoRA (Hu et al., 2021)), low-rank gradient projection (GaLore (Zhao et al., 2024)), and blockwise optimization (BAdam (Luo et al., 2024)) have been proposed. However, in all these algorithms, the effective rank of the weight updates remains low-rank, which can lead to a substantial loss of information from the gradient. This loss can be critically important, especially during the pre-training stage. In this paper, we introduce FRUGAL (Full-Rank Updates with GrAdient spLitting), a new memory-efficient optimization framework. FRUGAL leverages gradient splitting to perform low-dimensional updates using advanced algorithms (such as Adam), while updates along the remaining directions are executed via state-free methods like SGD or signSGD (Bernstein et al., 2018). Our framework can be integrated with various low-rank update selection techniques, including GaLore and BAdam. We provide theoretical convergence guarantees for our framework when using SGDM for low-dimensional updates and SGD for state-free updates. Additionally, our method consistently outperforms concurrent approaches across various fixed memory budgets, achieving state-of-the-art results in pre-training and fine-tuning tasks while balancing memory efficiency and performance metrics.

Optimizer states are a major source of memory consumption for training neural networks, limiting the maximum trainable model within given memory budget. Compressing the optimizer states from 32-bit floating points to lower bitwidth is promising to reduce the training memory footprint, while the current lowest achievable bitwidth is 8-bit. In this work, we push optimizer states bitwidth down to 4-bit through a detailed empirical analysis of first and second moments. Specifically, we find that moments have complicated outlier patterns, that current block-wise quantization cannot accurately approximate. We use a smaller block size and propose to utilize both row-wise and column-wise information for better quantization. We further identify a zero point problem of quantizing the second moment, and solve this problem with a linear quantizer that excludes the zero point. Our 4-bit optimizers are evaluated on a wide variety of benchmarks including natural language understanding, machine translation, image classification, and instruction tuning. On all the tasks our optimizers can achieve comparable accuracy with their full-precision counterparts, while enjoying better memory efficiency.

Model merging enables efficient multi-task models by combining task-specific fine-tuned checkpoints. However, storing multiple task-specific checkpoints requires significant memory, limiting scalability and restricting model merging to larger models and diverse tasks. In this paper, we propose quantizing task vectors (i.e., the difference between pre-trained and fine-tuned checkpoints) instead of quantizing fine-tuned checkpoints. We observe that task vectors exhibit a narrow weight range, enabling low precision quantization (up to 4 bit) within existing task vector merging frameworks. To further mitigate quantization errors within ultra-low bit precision (e.g., 2 bit), we introduce Residual Task Vector Quantization, which decomposes the task vector into a base vector and offset component. We allocate bits based on quantization sensitivity, ensuring precision while minimizing error within a memory budget. Experiments on image classification and dense prediction show our method maintains or improves model merging performance while using only 8% of the memory required for full-precision checkpoints.

Open-source foundation models have seen rapid adoption and development, enabling powerful general-purpose capabilities across diverse domains. However, fine-tuning large foundation models for domain-specific or personalized tasks remains prohibitively expensive for most users due to the significant memory overhead beyond that of inference. We introduce EMLoC, an Emulator-based Memory-efficient fine-tuning framework with LoRA Correction, which enables model fine-tuning within the same memory budget required for inference. EMLoC constructs a task-specific light-weight emulator using activation-aware singular value decomposition (SVD) on a small downstream calibration set. Fine-tuning then is performed on this lightweight emulator via LoRA. To tackle the misalignment between the original model and the compressed emulator, we propose a novel compensation algorithm to correct the fine-tuned LoRA module, which thus can be merged into the original model for inference. EMLoC supports flexible compression ratios and standard training pipelines, making it adaptable to a wide range of applications. Extensive experiments demonstrate that EMLoC outperforms other baselines across multiple datasets and modalities. Moreover, without quantization, EMLoC enables fine-tuning of a 38B model on a single 24GB consumer GPU-bringing efficient and practical model adaptation to individual users.

Owing to the huge success of generative artificial intelligence (AI), large language models (LLMs) have emerged as a core subclass, underpinning applications such as question answering, text generation, and code completion. While fine-tuning these models on domain-specific data can yield significant performance gains, it also poses daunting computational challenges, especially for researchers and small organizations with limited hardware resources. Although SSD offloading (i.e., ZeRO-Infinity) has emerged as a viable strategy to overcome the GPU memory barrier via leveraging both system memory (i.e., CPU DRAM) and storage space (i.e., solid-state devices, SSDs), its design primarily targets model-centric performance issues. As a result, key system-level issues, including system memory fragmentation, inefficient pinned buffer allocation, peak CPU usage spikes, and file system overhead, remain unaddressed, stifling scalability and inflating costs. Such an observation motivates this paper to introduce MemAscend, a framework that systematically tackles the underexplored system memory bottlenecks in SSD-offloaded LLM training, with a focus on resource-constrained environments. By streamlining pinned-memory allocation, eradicating fragmentation, and mitigating peak overhead, MemAscend reclaims a substantial system memory budget, enabling larger models, longer context windows, and higher batch sizes without exceeding modest hardware limits. Across diverse LLM benchmarks, MemAscend reduces peak system-memory consumption by an average of 55.7% compared with standard SSD offloading techniques, lowering the hardware barrier for fine-tuning and unlocking new possibilities for cost-effective large-scale training on limited-resource machines.

Large Language Models (LLMs) have revolutionized natural language processing, setting new standards across a wide range of applications. However, their relevant memory and computational demands make them impractical for deployment on technologically-constrained tiny devices such as wearable devices and Internet-of-Things units. To address this limitation, we introduce EmbBERT-Q, a novel tiny language model specifically designed for tiny devices with stringent memory constraints. EmbBERT-Q achieves state-of-the-art (SotA) accuracy in Natural Language Processing tasks in this scenario, with a total memory footprint (weights and activations) of just 781 kB, representing a 25x reduction in size with respect to SotA models. By combining architectural innovations with hardware-compatible 8-bit quantization, EmbBERT-Q consistently outperforms several baseline models scaled down to a 2 MB memory budget (i.e., the maximum memory typically available in tiny devices), including heavily compressed versions of BERT and MAMBA. Extensive experimental evaluations on both a selected benchmark dataset, TinyNLP, specifically curated to evaluate Tiny Language Models in NLP tasks and real-world scenarios, and the GLUE benchmark, demonstrate EmbBERT-Q ability to deliver competitive accuracy with respect to existing approaches, achieving an unmatched balance between memory and performance. To ensure the complete and immediate reproducibility of all our results, we release all code, scripts, and model checkpoints at https://github.com/RiccardoBravin/tiny-LLM.

Recent advancement in Automatic Speech Recognition (ASR) has produced large AI models, which become impractical for deployment in mobile devices. Model quantization is effective to produce compressed general-purpose models, however such models may only be deployed to a restricted sub-domain of interest. We show that ASR models can be personalized during quantization while relying on just a small set of unlabelled samples from the target domain. To this end, we propose myQASR, a mixed-precision quantization method that generates tailored quantization schemes for diverse users under any memory requirement with no fine-tuning. myQASR automatically evaluates the quantization sensitivity of network layers by analysing the full-precision activation values. We are then able to generate a personalised mixed-precision quantization scheme for any pre-determined memory budget. Results for large-scale ASR models show how myQASR improves performance for specific genders, languages, and speakers.

Learning-based 3D visual geometry models have benefited substantially from large-scale transformers. Among these, StreamVGGT leverages frame-wise causal attention for strong streaming reconstruction, but suffers from unbounded KV cache growth, leading to escalating memory consumption and inference latency as input frames accumulate. We propose XStreamVGGT, a tuning-free approach that systematically compresses the KV cache through joint pruning and quantization, enabling extremely memory-efficient streaming inference. Specifically, redundant KVs originating from multi-view inputs are pruned through efficient token importance identification, enabling a fixed memory budget. Leveraging the unique distribution of KV tensors, we incorporate KV quantization to further reduce memory consumption. Extensive evaluations show that XStreamVGGT achieves mostly negligible performance degradation while substantially reducing memory usage by 4.42times and accelerating inference by 5.48times, enabling scalable and practical streaming 3D applications. The code is available at https://github.com/ywh187/XStreamVGGT/.

Cross-attention transformers and other multimodal vision-language models excel at grounding and generation; however, their extensive, full-precision backbones make it challenging to deploy them on edge devices. Memory-augmented architectures enhance the utilization of past context; however, most works rarely pair them with aggressive edge-oriented quantization. We introduce BitMar, a quantized multimodal transformer that proposes an external human-like episodic memory for effective image-text generation on hardware with limited resources. BitMar utilizes 1.58-bit encoders, one for text (BitNet-style) and one for vision (DiNOv2-based), to create compact embeddings that are combined and used to query a fixed-size key-value episodic memory. During vector retrieval, the BitNet decoder applies per-layer conditioning, which increases the contextual relevance of generated content. The decoder also employs attention sinks with a sliding-window mechanism to process long or streaming inputs under tight memory budgets. The combination of per-layer conditioning and sliding-window attention achieves a strong quality-speed trade-off, delivering competitive captioning and multimodal understanding at low latency with a small model footprint. These characteristics make BitMar well-suited for edge deployment.

Due to the high price and heavy energy consumption of GPUs, deploying deep models on IoT devices such as microcontrollers makes significant contributions for ecological AI. Conventional methods successfully enable convolutional neural network inference of high resolution images on microcontrollers, while the framework for vision transformers that achieve the state-of-the-art performance in many vision applications still remains unexplored. In this paper, we propose a hardware-algorithm co-optimizations method called MCUFormer to deploy vision transformers on microcontrollers with extremely limited memory, where we jointly design transformer architecture and construct the inference operator library to fit the memory resource constraint. More specifically, we generalize the one-shot network architecture search (NAS) to discover the optimal architecture with highest task performance given the memory budget from the microcontrollers, where we enlarge the existing search space of vision transformers by considering the low-rank decomposition dimensions and patch resolution for memory reduction. For the construction of the inference operator library of vision transformers, we schedule the memory buffer during inference through operator integration, patch embedding decomposition, and token overwriting, allowing the memory buffer to be fully utilized to adapt to the forward pass of the vision transformer. Experimental results demonstrate that our MCUFormer achieves 73.62\% top-1 accuracy on ImageNet for image classification with 320KB memory on STM32F746 microcontroller. Code is available at https://github.com/liangyn22/MCUFormer.

Large Language Models (LLMs) have grown rapidly in size, creating significant challenges for efficient deployment on resource-constrained hardware. In this paper, we introduce Dynamic-Length Float (DFloat11), a lossless compression framework that reduces LLM size by 30% while preserving outputs that are bit-for-bit identical to the original model. DFloat11 is motivated by the low entropy in the BFloat16 weight representation of LLMs, which reveals significant inefficiency in existing storage format. By applying entropy coding, DFloat11 assigns dynamic-length encodings to weights based on frequency, achieving near information-optimal compression without any loss of precision. To facilitate efficient inference with dynamic-length encodings, we develop a custom GPU kernel for fast online decompression. Our design incorporates the following: (i) decomposition of memory-intensive lookup tables (LUTs) into compact LUTs that fit in GPU SRAM, (ii) a two-phase kernel for coordinating thread read/write positions using lightweight auxiliary variables, and (iii) transformer-block-level decompression to minimize latency. Experiments on recent models, including Llama-3.1, Qwen-2.5, and Gemma-3, validates our hypothesis that DFloat11 achieves around 30% model size reduction while preserving bit-for-bit exact outputs. Compared to a potential alternative of offloading parts of an uncompressed model to the CPU to meet memory constraints, DFloat11 achieves 1.9-38.8x higher throughput in token generation. With a fixed GPU memory budget, DFloat11 enables 5.3-13.17x longer context lengths than uncompressed models. Notably, our method enables lossless inference of Llama-3.1-405B, an 810GB model, on a single node equipped with 8x80GB GPUs. Our code and models are available at https://github.com/LeanModels/DFloat11.

Autonomous agents such as indoor drones must learn new object classes in real-time while limiting catastrophic forgetting, motivating Class-Incremental Learning (CIL). However, most unmanned aerial vehicle (UAV) datasets focus on outdoor scenes and offer limited temporally coherent indoor videos. We introduce an indoor dataset of 14,400 frames capturing inter-drone and ground vehicle footage, annotated via a semi-automatic workflow with a 98.6% first-pass labeling agreement before final manual verification. Using this dataset, we benchmark 3 replay-based CIL strategies: Experience Replay (ER), Maximally Interfered Retrieval (MIR), and Forgetting-Aware Replay (FAR), using YOLOv11-nano as a resource-efficient detector for deployment-constrained UAV platforms. Under tight memory budgets (5-10% replay), FAR performs better than the rest, achieving an average accuracy (ACC, mAP_{50-95} across increments) of 82.96% with 5% replay. Gradient-weighted class activation mapping (Grad-CAM) analysis shows attention shifts across classes in mixed scenes, which is associated with reduced localization quality for drones. The experiments further demonstrate that replay-based continual learning can be effectively applied to edge aerial systems. Overall, this work contributes an indoor UAV video dataset with preserved temporal coherence and an evaluation of replay-based CIL under limited replay budgets. Project page: https://spacetime-vision-robotics-laboratory.github.io/learning-on-the-fly-cl

Quantization has become a popular technique to compress neural networks and reduce compute cost, but most prior work focuses on studying quantization without changing the network size. Many real-world applications of neural networks have compute cost and memory budgets, which can be traded off with model quality by changing the number of parameters. In this work, we use ResNet as a case study to systematically investigate the effects of quantization on inference compute cost-quality tradeoff curves. Our results suggest that for each bfloat16 ResNet model, there are quantized models with lower cost and higher accuracy; in other words, the bfloat16 compute cost-quality tradeoff curve is Pareto-dominated by the 4-bit and 8-bit curves, with models primarily quantized to 4-bit yielding the best Pareto curve. Furthermore, we achieve state-of-the-art results on ImageNet for 4-bit ResNet-50 with quantization-aware training, obtaining a top-1 eval accuracy of 77.09%. We demonstrate the regularizing effect of quantization by measuring the generalization gap. The quantization method we used is optimized for practicality: It requires little tuning and is designed with hardware capabilities in mind. Our work motivates further research into optimal numeric formats for quantization, as well as the development of machine learning accelerators supporting these formats. As part of this work, we contribute a quantization library written in JAX, which is open-sourced at https://github.com/google-research/google-research/tree/master/aqt.

While large language models (LLMs) have enabled learning knowledge from the pre-training corpora, the acquired knowledge may be fundamentally incorrect or outdated over time, which necessitates rectifying the knowledge of the language model (LM) after the training. A promising approach involves employing a hyper-network to generate parameter shift, whereas existing hyper-networks suffer from inferior scalability in synchronous editing operation amount. To mitigate the problem, we propose the MAssive Language Model Editing Network (MALMEN), which formulates the parameter shift aggregation as the least square problem, subsequently updating the LM parameters using the normal equation. To accommodate editing multiple facts simultaneously with limited memory budgets, we separate the computation on the hyper-network and LM, enabling arbitrary batch size on both neural networks. Our method is evaluated by editing up to thousands of facts on LMs with different architectures, i.e., BERT-base, GPT-2, T5-XL (2.8B), and GPT-J (6B), across various knowledge-intensive NLP tasks, i.e., closed book fact-checking and question answering. Remarkably, MALMEN is capable of editing hundreds of times more facts than strong baselines with the identical hyper-network architecture and outperforms editor specifically designed for GPT. Our code is available at https://github.com/ChenmienTan/malmen.

Recent advances in large language models (LLMs) have extended context lengths, enabling assistants to sustain long histories for coherent, personalized responses. This ability, however, hinges on Key-Value (KV) caching, whose memory grows linearly with dialogue length and quickly dominates under strict resource constraints. An active line of research for reducing this overhead is KV cache compression, which seeks to limit cache size while preserving accuracy. Yet existing methods face two major limitations: (i) evicting entries after full-context prefill causes unbounded peak memory, and (ii) query-dependent eviction narrows the cache to a single query, leading to degraded accuracy in multi-turn conversations. We introduce EpiCache, a training-free KV cache management framework for long conversational question answering (LongConvQA) under fixed memory budgets. EpiCache bounds cache growth through block-wise prefill and preserves topic-relevant context via episodic KV compression, which clusters conversation history into coherent episodes and applies episode-specific KV cache eviction. We further design an adaptive layer-wise budget allocation strategy that measures each layer's sensitivity to eviction and distributes the memory budget across layers accordingly. Across three LongConvQA benchmarks, EpiCache improves accuracy by up to 40% over recent baselines, sustains near-full KV accuracy under 4-6x compression, and reduces latency and memory by up to 2.4x and 3.5x, thereby enabling efficient multi-turn interaction under strict resource constraints.

Recent advancements in speculative decoding have demonstrated considerable speedup across a wide array of large language model (LLM) tasks. Speculative decoding inherently relies on sacrificing extra memory allocations to generate several candidate tokens, of which acceptance rate drives the speedup. However, deploying speculative decoding on memory-constrained devices, such as mobile GPUs, remains as a significant challenge in real-world scenarios. In this work, we present a device-aware inference engine named SpecMemo that can smartly control memory allocations at finer levels to enable multi-turn chatbots with speculative decoding on such limited memory devices. Our methodology stems from theoretically modeling memory footprint of speculative decoding to determine a lower bound on the required memory budget while retaining speedup. SpecMemo empirically acquires a careful balance between minimizing redundant memory allocations for rejected candidate tokens and maintaining competitive performance gains from speculation. Notably, with SpecMemo's memory management, we maintain 96% of overall throughput from speculative decoding on MT-Bench, with reduced generation-memory by 65% on single Nvidia Titan RTX. Given multiple constrained GPUs, we build on top of previous speculative decoding architectures to facilitate big-model inference by distributing Llama-2-70B-Chat model, on which we provide novel batched speculative decoding to increase usability of multiple small server GPUs. This novel framework demonstrates 2x speedup over distributed and batched vanilla decoding with the base model on eight AMD MI250 GPUs. Moreover, inference throughput increases remarkably 8x with batch size 10. Our work contributes to democratized LLM applications in resource-constrained environments, providing a pathway for faster and cheaper deployment of real-world LLM applications with robust performance.

Compact pretrained bidirectional encoders remain the backbone of industrial NLP under tight compute and memory budgets. Their effectiveness stems from self-attention's ability to deliver high-quality bidirectional contextualization with sequence-level parallelism, as popularized by BERT-style architectures. Recently, Avey was introduced as an autoregressive, attention-free alternative that naturally admits an encoder-only adaptation. In this paper, we reformulate Avey for the encoder-only paradigm and propose several innovations to its architecture, including decoupled static and dynamic parameterizations, stability-oriented normalization, and neural compression. Results show that this reformulated architecture compares favorably to four widely used Transformer-based encoders, consistently outperforming them on standard token-classification and information-retrieval benchmarks while scaling more efficiently to long contexts.

Diffusion models have achieved impressive advancements in various vision tasks. However, these gains often rely on increasing model size, which escalates computational complexity and memory demands, complicating deployment, raising inference costs, and causing environmental impact. While some studies have explored pruning techniques to improve the memory efficiency of diffusion models, most existing methods require extensive retraining to retain the model performance. Retraining a modern large diffusion model is extremely costly and resource-intensive, which limits the practicality of these methods. In this work, we achieve low-cost diffusion pruning without retraining by proposing a model-agnostic structural pruning framework for diffusion models that learns a differentiable mask to sparsify the model. To ensure effective pruning that preserves the quality of the final denoised latent, we design a novel end-to-end pruning objective that spans the entire diffusion process. As end-to-end pruning is memory-intensive, we further propose time step gradient checkpointing, a technique that significantly reduces memory usage during optimization, enabling end-to-end pruning within a limited memory budget. Results on state-of-the-art U-Net diffusion models SDXL and diffusion transformers (FLUX) demonstrate that our method can effectively prune up to 20% parameters with minimal perceptible performance degradation, and notably, without the need for model retraining. We also showcase that our method can still prune on top of time step distilled diffusion models.

Deep models, e.g., CNNs and Vision Transformers, have achieved impressive achievements in many vision tasks in the closed world. However, novel classes emerge from time to time in our ever-changing world, requiring a learning system to acquire new knowledge continually. For example, a robot needs to understand new instructions, and an opinion monitoring system should analyze emerging topics every day. Class-Incremental Learning (CIL) enables the learner to incorporate the knowledge of new classes incrementally and build a universal classifier among all seen classes. Correspondingly, when directly training the model with new class instances, a fatal problem occurs -- the model tends to catastrophically forget the characteristics of former ones, and its performance drastically degrades. There have been numerous efforts to tackle catastrophic forgetting in the machine learning community. In this paper, we survey comprehensively recent advances in deep class-incremental learning and summarize these methods from three aspects, i.e., data-centric, model-centric, and algorithm-centric. We also provide a rigorous and unified evaluation of 16 methods in benchmark image classification tasks to find out the characteristics of different algorithms empirically. Furthermore, we notice that the current comparison protocol ignores the influence of memory budget in model storage, which may result in unfair comparison and biased results. Hence, we advocate fair comparison by aligning the memory budget in evaluation, as well as several memory-agnostic performance measures. The source code to reproduce these evaluations is available at https://github.com/zhoudw-zdw/CIL_Survey/

Recent advances in large language models (LLMs) have significantly boosted long-context processing. However, the increasing key-value (KV) cache size poses critical challenges to memory and execution efficiency. Most KV cache compression methods rely on heuristic token eviction using all attention heads in Grouped Query Attention (GQA)-based LLMs. This method ignores the different functionalities of attention heads, leading to the eviction of critical tokens and thus degrades the performance of LLMs. To address the issue above, instead of using all the attention heads in GQA-based LLMs to determine important tokens as in the previous work, we first identify the attention heads in each layer that are not only capable of retrieving the initial and final tokens of a prompt, but also capable of retrieving important tokens within the text and attending to their surrounding semantic context. Afterwards, we exploit such heads to determine the important tokens and retain their corresponding KV cache pairs. Furthermore, we analyze the cache eviction error of each layer individually and introduce a layer-adaptive KV cache allocation strategy. Experimental results demonstrate the proposed CompressKV consistently outperforms state-of-the-art approaches under various memory budgets on LongBench and Needle-in-a-Haystack benchmarks. Our code is publicly available at: https://github.com/TUDa-HWAI/CompressKV.git.

We present Mobile Video Networks (MoViNets), a family of computation and memory efficient video networks that can operate on streaming video for online inference. 3D convolutional neural networks (CNNs) are accurate at video recognition but require large computation and memory budgets and do not support online inference, making them difficult to work on mobile devices. We propose a three-step approach to improve computational efficiency while substantially reducing the peak memory usage of 3D CNNs. First, we design a video network search space and employ neural architecture search to generate efficient and diverse 3D CNN architectures. Second, we introduce the Stream Buffer technique that decouples memory from video clip duration, allowing 3D CNNs to embed arbitrary-length streaming video sequences for both training and inference with a small constant memory footprint. Third, we propose a simple ensembling technique to improve accuracy further without sacrificing efficiency. These three progressive techniques allow MoViNets to achieve state-of-the-art accuracy and efficiency on the Kinetics, Moments in Time, and Charades video action recognition datasets. For instance, MoViNet-A5-Stream achieves the same accuracy as X3D-XL on Kinetics 600 while requiring 80% fewer FLOPs and 65% less memory. Code will be made available at https://github.com/tensorflow/models/tree/master/official/vision.

Long sequence modeling has gained broad interest as large language models (LLMs) continue to advance. Recent research has identified that a large portion of hidden states within the key-value caches of Transformer models can be discarded (also termed evicted) without affecting the perplexity performance in generating long sequences. However, we show that these methods, despite preserving perplexity performance, often drop information that is important for solving downstream tasks, a problem which we call information neglect. To address this issue, we introduce Chunked Instruction-aware State Eviction (CItruS), a novel modeling technique that integrates the attention preferences useful for a downstream task into the eviction process of hidden states. In addition, we design a method for chunked sequence processing to further improve efficiency. Our training-free method exhibits superior performance on long sequence comprehension and retrieval tasks over several strong baselines under the same memory budget, while preserving language modeling perplexity.

Alignment of Large Language Models (LLMs) typically relies on Reinforcement Learning from Human Feedback (RLHF) with gradient-based optimizers such as Proximal Policy Optimization (PPO) or Group Relative Policy Optimization (GRPO). While effective, these methods require complex distributed training, large memory budgets, and careful hyperparameter tuning, all of which become increasingly difficult at billion-parameter scale. We present ESSA, Evolutionary Strategies for Scalable Alignment, a gradient-free framework that aligns LLMs using only forward inference and black-box optimization. ESSA focuses optimization on Low-Rank Adapters (LoRA) and further compresses their parameter space by optimizing only the singular values from an SVD decomposition of each adapter matrix. This dimensionality reduction makes evolutionary search practical even for very large models and allows efficient operation in quantized INT4 and INT8 inference mode. Across these benchmarks ESSA improves the test accuracy of Qwen2.5-Math-7B by 12.6% on GSM8K and 14.8% on PRM800K, and raises the accuracy of LLaMA3.1-8B on IFEval by 22.5%, all compared with GRPO. In large-scale settings ESSA shows stronger scaling than gradient-based methods: on Qwen2.5-32B for PRM800K it reaches near-optimal accuracy twice as fast on 16 GPUs and six times as fast on 128 GPUs compared with GRPO. These results position evolutionary strategies as a compelling, hardware-friendly alternative to gradient-based LLM alignment, combining competitive quality with substantially reduced wall-clock time and engineering overhead.

We investigate parameter-efficient fine-tuning (PEFT) methods that can provide good accuracy under limited computational and memory budgets in the context of large language models (LLMs). We present a new PEFT method called Robust Adaptation (RoSA) inspired by robust principal component analysis (PCA) that jointly trains low-rank and highly-sparse components on top of a set of fixed pretrained weights to efficiently approximate the performance of a full-fine-tuning (FFT) solution. Across a series of challenging generative tasks such as grade-school math and SQL query generation, which require fine-tuning for good performance, we show that RoSA outperforms both LoRA and pure sparse fine-tuning, at the same parameter budget. We provide system support for RoSA to complement the training algorithm, specifically in the form of sparse GPU kernels which enable memory- and computationally-efficient training. Our code will be made available at https://github.com/IST-DASLab/RoSA.

Parameter-efficient fine-tuning (PEFT) has become the standard approach for adapting large language models under limited compute and memory budgets. Although previous methods improve efficiency through low-rank updates, quantization, or heuristic budget reallocation, they often decouple the allocation of capacity from the way updates evolve during training. In this work, we introduce CTR-LoRA, a framework guided by curvature trust region that integrates rank scheduling with stability-aware optimization. CTR-LoRA allocates parameters based on marginal utility derived from lightweight second-order proxies and constrains updates using a Fisher/Hessian-metric trust region. Experiments on multiple open-source backbones (7B-13B), evaluated on both in-distribution and out-of-distribution benchmarks, show consistent improvements over strong PEFT baselines. In addition to increased accuracy, CTR-LoRA enhances training stability, reduces memory requirements, and achieves higher throughput, positioning it on the Pareto frontier of performance and efficiency. These results highlight a principled path toward more robust and deployable PEFT.

Efficiently adapting large foundation models is critical, especially with tight compute and memory budgets. Parameter-Efficient Fine-Tuning (PEFT) methods such as LoRA offer limited granularity and effectiveness in few-parameter regimes. We propose Wavelet Fine-Tuning (WaveFT), a novel PEFT method that learns highly sparse updates in the wavelet domain of residual matrices. WaveFT allows precise control of trainable parameters, offering fine-grained capacity adjustment and excelling with remarkably low parameter count, potentially far fewer than LoRA's minimum, ideal for extreme parameter-efficient scenarios. Evaluated on personalized text-to-image generation using Stable Diffusion XL as baseline, WaveFT significantly outperforms LoRA and other PEFT methods, especially at low parameter counts; achieving superior subject fidelity, prompt alignment, and image diversity.

Large Language Models (LLMs) have demonstrated impressive performance across various tasks, and their application in edge scenarios has attracted significant attention. However, sparse-activated Mixture-of-Experts (MoE) models, which are well suited for edge scenarios, have received relatively little attention due to their high memory demands. Offload-based methods have been proposed to address this challenge, but they face difficulties with expert prediction. Inaccurate expert predictions can result in prolonged inference delays. To promote the application of MoE models in edge scenarios, we propose Fate, an offloading system designed for MoE models to enable efficient inference in resource-constrained environments. The key insight behind Fate is that gate inputs from adjacent layers can be effectively used for expert prefetching, achieving high prediction accuracy without additional GPU overhead. Furthermore, Fate employs a shallow-favoring expert caching strategy that increases the expert hit rate to 99\%. Additionally, Fate integrates tailored quantization strategies for cache optimization and IO efficiency. Experimental results show that, compared to Load on Demand and Expert Activation Path-based method, Fate achieves up to 4.5x and 1.9x speedups in prefill speed and up to 4.1x and 2.2x speedups in decoding speed, respectively, while maintaining inference quality. Moreover, Fate's performance improvements are scalable across different memory budgets.

How to efficiently serve Large Language Models (LLMs) has become a pressing issue because of their huge computational cost in their autoregressive generation process. To mitigate computational costs, LLMs often employ the KV Cache technique to improve the generation speed. While improving the computational efficiency, the storage requirements of the KV cache are substantial, particularly in long-context scenarios, leading to significant memory consumption. Existing KV cache eviction methods often degrade the performance of LLMs in long-context scenarios due to the information loss introduced by eviction. In this paper, we propose a novel KV cache merging approach, called KVMerger, to achieve adaptive KV cache compression for long-context tasks without significant performance degradation under constrained memory budgets. Our approach is inspired by the intriguing observation that key states exhibit high similarity at the token level within a single sequence. To facilitate merging, we develop an effective yet straightforward merging set identification algorithm to identify suitable KV states for merging. Our merging set identification algorithm stimulates the second observation that KV cache sparsity, from similarity perspective, is independent of the dataset and remains persistent at the model level. Subsequently, we propose a Gaussian kernel weighted merging algorithm to selectively merge all states within each merging set. We conduct extensive experiments to demonstrate the effectiveness of KVMerger for long-context tasks under constrained memory budgets, applying it to models including Llama2-7B-chat and Llama2-13B-chat. Using the LongBench and ZeroScroll benchmarks, we compare our method with other KV cache compression techniques, including H2O and CaM, showing that our method achieves superior performance across tasks with both 50% and 35% KV cache budgets.

The deployment and training of neural networks on edge computing devices pose many challenges. The low memory nature of edge devices is often one of the biggest limiting factors encountered in the deployment of large neural network models. Tensor rematerialization or recompute is a way to address high memory requirements for neural network training and inference. In this paper we consider the problem of execution time minimization of compute graphs subject to a memory budget. In particular, we develop a new constraint programming formulation called Moccasin with only O(n) integer variables, where n is the number of nodes in the compute graph. This is a significant improvement over the works in the recent literature that propose formulations with O(n^2) Boolean variables. We present numerical studies that show that our approach is up to an order of magnitude faster than recent work especially for large-scale graphs.

Mixture-of-Experts (MoE) models achieve a favorable trade-off between performance and inference efficiency by activating only a subset of experts. However, the memory overhead of storing all experts remains a major limitation, especially in large-scale MoE models such as DeepSeek-R1(671B). In this study, we investigate domain specialization and expert redundancy in large-scale MoE models and uncover a consistent behavior we term few-shot expert localization, with only a few in-domain demonstrations, the model consistently activates a sparse and stable subset of experts on tasks within the same domain. Building on this observation, we propose a simple yet effective pruning framework, EASY-EP, that leverages a few domain-specific demonstrations to identify and retain only the most relevant experts. EASY-EP comprises two key components: output-aware expert importance assessment and expert-level token contribution estimation. The former evaluates the importance of each expert for the current token by considering the gating scores and L2 norm of the outputs of activated experts, while the latter assesses the contribution of tokens based on representation similarities before and after routed experts. Experiments on DeepSeek-R1 and DeepSeek-V3-0324 show that our method can achieve comparable performances and 2.99times throughput under the same memory budget with full model with only half the experts.

Diffusion Large Language Models (DLLMs) have emerged as a new paradigm of language modeling beyond autoregressive next-token prediction. Thanks to their bidirectional attention mechanism, DLLMs are more capable of capturing the connection of context, and thus show unique advantages in challenges like the famous "reversal curse" or learning under data-constrained scenarios. In addition, taking advantage of their inherent modeling foundations, DLLMs have the great potential of efficient inference with parallel decoding algorithms, which enable multi-token prediction per step. However, the high generation quality often requires the number of decoding steps equal to the sequence length, which performs a one-token-per-step decoding, and existing parallel decoding algorithms, which yield suboptimal decoding paths, bring inference speedup at the cost of non-negligible performance degradation. To overcome this challenge, we introduce Free Draft-and-Verification (FreeDave), a novel fast decoding algorithm tailored for DLLMs that achieves lossless parallel decoding without any model modification or extra modules. Specifically, we propose an algorithm of parallel-decoded candidate generation and verification, which is theoretically guaranteed to use the fewest model forward calls to reproduce the same sequence generated by static decoding when enough computation and memory budget is provided. By extensive evaluations on math reasoning and code generation benchmarks across different DLLMs, FreeDave is proven to boost the inference throughput up to 3.78times without performance degradation.

Large Vision and Language Models have exhibited remarkable human-like intelligence in tasks such as natural language comprehension, problem-solving, logical reasoning, and knowledge retrieval. However, training and serving these models require substantial computational resources, posing a significant barrier to their widespread application and further research. To mitigate this challenge, various model compression techniques have been developed to reduce computational requirements. Nevertheless, existing methods often employ uniform quantization configurations, failing to account for the varying difficulties across different layers in quantizing large neural network models. This paper tackles this issue by leveraging layer-sensitivity features, such as activation sensitivity and weight distribution Kurtosis, to identify layers that are challenging to quantize accurately and allocate additional memory budget. The proposed methods, named SensiBoost and KurtBoost, respectively, demonstrate notable improvement in quantization accuracy, achieving up to 9% lower perplexity with only a 2% increase in memory budget on LLama models compared to the baseline.

Incrementally fine-tuning foundational models on new tasks or domains is now the de facto approach in NLP. A known pitfall of this approach is the catastrophic forgetting of prior knowledge that happens during fine-tuning. A common approach to alleviate such forgetting is to rehearse samples from prior tasks during fine-tuning. Several existing works assume a fixed memory buffer to store prior task examples, while relying on inferences (forward passes) with the model at hand for choosing examples for rehearsal from the buffer. However, given the increasing computational cost of model inference, and decreasing cost of data storage, we focus on the setting to rehearse samples with a fixed computational budget instead of a fixed memory budget. We propose a sampling scheme, \bf mix-cd, that prioritizes rehearsal of ``collateral damage'' samples, which are samples predicted correctly by the prior model but forgotten by the incrementally tuned one. The crux of our scheme is a procedure to efficiently estimate the density of collateral damage samples without incurring additional model inferences. Our approach is computationally efficient, easy to implement, and outperforms several leading continual learning methods in compute-constrained settings. All the code will be publicly available at https://github.com/jybai/mix-cd-rehearsal.

Real-time rendering and animation of humans is a core function in games, movies, and telepresence applications. Existing methods have a number of drawbacks we aim to address with our work. Triangle meshes have difficulty modeling thin structures like hair, volumetric representations like Neural Volumes are too low-resolution given a reasonable memory budget, and high-resolution implicit representations like Neural Radiance Fields are too slow for use in real-time applications. We present Mixture of Volumetric Primitives (MVP), a representation for rendering dynamic 3D content that combines the completeness of volumetric representations with the efficiency of primitive-based rendering, e.g., point-based or mesh-based methods. Our approach achieves this by leveraging spatially shared computation with a deconvolutional architecture and by minimizing computation in empty regions of space with volumetric primitives that can move to cover only occupied regions. Our parameterization supports the integration of correspondence and tracking constraints, while being robust to areas where classical tracking fails, such as around thin or translucent structures and areas with large topological variability. MVP is a hybrid that generalizes both volumetric and primitive-based representations. Through a series of extensive experiments we demonstrate that it inherits the strengths of each, while avoiding many of their limitations. We also compare our approach to several state-of-the-art methods and demonstrate that MVP produces superior results in terms of quality and runtime performance.

Current architectures for video understanding mainly build upon 3D convolutional blocks or 2D convolutions with additional operations for temporal modeling. However, these methods all regard the temporal axis as a separate dimension of the video sequence, which requires large computation and memory budgets and thus limits their usage on mobile devices. In this paper, we propose to squeeze the time axis of a video sequence into the channel dimension and present a lightweight video recognition network, term as SqueezeTime, for mobile video understanding. To enhance the temporal modeling capability of the proposed network, we design a Channel-Time Learning (CTL) Block to capture temporal dynamics of the sequence. This module has two complementary branches, in which one branch is for temporal importance learning and another branch with temporal position restoring capability is to enhance inter-temporal object modeling ability. The proposed SqueezeTime is much lightweight and fast with high accuracies for mobile video understanding. Extensive experiments on various video recognition and action detection benchmarks, i.e., Kinetics400, Kinetics600, HMDB51, AVA2.1 and THUMOS14, demonstrate the superiority of our model. For example, our SqueezeTime achieves +1.2% accuracy and +80% GPU throughput gain on Kinetics400 than prior methods. Codes are publicly available at https://github.com/xinghaochen/SqueezeTime and https://github.com/mindspore-lab/models/tree/master/research/huawei-noah/SqueezeTime.

Adversarial examples provoke weak reliability and potential security issues in deep neural networks. Although adversarial training has been widely studied to improve adversarial robustness, it works in an over-parameterized regime and requires high computations and large memory budgets. To bridge adversarial robustness and model compression, we propose a novel adversarial pruning method, Masking Adversarial Damage (MAD) that employs second-order information of adversarial loss. By using it, we can accurately estimate adversarial saliency for model parameters and determine which parameters can be pruned without weakening adversarial robustness. Furthermore, we reveal that model parameters of initial layer are highly sensitive to the adversarial examples and show that compressed feature representation retains semantic information for the target objects. Through extensive experiments on three public datasets, we demonstrate that MAD effectively prunes adversarially trained networks without loosing adversarial robustness and shows better performance than previous adversarial pruning methods.

Cooperative perception through Vehicle-to-Everything (V2X) communication offers significant potential for enhancing vehicle perception by mitigating occlusions and expanding the field of view. However, past research has predominantly focused on improving accuracy metrics without addressing the crucial system-level considerations of efficiency, latency, and real-world deployability. Noticeably, most existing systems rely on full-precision models, which incur high computational and transmission costs, making them impractical for real-time operation in resource-constrained environments. In this paper, we introduce QuantV2X, the first fully quantized multi-agent system designed specifically for efficient and scalable deployment of multi-modal, multi-agent V2X cooperative perception. QuantV2X introduces a unified end-to-end quantization strategy across both neural network models and transmitted message representations that simultaneously reduces computational load and transmission bandwidth. Remarkably, despite operating under low-bit constraints, QuantV2X achieves accuracy comparable to full-precision systems. More importantly, when evaluated under deployment-oriented metrics, QuantV2X reduces system-level latency by 3.2times and achieves a +9.5 improvement in mAP30 over full-precision baselines. Furthermore, QuantV2X scales more effectively, enabling larger and more capable models to fit within strict memory budgets. These results highlight the viability of a fully quantized multi-agent intermediate fusion system for real-world deployment. The system will be publicly released to promote research in this field: https://github.com/ucla-mobility/QuantV2X.

We propose a simple approach for memory-efficient adaptation of pretrained language models. Our approach uses an iterative algorithm to decompose each pretrained matrix into a high-precision low-rank component and a memory-efficient quantized component. During finetuning, the quantized component remains fixed and only the low-rank component is updated. We present an integer linear programming formulation of the quantization component which enables dynamic configuration of quantization parameters (e.g., bit-width, block size) for each matrix given an overall target memory budget. We further explore a data-aware version of the algorithm which uses an approximation of the Fisher information matrix to weight the reconstruction objective during matrix decomposition. Experiments on adapting RoBERTa and LLaMA-2 (7B and 70B) demonstrate that our low-rank plus quantized matrix decomposition approach (LQ-LoRA) outperforms strong QLoRA and GPTQ-LoRA baselines and moreover enables more aggressive quantization. For example, on the OpenAssistant benchmark LQ-LoRA is able to learn a 2.5-bit LLaMA-2 model that is competitive with a model finetuned with 4-bit QLoRA. When finetuned on a language modeling calibration dataset, LQ-LoRA can also be used for model compression; in this setting our 2.75-bit LLaMA-2-70B model (which has 2.85 bits on average when including the low-rank components and requires 27GB of GPU memory) is competitive with the original model in full precision.

Approximate Nearest Neighbor Search (ANNS), as the core of vector databases (VectorDBs), has become widely used in modern AI and ML systems, powering applications from information retrieval to bio-informatics. While graph-based ANNS methods achieve high query efficiency, their scalability is constrained by the available host memory. Recent disk-based ANNS approaches mitigate memory usage by offloading data to Solid-State Drives (SSDs). However, they still suffer from issues such as long I/O traversal path, misalignment with storage I/O granularity, and high in-memory indexing overhead, leading to significant I/O latency and ultimately limiting scalability for large-scale vector search. In this paper, we propose PageANN, a disk-based approximate nearest neighbor search (ANNS) framework designed for high performance and scalability. PageANN introduces a page-node graph structure that aligns logical graph nodes with physical SSD pages, thereby shortening I/O traversal paths and reducing I/O operations. Specifically, similar vectors are clustered into page nodes, and a co-designed disk data layout leverages this structure with a merging technique to store only representative vectors and topology information, avoiding unnecessary reads. To further improve efficiency, we design a memory management strategy that combines lightweight indexing with coordinated memory-disk data allocation, maximizing host memory utilization while minimizing query latency and storage overhead. Experimental results show that PageANN significantly outperforms state-of-the-art (SOTA) disk-based ANNS methods, achieving 1.85x-10.83x higher throughput and 51.7%-91.9% lower latency across different datasets and memory budgets, while maintaining comparable high recall accuracy.

Large language models (LLMs) rely on key-value cache (KV cache) to accelerate decoding by reducing redundant computations. However, the KV cache memory usage grows substantially with longer text sequences, posing challenges for efficient deployment. Existing KV cache eviction methods prune tokens using prefilling-stage attention scores, causing inconsistency with actual inference queries, especially under tight memory budgets. In this paper, we propose Lookahead Q-Cache (LAQ), a novel eviction framework that generates low-cost pseudo lookahead queries to better approximate the true decoding-stage queries. By using these lookahead queries as the observation window for importance estimation, LAQ achieves more consistent and accurate KV cache eviction aligned with real inference scenarios. Experimental results on LongBench and Needle-in-a-Haystack benchmarks show that LAQ outperforms existing methods across various budget levels, achieving a 1 sim 4 point improvement on LongBench under limited cache budget. Moreover, LAQ is complementary to existing approaches and can be flexibly combined to yield further improvements.

The mixture of experts (MoE) model is a sparse variant of large language models (LLMs), designed to hold a better balance between intelligent capability and computational overhead. Despite its benefits, MoE is still too expensive to deploy on resource-constrained edge devices, especially with the demands of on-device inference services. Recent research efforts often apply model compression techniques, such as quantization, pruning and merging, to restrict MoE complexity. Unfortunately, due to their predefined static model optimization strategies, they cannot always achieve the desired quality-overhead trade-off when handling multiple requests, finally degrading the on-device quality of service. These limitations motivate us to propose the D^2MoE, an algorithm-system co-design framework that matches diverse task requirements by dynamically allocating the most proper bit-width to each expert. Specifically, inspired by the nested structure of matryoshka dolls, we propose the matryoshka weight quantization (MWQ) to progressively compress expert weights in a bit-nested manner and reduce the required runtime memory. On top of it, we further optimize the I/O-computation pipeline and design a heuristic scheduling algorithm following our hottest-expert-bit-first (HEBF) principle, which maximizes the expert parallelism between I/O and computation queue under constrained memory budgets, thus significantly reducing the idle temporal bubbles waiting for the experts to load. Evaluations on real edge devices show that D^2MoE improves the overall inference throughput by up to 1.39times and reduces the peak memory footprint by up to 53% over the latest on-device inference frameworks, while still preserving comparable serving accuracy as its INT8 counterparts.

Large language models (LLMs) excel at processing long sequences, boosting demand for key-value (KV) caching. While recent efforts to evict KV cache have alleviated the inference burden, they often fail to allocate resources rationally across layers with different attention patterns. In this paper, we introduce Cascading and Adaptive KV cache Eviction (CAKE), a novel approach that frames KV cache eviction as a "cake-slicing problem." CAKE assesses layer-specific preferences by considering attention dynamics in both spatial and temporal dimensions, allocates rational cache size for layers accordingly, and manages memory constraints in a cascading manner. This approach enables a global view of cache allocation, adaptively distributing resources across diverse attention mechanisms while maintaining memory budgets. CAKE also employs a new eviction indicator that considers the shifting importance of tokens over time, addressing limitations in existing methods that overlook temporal dynamics. Comprehensive experiments on LongBench and NeedleBench show that CAKE maintains model performance with only 3.2% of the KV cache and consistently outperforms current baselines across various models and memory constraints, particularly in low-memory settings. Additionally, CAKE achieves over 10x speedup in decoding latency compared to full cache when processing contexts of 128K tokens with FlashAttention-2. Our code is available at https://github.com/antgroup/cakekv.

Learnable embedding vector is one of the most important applications in machine learning, and is widely used in various database-related domains. However, the high dimensionality of sparse data in recommendation tasks and the huge volume of corpus in retrieval-related tasks lead to a large memory consumption of the embedding table, which poses a great challenge to the training and deployment of models. Recent research has proposed various methods to compress the embeddings at the cost of a slight decrease in model quality or the introduction of other overheads. Nevertheless, the relative performance of these methods remains unclear. Existing experimental comparisons only cover a subset of these methods and focus on limited metrics. In this paper, we perform a comprehensive comparative analysis and experimental evaluation of embedding compression. We introduce a new taxonomy that categorizes these techniques based on their characteristics and methodologies, and further develop a modular benchmarking framework that integrates 14 representative methods. Under a uniform test environment, our benchmark fairly evaluates each approach, presents their strengths and weaknesses under different memory budgets, and recommends the best method based on the use case. In addition to providing useful guidelines, our study also uncovers the limitations of current methods and suggests potential directions for future research.

Kernel continual learning by derakhshani2021kernel has recently emerged as a strong continual learner due to its non-parametric ability to tackle task interference and catastrophic forgetting. Unfortunately its success comes at the expense of an explicit memory to store samples from past tasks, which hampers scalability to continual learning settings with a large number of tasks. In this paper, we introduce generative kernel continual learning, which explores and exploits the synergies between generative models and kernels for continual learning. The generative model is able to produce representative samples for kernel learning, which removes the dependence on memory in kernel continual learning. Moreover, as we replay only on the generative model, we avoid task interference while being computationally more efficient compared to previous methods that need replay on the entire model. We further introduce a supervised contrastive regularization, which enables our model to generate even more discriminative samples for better kernel-based classification performance. We conduct extensive experiments on three widely-used continual learning benchmarks that demonstrate the abilities and benefits of our contributions. Most notably, on the challenging SplitCIFAR100 benchmark, with just a simple linear kernel we obtain the same accuracy as kernel continual learning with variational random features for one tenth of the memory, or a 10.1\% accuracy gain for the same memory budget.

Recently, dataset distillation has paved the way towards efficient machine learning, especially for image datasets. However, the distillation for videos, characterized by an exclusive temporal dimension, remains an underexplored domain. In this work, we provide the first systematic study of video distillation and introduce a taxonomy to categorize temporal compression. Our investigation reveals that the temporal information is usually not well learned during distillation, and the temporal dimension of synthetic data contributes little. The observations motivate our unified framework of disentangling the dynamic and static information in the videos. It first distills the videos into still images as static memory and then compensates the dynamic and motion information with a learnable dynamic memory block. Our method achieves state-of-the-art on video datasets at different scales, with a notably smaller memory storage budget. Our code is available at https://github.com/yuz1wan/video_distillation.

Memory is increasingly central to Large Language Model (LLM) agents operating beyond a single context window, yet most existing systems rely on offline, query-agnostic memory construction that can be inefficient and may discard query-critical information. Although runtime memory utilization is a natural alternative, prior work often incurs substantial overhead and offers limited explicit control over the performance-cost trade-off. In this work, we present BudgetMem, a runtime agent memory framework for explicit, query-aware performance-cost control. BudgetMem structures memory processing as a set of memory modules, each offered in three budget tiers (i.e., Low/Mid/High). A lightweight router performs budget-tier routing across modules to balance task performance and memory construction cost, which is implemented as a compact neural policy trained with reinforcement learning. Using BudgetMem as a unified testbed, we study three complementary strategies for realizing budget tiers: implementation (method complexity), reasoning (inference behavior), and capacity (module model size). Across LoCoMo, LongMemEval, and HotpotQA, BudgetMem surpasses strong baselines when performance is prioritized (i.e., high-budget setting), and delivers better accuracy-cost frontiers under tighter budgets. Moreover, our analysis disentangles the strengths and weaknesses of different tiering strategies, clarifying when each axis delivers the most favorable trade-offs under varying budget regimes.

Long contexts challenge transformers: attention scores dilute across thousands of tokens, critical information is often lost in the middle, and models struggle to adapt to novel patterns at inference time. Recent work on test-time adaptation addresses this by maintaining a form of working memory -- transient parameters updated on the current context -- but existing approaches rely on uniform write policies that waste computation on low-utility regions and suffer from high gradient variance across semantically heterogeneous contexts. In this work, we reframe test-time adaptation as a budget-constrained memory consolidation problem, focusing on which parts of the context should be consolidated into working memory under limited computation. We propose Gdwm (Gated Differentiable Working Memory), a framework that introduces a write controller to gate the consolidation process. The controller estimates Contextual Utility, an information-theoretic measure of long-range contextual dependence, and allocates gradient steps accordingly while maintaining global coverage. Experiments on ZeroSCROLLS and LongBench v2 demonstrate that Gdwm achieves comparable or superior performance with 4times fewer gradient steps than uniform baselines, establishing a new efficiency-performance Pareto frontier for test-time adaptation.

Mixture of Experts (MoE) architectures have significantly increased computational efficiency in both research and real-world applications of large-scale machine learning models. However, their scalability and efficiency under memory constraints remain relatively underexplored. In this work, we present joint scaling laws for dense and MoE models, incorporating key factors such as the number of active parameters, dataset size, and the number of experts. Our findings provide a principled framework for selecting the optimal MoE configuration under fixed memory and compute budgets. Surprisingly, we show that MoE models can be more memory-efficient than dense models, contradicting conventional wisdom. To derive and validate the theoretical predictions of our scaling laws, we conduct over 280 experiments with up to 2.7B active parameters and up to 5B total parameters. These results offer actionable insights for designing and deploying MoE models in practical large-scale training scenarios.

Retrieval-Augmented Generation (RAG) systems often rely on fixed top-k document selection mechanisms that ignore downstream generation quality and impose computational overheads. We propose SRAS (Sparse Reward-Aware Selector), a lightweight document selector trained via reinforcement learning (RL) for edge-native RAG deployment. Unlike prior RL-based retrievers that assume large memory and latency budgets, SRAS learns a compact (~0.76MB) policy using Proximal Policy Optimization (PPO), guided by a hybrid reward signal combining Relaxed F1 and BERTScore. Our method operates under tight token and compute constraints, maintaining <1s latency on CPU. SRAS outperforms supervised and random selectors on a synthetic QA benchmark, and generalizes to real-world data, achieving BERTScore F1 of 0.8546 on SQuAD v2 without domain-specific tuning. This work is the first to demonstrate that RL-based document selection can be made ultra-lightweight, latency-aware, and effective for on-device RAG pipelines.

Matryoshka Quantization (MatQuant) is a recent quantization approach showing that a single integer-quantized model can be served across multiple precisions, by slicing the most significant bits (MSB) at inference time. This enables a single checkpoint to cover a wide range of memory and latency budgets, but renders quantization much more challenging. In particular, the initial MatQuant relies on expensive quantization-aware training (QAT) variants, rather than fast one-shot post training quantization (PTQ), and lacks open-source and kernel support. We address all of these limitations by introducing Post-Training Matryoshka Quantization (MatGPTQ), a new PTQ pipeline that produces a single parent model jointly optimized for multiple target precisions in one-shot, based on a small calibration set. MatGPTQ casts Matryoshka quantization as a multi-precision objective with bit-slicing and cross-bit error compensation, resulting in an algorithm that produces a multi-bit-width, "sliceable" model in a single pass. We also incorporate a new budget-aware search for heterogeneous per-layer bit-witdhs and provide efficient kernels that implement slicing and mixed-precision execution. Across standard LLMs and benchmarks, MatGPTQ preserves high-bit accuracy while substantially improving performance at low-bit-witdh settings. Overall, we establish a new state of the art for Matryoshka-style post-training quantization and make single-checkpoint, multi-precision deployment open and practical. Code is available at https://github.com/IST-DASLab/MatGPTQ.

In this work, we demonstrate that distinctive keys during LLM inference tend to have high attention scores. We explore this phenomenon and propose KeyDiff, a training-free KV cache eviction method based on key similarity. This method facilitates the deployment of LLM-based application requiring long input prompts in resource-constrained environments with limited memory and compute budgets. Unlike other KV cache eviction methods, KeyDiff can process arbitrarily long prompts within strict resource constraints and efficiently generate responses. We demonstrate that KeyDiff computes the optimal solution to a KV cache selection problem that maximizes key diversity, providing a theoretical understanding of KeyDiff. Notably,KeyDiff does not rely on attention scores, allowing the use of optimized attention mechanisms like FlashAttention. We demonstrate the effectiveness of KeyDiff across diverse tasks and models, illustrating a performance gap of less than 0.04\% with 8K cache budget (sim 23\% KV cache reduction) from the non-evicting baseline on the LongBench benchmark for Llama 3.1-8B and Llama 3.2-3B.

Neural network pruning is an essential technique for reducing the size and complexity of deep neural networks, enabling large-scale models on devices with limited resources. However, existing pruning approaches heavily rely on training data for guiding the pruning strategies, making them ineffective for federated learning over distributed and confidential datasets. Additionally, the memory- and computation-intensive pruning process becomes infeasible for recourse-constrained devices in federated learning. To address these challenges, we propose FedTiny, a distributed pruning framework for federated learning that generates specialized tiny models for memory- and computing-constrained devices. We introduce two key modules in FedTiny to adaptively search coarse- and finer-pruned specialized models to fit deployment scenarios with sparse and cheap local computation. First, an adaptive batch normalization selection module is designed to mitigate biases in pruning caused by the heterogeneity of local data. Second, a lightweight progressive pruning module aims to finer prune the models under strict memory and computational budgets, allowing the pruning policy for each layer to be gradually determined rather than evaluating the overall model structure. The experimental results demonstrate the effectiveness of FedTiny, which outperforms state-of-the-art approaches, particularly when compressing deep models to extremely sparse tiny models. FedTiny achieves an accuracy improvement of 2.61% while significantly reducing the computational cost by 95.91% and the memory footprint by 94.01% compared to state-of-the-art methods.

Recent advances in large language models (LLMs) enable agentic systems trained with reinforcement learning (RL) over multi-turn interaction trajectories, but practical deployment is bottlenecked by rapidly growing textual histories that inflate token budgets and memory usage. We introduce AgentOCR, a framework that exploits the superior information density of visual tokens by representing the accumulated observation-action history as a compact rendered image. To make multi-turn rollouts scalable, AgentOCR proposes segment optical caching. By decomposing history into hashable segments and maintaining a visual cache, this mechanism eliminates redundant re-rendering. Beyond fixed rendering, AgentOCR introduces agentic self-compression, where the agent actively emits a compression rate and is trained with compression-aware reward to adaptively balance task success and token efficiency. We conduct extensive experiments on challenging agentic benchmarks, ALFWorld and search-based QA. Remarkably, results demonstrate that AgentOCR preserves over 95\% of text-based agent performance while substantially reducing token consumption (>50\%), yielding consistent token and memory efficiency. Our further analysis validates a 20x rendering speedup from segment optical caching and the effective strategic balancing of self-compression.

The rising demand for energy-efficient edge AI systems (e.g., mobile agents/robots) has increased the interest in neuromorphic computing, since it offers ultra-low power/energy AI computation through spiking neural network (SNN) algorithms on neuromorphic processors. However, their efficient implementation strategy has not been comprehensively studied, hence limiting SNN deployments for edge AI systems. Toward this, we propose a design methodology to enable efficient SNN processing on commodity neuromorphic processors. To do this, we first study the key characteristics of targeted neuromorphic hardware (e.g., memory and compute budgets), and leverage this information to perform compatibility analysis for network selection. Afterward, we employ a mapping strategy for efficient SNN implementation on the targeted processor. Furthermore, we incorporate an efficient on-chip learning mechanism to update the systems' knowledge for adapting to new input classes and dynamic environments. The experimental results show that the proposed methodology leads the system to achieve low latency of inference (i.e., less than 50ms for image classification, less than 200ms for real-time object detection in video streaming, and less than 1ms in keyword recognition) and low latency of on-chip learning (i.e., less than 2ms for keyword recognition), while incurring less than 250mW of processing power and less than 15mJ of energy consumption across the respective different applications and scenarios. These results show the potential of the proposed methodology in enabling efficient edge AI systems for diverse application use-cases.

Long-horizon conversational agents require persistent memory for coherent reasoning, yet uncontrolled accumulation causes temporal decay and false memory propagation. Benchmarks such as LOCOMO and LOCCO report performance degradation from 0.455 to 0.05 across stages, while MultiWOZ shows 78.2% accuracy with 6.8% false memory rate under persistent retention. This work introduces an adaptive budgeted forgetting framework that regulates memory through relevanceguided scoring and bounded optimization. The approach integrates recency, frequency, and semantic alignment to maintain stability under constrained context. Comparative analysis demonstrates improved long-horizon F1 beyond 0.583 baseline levels, higher retention consistency, and reduced false memory behavior without increasing context usage. These findings confirm that structured forgetting preserves reasoning performance while preventing unbounded memory growth in extended conversational settings.

Autoregressive video diffusion models have proved effective for world modeling and interactive scene generation, with Minecraft gameplay as a representative application. To faithfully simulate play, a model must generate natural content while exploring new scenes and preserve spatial consistency when revisiting explored areas. Under limited computation budgets, it must compress and exploit historical cues within a finite context window, which exposes a trade-off: Temporal-only memory lacks long-term spatial consistency, whereas adding spatial memory strengthens consistency but may degrade new scene generation quality when the model over-relies on insufficient spatial context. We present Memory Forcing, a learning framework that pairs training protocols with a geometry-indexed spatial memory. Hybrid Training exposes distinct gameplay regimes, guiding the model to rely on temporal memory during exploration and incorporate spatial memory for revisits. Chained Forward Training extends autoregressive training with model rollouts, where chained predictions create larger pose variations and encourage reliance on spatial memory for maintaining consistency. Point-to-Frame Retrieval efficiently retrieves history by mapping currently visible points to their source frames, while Incremental 3D Reconstruction maintains and updates an explicit 3D cache. Extensive experiments demonstrate that Memory Forcing achieves superior long-term spatial consistency and generative quality across diverse environments, while maintaining computational efficiency for extended sequences.

Training large language models (LLMs) for complex reasoning via reinforcement learning requires reward labels that specify whether the generated rollouts are correct. However, obtaining reward labels at scale often requires expensive human labeling or time-consuming verification procedures; for instance, evaluating mathematical proofs demands expert review, while open-ended question answering lacks definitive ground truth. When reward labels are limited, the effectiveness of reinforcement learning fine-tuning is constrained by the scarcity of reward labels. We introduce MemReward, a graph-based experience memory framework: an initial LLM policy generates rollouts for each query, each comprising a thinking process and a final answer, and these rollouts are stored as experience memory. Queries, thinking processes, and answers form nodes in a heterogeneous graph with similarity and structural edges; a GNN trained on labeled nodes propagates rewards to unlabeled rollouts during online optimization. Experiments on Qwen2.5-3B and 1.5B across mathematics, question answering, and code generation demonstrate that MemReward, with only 20% labels, achieves 97.3% of Oracle performance on 3B and 96.6% on 1.5B, surpassing Oracle on out-of-domain tasks. Performance scales smoothly with label budget, reaching 99.4% of Oracle at 70% labels.

A primary cost driver for training large models is wall-clock training time. We show that popular time estimates based on FLOPs are poor estimates, and construct a more accurate proxy based on memory copies. We show that with some simple accounting, we can estimate the training speed of a transformer model from its hyperparameters. Combined with a scaling law curve like Chinchilla, this lets us estimate the final loss of the model. We fit our estimate to real data with a linear regression, and apply the result to rewrite Chinchilla in terms of a model's estimated training time as opposed to the amount of training data. This gives an expression for the loss in terms of the model's hyperparameters alone. We show that this expression is accurate across a wide range of model hyperparameter values, enabling us to analytically make architectural decisions and train models more efficiently.

Key-value (KV) caching is widely used to accelerate transformer inference, but its memory cost grows linearly with input length, limiting long-context deployment. Existing token eviction methods reduce memory by discarding less important tokens, which can be viewed as a coarse form of dimensionality reduction that assigns each token either zero or full dimension. We propose MixedDimKV, a mixed-dimension KV cache compression method that allocates dimensions to tokens at a more granular level, and MixedDimKV-H, which further integrates head-level importance information. Experiments on long-context benchmarks show that MixedDimKV outperforms prior KV cache compression methods that do not rely on head-level importance profiling. When equipped with the same head-level importance information, MixedDimKV-H consistently outperforms HeadKV. Notably, our approach achieves comparable performance to full attention on LongBench with only 6.25% of the KV cache. Furthermore, in the Needle-in-a-Haystack test, our solution maintains 100% accuracy at a 50K context length while using as little as 0.26% of the cache.

Multi-agent LLM systems on edge devices face a memory management problem: device RAM is too small to hold every agent's KV cache simultaneously. On Apple M4 Pro with 10.2 GB of cache budget, only 3 agents fit at 8K context in FP16. A 10-agent workflow must constantly evict and reload caches. Without persistence, every eviction forces a full re-prefill through the model -- 15.7 seconds per agent at 4K context. We address this by persisting each agent's KV cache to disk in 4-bit quantized format and reloading it directly into the attention layer, eliminating redundant O(n) prefill computation via direct cache restoration. The system comprises three components: a block pool providing per-agent isolated Q4 KV caches in safetensors format, a BatchQuantizedKVCache for concurrent inference over multiple agents' quantized caches, and cross-phase context injection that accumulates attention state across conversation phases without re-computation. Evaluated on three architectures (Gemma 3 12B, dense GQA, 48 layers; DeepSeek-Coder-V2-Lite 16B, MoE MLA, 27 layers; Llama 3.1 8B, dense GQA, 32 layers), cache restoration reduces time-to-first-token by up to 136x (Gemma: 22--136x at 4K--32K; DeepSeek: 11--76x at 4K--32K; Llama: 24--111x at 4K--16K; 3--10x at 1K). Q4 quantization fits 4x more agent contexts into fixed device memory than FP16. Perplexity measured with actual Q4 KV caches shows -0.7% for Gemma, +2.8% for Llama, and +3.0% for DeepSeek. Open-source at https://github.com/yshk-mxim/agent-memory

Large language models (LLMs) are increasingly deployed in multi-turn dialogue settings, but their behavior is still bottlenecked by fixed context windows and naive memory strategies. Replaying the full conversation at every turn is simple but expensive, while static summarization or recency-only heuristics often erase safety-critical user details. We present Adaptive Focus Memory (AFM), a dynamic context manager that assigns each past message one of three fidelity levels -- FULL, COMPRESSED, or PLACEHOLDER -- based on semantic similarity to the current query, half-life recency weighting, and importance classification. AFM packs messages chronologically under a strict token budget, preferring high fidelity for the most relevant turns while aiming to preserve a cheap trace of the dialogue. In a safety-oriented benchmark involving a user with a severe peanut allergy planning a trip to Thailand, AFM retains the allergy across both short and medium-length conversations, matches the safety performance of naive replay, and cuts average token usage by 66% relative to a replay baseline. We release a modular Python implementation of AFM designed for OpenAI-compatible APIs and offline operation, enabling practitioners to reduce inference cost without sacrificing safety or factual continuity in the evaluated scenario.

Streaming video understanding requires models to robustly encode, store, and retrieve information from a continuous video stream to support accurate video question answering (VQA). Existing state-of-the-art approaches rely on key-value caching to accumulate frame-level information over time, but use a limited number of tokens per frame, leading to the loss of fine-grained visual details. In this work, we propose scaling the token budget to enable more granular spatiotemporal understanding and reasoning. First, we find that current methods are ill-equipped to handle dense streams: their feature encoding causes query-frame similarity scores to increase over time, biasing retrieval toward later frames. To address this, we introduce an adaptive selection strategy that reduces token redundancy while preserving local spatiotemporal information. We further propose a training-free retrieval mixture-of-experts that leverages external models to better identify relevant frames. Our method, MemStream, achieves +8.0% on CG-Bench, +8.5% on LVBench, and +2.4% on VideoMME (Long) over ReKV with Qwen2.5-VL-7B.

Diffusion Large Language Models (dLLMs) have emerged as a promising alternative to Autoregressive Models (ARMs), utilizing parallel decoding to overcome sequential bottlenecks. However, existing research focuses primarily on kernel-level optimizations, lacking a holistic serving framework that addresses the unique memory dynamics of diffusion processes in production. We identify a critical "memory footprint crisis" specific to dLLMs, driven by monolithic logit tensors and the severe resource oscillation between compute-bound "Refresh" phases and bandwidth-bound "Reuse" phases. To bridge this gap, we present dLLM-Serve, an efficient dLLM serving system that co-optimizes memory footprint, computational scheduling, and generation quality. dLLM-Serve introduces Logit-Aware Activation Budgeting to decompose transient tensor peaks, a Phase-Multiplexed Scheduler to interleave heterogeneous request phases, and Head-Centric Sparse Attention to decouple logical sparsity from physical storage. We evaluate dLLM-Serve on diverse workloads (LiveBench, Burst, OSC) and GPUs (RTX 4090, L40S). Relative to the state-of-the-art baseline, dLLM-Serve improves throughput by 1.61times-1.81times on the consumer-grade RTX 4090 and 1.60times-1.74times on the server-grade NVIDIA L40S, while reducing tail latency by nearly 4times under heavy contention. dLLM-Serve establishes the first blueprint for scalable dLLM inference, converting theoretical algorithmic sparsity into tangible wall-clock acceleration across heterogeneous hardware.

Large Language Model (LLM) serving faces a fundamental tension between stringent latency Service Level Objectives (SLOs) and limited GPU memory capacity. When high request rates exhaust the KV cache budget, existing LLM inference systems often suffer severe head-of-line (HOL) blocking. While prior work explored PCIe-based offloading, these approaches cannot sustain responsiveness under high request rates, often failing to meet tight Time-To-First-Token (TTFT) and Time-Between-Tokens (TBT) SLOs. We present SuperInfer, a high-performance LLM inference system designed for emerging Superchips (e.g., NVIDIA GH200) with tightly coupled GPU-CPU architecture via NVLink-C2C. SuperInfer introduces RotaSched, the first proactive, SLO-aware rotary scheduler that rotates requests to maintain responsiveness on Superchips, and DuplexKV, an optimized rotation engine that enables full-duplex transfer over NVLink-C2C. Evaluations on GH200 using various models and datasets show that SuperInfer improves TTFT SLO attainment rates by up to 74.7% while maintaining comparable TBT and throughput compared to state-of-the-art systems, demonstrating that SLO-aware scheduling and memory co-design unlocks the full potential of Superchips for responsive LLM serving.

Federated learning (FL) promotes decentralized training while prioritizing data confidentiality. However, its application on resource-constrained devices is challenging due to the high demand for computation and memory resources to train deep learning models. Neural network pruning techniques, such as dynamic pruning, could enhance model efficiency, but directly adopting them in FL still poses substantial challenges, including post-pruning performance degradation, high activation memory usage, etc. To address these challenges, we propose FedMef, a novel and memory-efficient federated dynamic pruning framework. FedMef comprises two key components. First, we introduce the budget-aware extrusion that maintains pruning efficiency while preserving post-pruning performance by salvaging crucial information from parameters marked for pruning within a given budget. Second, we propose scaled activation pruning to effectively reduce activation memory footprints, which is particularly beneficial for deploying FL to memory-limited devices. Extensive experiments demonstrate the effectiveness of our proposed FedMef. In particular, it achieves a significant reduction of 28.5% in memory footprint compared to state-of-the-art methods while obtaining superior accuracy.

This paper introduces a structured memory which can be easily integrated into a neural network. The memory is very large by design and significantly increases the capacity of the architecture, by up to a billion parameters with a negligible computational overhead. Its design and access pattern is based on product keys, which enable fast and exact nearest neighbor search. The ability to increase the number of parameters while keeping the same computational budget lets the overall system strike a better trade-off between prediction accuracy and computation efficiency both at training and test time. This memory layer allows us to tackle very large scale language modeling tasks. In our experiments we consider a dataset with up to 30 billion words, and we plug our memory layer in a state-of-the-art transformer-based architecture. In particular, we found that a memory augmented model with only 12 layers outperforms a baseline transformer model with 24 layers, while being twice faster at inference time. We release our code for reproducibility purposes.

LLM-based agents show strong potential for long-horizon reasoning, yet their context size is limited by deployment factors (e.g., memory, latency, and cost), yielding a constrained context budget. As interaction histories grow, this induces a trade-off between retaining past information and staying within the context limit. To address this challenge, we propose Budget-Aware Context Management (BACM), which formulates context management as a sequential decision problem with a context budget constraint. It enables agents to assess the available budget before incorporating new observations and decide when and how much of the interaction history to compress. We further develop BACM-RL, an end-to-end curriculum-based reinforcement learning approach that learns compression strategies under varying context budgets. Experiments on compositional multi-objective QA and long-horizon web browsing benchmarks show that BACM-RL consistently outperforms prior methods across model scales and task complexities, achieving over 1.6times gains over strong baselines in high-complexity settings, while maintaining strong advantages as budgets shrink, where most methods exhibit a downward performance trend.

Memory layers use a trainable key-value lookup mechanism to add extra parameters to a model without increasing FLOPs. Conceptually, sparsely activated memory layers complement compute-heavy dense feed-forward layers, providing dedicated capacity to store and retrieve information cheaply. This work takes memory layers beyond proof-of-concept, proving their utility at contemporary scale. On downstream tasks, language models augmented with our improved memory layer outperform dense models with more than twice the computation budget, as well as mixture-of-expert models when matched for both compute and parameters. We find gains are especially pronounced for factual tasks. We provide a fully parallelizable memory layer implementation, demonstrating scaling laws with up to 128B memory parameters, pretrained to 1 trillion tokens, comparing to base models with up to 8B parameters.

Key-Value (KV) cache plays a crucial role in accelerating inference in large language models (LLMs) by storing intermediate attention states and avoiding redundant computation during autoregressive generation. However, its memory footprint scales linearly with sequence length, often leading to severe memory bottlenecks on resource-constrained hardware. Prior work has explored offloading KV cache to the CPU while retaining only a subset on the GPU, but these approaches often rely on imprecise token selection and suffer performance degradation in long-generation tasks such as chain-of-thought reasoning. In this paper, we propose a novel KV cache management strategy, IceCache, which integrates semantic token clustering with PagedAttention. By organizing semantically related tokens into contiguous memory regions managed by a hierarchical, dynamically updatable data structure, our method enables more efficient token selection and better utilization of memory bandwidth during CPU-GPU transfers. Experimental results on LongBench show that, with a 256-token budget, IceCache maintains 99% of the original accuracy achieved by the full KV cache model. Moreover, compared to other offloading-based methods, IceCache attains competitive or even superior latency and accuracy while using only 25% of the KV cache token budget, demonstrating its effectiveness in long-sequence scenarios. The code is available on our project website at https://yuzhenmao.github.io/IceCache/.

Optimizing the Key-Value (KV) cache of the Large Language Model (LLM) has been considered critical to saving the cost of inference. Most of the existing KV-cache compression algorithms attempted to sparsify the sequence of tokens by taking advantage of the different importance of tokens. In this work, we found that by identifying the importance of attention layers, we could optimize the KV-cache jointly from two dimensions. Based on our observations regarding layer-wise importance in inference, we propose SqueezeAttention to precisely optimize the allocation of KV-cache budget among layers on-the-fly and then incorporate three representative token sparsification algorithms to compress the KV-cache for each layer with its very own budget. By optimizing the KV-cache from both sequence's and layer's dimensions, SqueezeAttention achieves around 30% to 70% of the memory reductions and up to 2.2 times of throughput improvements in a wide range of LLMs and benchmarks. The code is available at https://github.com/hetailang/SqueezeAttention.

Recent years have witnessed increasing interest in extending large language models into agentic systems. While the effectiveness of agents has continued to improve, efficiency, which is crucial for real-world deployment, has often been overlooked. This paper therefore investigates efficiency from three core components of agents: memory, tool learning, and planning, considering costs such as latency, tokens, steps, etc. Aimed at conducting comprehensive research addressing the efficiency of the agentic system itself, we review a broad range of recent approaches that differ in implementation yet frequently converge on shared high-level principles including but not limited to bounding context via compression and management, designing reinforcement learning rewards to minimize tool invocation, and employing controlled search mechanisms to enhance efficiency, which we discuss in detail. Accordingly, we characterize efficiency in two complementary ways: comparing effectiveness under a fixed cost budget, and comparing cost at a comparable level of effectiveness. This trade-off can also be viewed through the Pareto frontier between effectiveness and cost. From this perspective, we also examine efficiency oriented benchmarks by summarizing evaluation protocols for these components and consolidating commonly reported efficiency metrics from both benchmark and methodological studies. Moreover, we discuss the key challenges and future directions, with the goal of providing promising insights.

Extensive system scales (i.e. thousands of GPU/TPUs) and prolonged training periods (i.e. months of pretraining) significantly escalate the probability of failures when training large language models (LLMs). Thus, efficient and reliable fault-tolerance methods are in urgent need. Checkpointing is the primary fault-tolerance method to periodically save parameter snapshots from GPU memory to disks via CPU memory. In this paper, we identify the frequency of existing checkpoint-based fault-tolerance being significantly limited by the storage I/O overheads, which results in hefty re-training costs on restarting from the nearest checkpoint. In response to this gap, we introduce an in-memory fault-tolerance framework for large-scale LLM pretraining. The framework boosts the efficiency and reliability of fault tolerance from three aspects: (1) Reduced Data Transfer and I/O: By asynchronously caching parameters, i.e., sharded model parameters, optimizer states, and RNG states, to CPU volatile memory, Our framework significantly reduces communication costs and bypasses checkpoint I/O. (2) Enhanced System Reliability: Our framework enhances parameter protection with a two-layer hierarchy: snapshot management processes (SMPs) safeguard against software failures, together with Erasure Coding (EC) protecting against node failures. This double-layered protection greatly improves the survival probability of the parameters compared to existing checkpointing methods. (3) Improved Snapshotting Frequency: Our framework achieves more frequent snapshotting compared with asynchronous checkpointing optimizations under the same saving time budget, which improves the fault tolerance efficiency. Empirical results demonstrate that Our framework minimizes the overhead of fault tolerance of LLM pretraining by effectively leveraging redundant CPU resources.

It is widely acknowledged that the performance of Transformer models is exponentially related to their number of parameters and computational complexity. While approaches like Mixture of Experts (MoE) decouple parameter count from computational complexity, they still face challenges in inference due to high memory access costs. This work introduces UltraMem, incorporating large-scale, ultra-sparse memory layer to address these limitations. Our approach significantly reduces inference latency while maintaining model performance. We also investigate the scaling laws of this new architecture, demonstrating that it not only exhibits favorable scaling properties but outperforms traditional models. In our experiments, we train networks with up to 20 million memory slots. The results show that our method achieves state-of-the-art inference speed and model performance within a given computational budget.

Large language model (LLM) agents increasingly operate in settings where a single context window is far too small to capture what has happened, what was learned, and what should not be repeated. Memory -- the ability to persist, organize, and selectively recall information across interactions -- is what turns a stateless text generator into a genuinely adaptive agent. This survey offers a structured account of how memory is designed, implemented, and evaluated in modern LLM-based agents, covering work from 2022 through early 2026. We formalize agent memory as a write--manage--read loop tightly coupled with perception and action, then introduce a three-dimensional taxonomy spanning temporal scope, representational substrate, and control policy. Five mechanism families are examined in depth: context-resident compression, retrieval-augmented stores, reflective self-improvement, hierarchical virtual context, and policy-learned management. On the evaluation side, we trace the shift from static recall benchmarks to multi-session agentic tests that interleave memory with decision-making, analyzing four recent benchmarks that expose stubborn gaps in current systems. We also survey applications where memory is the differentiating factor -- personal assistants, coding agents, open-world games, scientific reasoning, and multi-agent teamwork -- and address the engineering realities of write-path filtering, contradiction handling, latency budgets, and privacy governance. The paper closes with open challenges: continual consolidation, causally grounded retrieval, trustworthy reflection, learned forgetting, and multimodal embodied memory.

Semantic top-K selection with cross-encoder rerankers underpins of on-device AI services, such as retrieval-augmented generation, agent memory, and personalized recommendation. However, its latency and memory demands dominate end-to-end budgets on edge hardware. Revisiting the objective of top-K selection, we reveal that only relative rankings matter, not exact per-candidate scores. We further observe sequence-level sparsity: relative rankings stabilize early in intermediate layers, allowing pruning opportunities prior to completing full inference. Building on this insight, we propose monolithic forwarding and develop a training-free inference system, GRATING. By maintaining a global view of all candidates, it reduces latency through progressive cluster pruning. It also bounds peak memory usage by strategically overlapping I/O with computation via dual-layer sliding window and chunked execution. We evaluate GRATING against state-of-the-art baselines on rerankers from 0.6B to 8B parameters across Apple M2 and RTX 5070. GRATING consistently reduces latency by up to 89.0% and peak memory by up to 94.9% in microbenchmarks, without any loss in precision. Across three real-world on-device AI applications, GRATING lowers latency by 11.6%-51.0% and peak memory by 18.6%-77.8%, demonstrating substantial improvements in efficiency and deployability.

Speculative decoding accelerates Large Language Model (LLM) inference by verifying multiple drafted tokens in parallel. However, for Mixture-of-Experts (MoE) models, this parallelism introduces a severe bottleneck: large draft trees activate many unique experts, significantly increasing memory pressure and diminishing speedups from speculative decoding relative to autoregressive decoding. Prior methods reduce speculation depth when MoE verification becomes expensive. We propose MoE-Spec, a training-free verification-time expert budgeting method that decouples speculation depth from memory cost by enforcing a fixed expert capacity limit at each layer, loading only the experts that contribute most to verification and dropping the long tail of rarely used experts that drive bandwidth overhead. Experiments across multiple model scales and datasets show that this method yields 10--30\% higher throughput than state-of-the-art speculative decoding baselines (EAGLE-3) at comparable quality, with flexibility to trade accuracy for further latency reductions through tighter budgets.

Training LLMs presents significant memory challenges due to growing size of data, weights, and optimizer states. Techniques such as data and model parallelism, gradient checkpointing, and offloading strategies address this issue but are often infeasible due to hardware constraints. To mitigate memory usage, alternative methods like Parameter-Efficient-Fine-Tuning (PEFT) and GaLore approximate weights or optimizer states. PEFT methods, such as LoRA, have gained popularity for fine-tuning LLMs, though they require a full-rank warm start. In contrast, GaLore allows full-parameter learning while being more memory-efficient. This work introduces Natural GaLore, a simple drop in replacement for AdamW, which efficiently applies the inverse Empirical Fisher Information Matrix to low-rank gradients using Woodbury's Identity. We demonstrate that incorporating second-order information speeds up optimization significantly, especially when the iteration budget is limited. Empirical pretraining on 60M, 130M, 350M, and 1.1B parameter Llama models on C4 data demonstrate significantly lower perplexity over GaLore without additional memory overhead. By fine-tuning RoBERTa on the GLUE benchmark using Natural GaLore, we demonstrate significant reduction in gap 86.05% vs 86.28% for full-finetuning. Furthermore, fine-tuning the TinyLlama 1.1B model for function calling using the TinyAgent framework shows that Natural GaLore achieving 83.09% accuracy on the TinyAgent dataset, significantly outperforms 16-bit LoRA at 80.06% and even surpasses GPT4-Turbo by 4%, all while using 30% less memory. All code to reproduce the results are available at: https://github.com/selfsupervised-ai/Natural-GaLore.git

Complex reasoning in tool-augmented agent frameworks is inherently long-horizon, causing reasoning traces and transient tool artifacts to accumulate and strain the bounded working context of large language models. Without explicit memory mechanisms, such accumulation disrupts logical continuity and undermines task alignment. This positions memory not as an auxiliary efficiency concern, but as a core component for sustaining coherent, goal-directed reasoning over long horizons. We propose MemoBrain, an executive memory model for tool-augmented agents that constructs a dependency-aware memory over reasoning steps, capturing salient intermediate states and their logical relations. Operating as a co-pilot alongside the reasoning agent, MemoBrain organizes reasoning progress without blocking execution and actively manages the working context. Specifically, it prunes invalid steps, folds completed sub-trajectories, and preserves a compact, high-salience reasoning backbone under a fixed context budget. Together, these mechanisms enable explicit cognitive control over reasoning trajectories rather than passive context accumulation. We evaluate MemoBrain on challenging long-horizon benchmarks, including GAIA, WebWalker, and BrowseComp-Plus, demonstrating consistent improvements over strong baselines.

Retrieval-augmented language models such as Fusion-in-Decoder are powerful, setting the state of the art on a variety of knowledge-intensive tasks. However, they are also expensive, due to the need to encode a large number of retrieved passages. Some work avoids this cost by pre-encoding a text corpus into a memory and retrieving dense representations directly. However, pre-encoding memory incurs a severe quality penalty as the memory representations are not conditioned on the current input. We propose LUMEN, a hybrid between these two extremes, pre-computing the majority of the retrieval representation and completing the encoding on the fly using a live encoder that is conditioned on the question and fine-tuned for the task. We show that LUMEN significantly outperforms pure memory on multiple question-answering tasks while being much cheaper than FiD, and outperforms both for any given compute budget. Moreover, the advantage of LUMEN over FiD increases with model size.

Large language models (LLMs) excel at complex reasoning, yet their efficiency is limited by the surging cognitive overhead of long thought traces. In this paper, we propose LightThinker, a method that enables LLMs to dynamically compress intermediate thoughts into compact semantic representations. However, static compression often struggles with complex reasoning where the irreversible loss of intermediate details can lead to logical bottlenecks. To address this, we evolve the framework into LightThinker++, introducing Explicit Adaptive Memory Management. This paradigm shifts to behavioral-level management by incorporating explicit memory primitives, supported by a specialized trajectory synthesis pipeline to train purposeful memory scheduling. Extensive experiments demonstrate the framework's versatility across three dimensions. (1) LightThinker reduces peak token usage by 70% and inference time by 26% with minimal accuracy loss. (2) In standard reasoning, LightThinker++ slashes peak token usage by 69.9% while yielding a +2.42% accuracy gain under the same context budget for maximum performance. (3) Most notably, in long-horizon agentic tasks, it maintains a stable footprint beyond 80 rounds (a 60%-70% reduction), achieving an average performance gain of 14.8% across different complex scenarios. Overall, our work provides a scalable direction for sustaining deep LLM reasoning over extended horizons with minimal overhead.

Long-horizon applications increasingly require large language models (LLMs) to answer queries when relevant evidence is sparse and dispersed across very long contexts. Existing memory systems largely follow two paradigms: explicit structured memories offer interpretability but often become brittle under long-context overload, while latent memory mechanisms are efficient and stable yet difficult to inspect. We propose LatentGraphMem, a memory framework that combines implicit graph memory with explicit subgraph retrieval. LatentGraphMem stores a graph-structured memory in latent space for stability and efficiency, and exposes a task-specific subgraph retrieval interface that returns a compact symbolic subgraph under a fixed budget for downstream reasoning and human inspection. During training, an explicit graph view is materialized to interface with a frozen reasoner for question-answering supervision. At inference time, retrieval is performed in latent space and only the retrieved subgraph is externalized. Experiments on long-horizon benchmarks across multiple model scales show that LatentGraphMem consistently outperforms representative explicit-graph and latent-memory baselines, while enabling parameter-efficient adaptation and flexible scaling to larger reasoners without introducing large symbolic artifacts.

Effective long-term memory in conversational AI requires synthesizing information across multiple sessions. However, current systems place excessive reasoning burden on response generation, making performance significantly dependent on model sizes. We introduce PREMem (Pre-storage Reasoning for Episodic Memory), a novel approach that shifts complex reasoning processes from inference to memory construction. PREMem extracts fine-grained memory fragments categorized into factual, experiential, and subjective information; it then establishes explicit relationships between memory items across sessions, capturing evolution patterns like extensions, transformations, and implications. By performing this reasoning during pre-storage rather than when generating a response, PREMem creates enriched representations while reducing computational demands during interactions. Experiments show significant performance improvements across all model sizes, with smaller models achieving results comparable to much larger baselines while maintaining effectiveness even with constrained token budgets. Code and dataset are available at https://github.com/sangyeop-kim/PREMem.

Large language model (LLM) agents are fundamentally bottlenecked by finite context windows on long-horizon tasks. As trajectories grow, retaining tool outputs and intermediate reasoning in-context quickly becomes infeasible: the working context becomes prohibitively long, eventually exceeds the context budget, and makes distant evidence harder to use even when it is still present. Existing solutions typically shorten context through truncation or running summaries, but these methods are fundamentally lossy because they compress or discard past evidence itself. We introduce Memex, an indexed experience memory mechanism that instead compresses context without discarding evidence. Memex maintains a compact working context consisting of concise structured summaries and stable indices, while storing full-fidelity underlying interactions in an external experience database under those indices. The agent can then decide when to dereference an index and recover the exact past evidence needed for the current subgoal. We optimize both write and read behaviors with our reinforcement learning framework MemexRL, using reward shaping tailored to indexed memory usage under a context budget, so the agent learns what to summarize, what to archive, how to index it, and when to retrieve it. This yields a substantially less lossy form of long-horizon memory than summary-only approaches. We further provide a theoretical analysis showing the potential of the Memex loop to preserve decision quality with bounded dereferencing while keeping effective in-context computation bounded as history grows. Empirically, on challenging long-horizon tasks, Memex agent trained with MemexRL improves task success while using a significantly smaller working context.

Pretraining large language models (LLMs) typically requires centralized clusters with thousands of high-memory GPUs (e.g., H100/A100). Recent decentralized training methods reduce communication overhead by employing federated optimization; however, they still need to train the entire model on each node, remaining constrained by GPU memory limitations. In this work, we propose SParse Expert Synchronization (SPES), a memory-efficient decentralized framework for pretraining mixture-of-experts (MoE) LLMs. SPES trains only a subset of experts per node, substantially lowering the memory footprint. Each node updates its local experts and periodically synchronizes with other nodes, eliminating full-parameter transmission while ensuring efficient knowledge sharing. To accelerate convergence, we introduce an expert-merging warm-up strategy, where experts exchange knowledge early in training, to rapidly establish foundational capabilities. With SPES, we train a 2B-parameter MoE LLM using 16 standalone 48GB GPUs over internet connections, which achieves competitive performance with centrally trained LLMs under similar computational budgets. We further demonstrate scalability by training a 7B model from scratch and a 9B model upcycled from a dense checkpoint, both of which match prior centralized baselines. Our code is available at https://github.com/zjr2000/SPES.

Autoregressive sequence modeling stands as the cornerstone of modern Generative AI, powering results across diverse modalities ranging from text generation to image generation. However, a fundamental limitation of this paradigm is the rigid structural coupling of model capacity to computational cost: expanding a model's parametric memory -- its repository of factual knowledge or visual patterns -- traditionally requires deepening or widening the network, which incurs a proportional rise in active FLOPs. In this work, we introduce MoVE (Mixture of Value Embeddings), a mechanism that breaks this coupling and establishes a new axis for scaling capacity. MoVE decouples memory from compute by introducing a global bank of learnable value embeddings shared across all attention layers. For every step in the sequence, the model employs a differentiable soft gating mechanism to dynamically mix retrieved concepts from this bank into the standard value projection. This architecture allows parametric memory to be scaled independently of network depth by simply increasing the number of embedding slots. We validate MoVE through strictly controlled experiments on two representative applications of autoregressive modeling: Text Generation and Image Generation. In both domains, MoVE yields consistent performance improvements over standard and layer-wise memory baselines, enabling the construction of "memory-dense" models that achieve lower perplexity and higher fidelity than their dense counterparts at comparable compute budgets.

We propose Infinite-World, a robust interactive world model capable of maintaining coherent visual memory over 1000+ frames in complex real-world environments. While existing world models can be efficiently optimized on synthetic data with perfect ground-truth, they lack an effective training paradigm for real-world videos due to noisy pose estimations and the scarcity of viewpoint revisits. To bridge this gap, we first introduce a Hierarchical Pose-free Memory Compressor (HPMC) that recursively distills historical latents into a fixed-budget representation. By jointly optimizing the compressor with the generative backbone, HPMC enables the model to autonomously anchor generations in the distant past with bounded computational cost, eliminating the need for explicit geometric priors. Second, we propose an Uncertainty-aware Action Labeling module that discretizes continuous motion into a tri-state logic. This strategy maximizes the utilization of raw video data while shielding the deterministic action space from being corrupted by noisy trajectories, ensuring robust action-response learning. Furthermore, guided by insights from a pilot toy study, we employ a Revisit-Dense Finetuning Strategy using a compact, 30-minute dataset to efficiently activate the model's long-range loop-closure capabilities. Extensive experiments, including objective metrics and user studies, demonstrate that Infinite-World achieves superior performance in visual quality, action controllability, and spatial consistency.

Multimodal Large Language Models have achieved significant success in offline video understanding, yet their application to streaming videos is severely limited by the linear explosion of visual tokens, which often leads to Out-of-Memory (OOM) errors or catastrophic forgetting. Existing visual retention and memory management methods typically rely on uniform sampling, low-level physical metrics, or passive cache eviction. However, these strategies often lack intrinsic semantic awareness, potentially disrupting contextual coherence and blurring transient yet critical semantic transitions. To address these limitations, we propose CurveStream, a training-free, curvature-aware hierarchical visual memory management framework. Our approach is motivated by the key observation that high-curvature regions along continuous feature trajectories closely align with critical global semantic transitions. Based on this geometric insight, CurveStream evaluates real-time semantic intensity via a Curvature Score and integrates an online K-Sigma dynamic threshold to adaptively route frames into clear and fuzzy memory states under a strict token budget. Evaluations across diverse temporal scales confirm that this lightweight framework, CurveStream, consistently yields absolute performance gains of over 10% (e.g., 10.69% on StreamingBench and 13.58% on OVOBench) over respective baselines, establishing new state-of-the-art results for streaming video perception.The code will be released at https://github.com/streamingvideos/CurveStream.

Multi-agent large language model (LLM) systems have shown strong potential in complex reasoning and collaborative decision-making tasks. However, most existing coordination schemes rely on static or full-context routing strategies, which lead to excessive token consumption, redundant memory exposure, and limited adaptability across interaction rounds. We introduce RCR-Router, a modular and role-aware context routing framework designed to enable efficient, adaptive collaboration in multi-agent LLMs. To our knowledge, this is the first routing approach that dynamically selects semantically relevant memory subsets for each agent based on its role and task stage, while adhering to a strict token budget. A lightweight scoring policy guides memory selection, and agent outputs are iteratively integrated into a shared memory store to facilitate progressive context refinement. To better evaluate model behavior, we further propose an Answer Quality Score metric that captures LLM-generated explanations beyond standard QA accuracy. Experiments on three multi-hop QA benchmarks -- HotPotQA, MuSiQue, and 2WikiMultihop -- demonstrate that RCR-Router reduces token usage (up to 30%) while improving or maintaining answer quality. These results highlight the importance of structured memory routing and output-aware evaluation in advancing scalable multi-agent LLM systems.

Large Language models (LLMs) have become a research hotspot. To accelerate the inference of LLMs, storing computed caches in memory has become the standard technique. However, as the inference length increases, growing KV caches might lead to out-of-memory issues. Many existing methods address this issue through KV cache compression, primarily by preserving key tokens throughout all layers to reduce information loss. Most of them allocate a uniform budget size for each layer to retain. However, we observe that the minimum budget sizes needed to retain essential information vary across layers and models based on the perspectives of attention and hidden state output. Building on this observation, this paper proposes a simple yet effective KV cache compression method that leverages layer uncertainty to allocate budget size for each layer. Experimental results show that the proposed method can reduce memory usage of the KV caches to only sim20\% when compared to Full KV inference while achieving nearly lossless performance.

Automating AI research differs from general software engineering due to computationally expensive evaluation (e.g., model training) and opaque performance attribution. Current LLM-based agents struggle here, often generating monolithic scripts that ignore execution costs and causal factors. We introduce MARS (Modular Agent with Reflective Search), a framework optimized for autonomous AI research. MARS relies on three pillars: (1) Budget-Aware Planning via cost-constrained Monte Carlo Tree Search (MCTS) to explicitly balance performance with execution expense; (2) Modular Construction, employing a "Design-Decompose-Implement" pipeline to manage complex research repositories; and (3) Comparative Reflective Memory, which addresses credit assignment by analyzing solution differences to distill high-signal insights. MARS achieves state-of-the-art performance among open-source frameworks on MLE-Bench under comparable settings, maintaining competitiveness with the global leaderboard's top methods. Furthermore, the system exhibits qualitative "Aha!" moments, where 63% of all utilized lessons originate from cross-branch transfer, demonstrating that the agent effectively generalizes insights across search paths.

This paper introduces an information-aware quantization framework that adaptively compresses the key-value (KV) cache in large language models (LLMs). Although prior work has underscored the distinct roles of key and value cache during inference, our systematic analysis -- examining singular value distributions, spectral norms, and Frobenius norms -- reveals, for the first time, that key matrices consistently exhibit higher norm values and are more sensitive to quantization than value matrices. Furthermore, our theoretical analysis shows that matrices with higher spectral norms amplify quantization errors more significantly. Motivated by these insights, we propose a mixed-precision quantization strategy, KV-AdaQuant, which allocates more bit-width for keys and fewer for values since key matrices have higher norm values. With the same total KV bit budget, this approach effectively mitigates error propagation across transformer layers while achieving significant memory savings. Our extensive experiments on multiple LLMs (1B--70B) demonstrate that our mixed-precision quantization scheme maintains high model accuracy even under aggressive compression. For instance, using 4-bit for Key and 2-bit for Value achieves an accuracy of 75.2%, whereas reversing the assignment (2-bit for Key and 4-bit for Value) yields only 54.7% accuracy. The code is available at https://tinyurl.com/kv-adaquant

Despite rapid progress in autoregressive video diffusion, an emerging system algorithm bottleneck limits both deployability and generation capability: KV cache memory. In autoregressive video generation models, the KV cache grows with generation history and quickly dominates GPU memory, often exceeding 30 GB, preventing deployment on widely available hardware. More critically, constrained KV cache budgets restrict the effective working memory, directly degrading long horizon consistency in identity, layout, and motion. To address this challenge, we present Quant VideoGen (QVG), a training free KV cache quantization framework for autoregressive video diffusion models. QVG leverages video spatiotemporal redundancy through Semantic Aware Smoothing, producing low magnitude, quantization friendly residuals. It further introduces Progressive Residual Quantization, a coarse to fine multi stage scheme that reduces quantization error while enabling a smooth quality memory trade off. Across LongCat Video, HY WorldPlay, and Self Forcing benchmarks, QVG establishes a new Pareto frontier between quality and memory efficiency, reducing KV cache memory by up to 7.0 times with less than 4% end to end latency overhead while consistently outperforming existing baselines in generation quality.

Training a family of large language models targeting multiple scales and deployment objectives is prohibitively expensive, requiring separate training runs for each different size. Recent work on model compression through pruning and knowledge distillation has reduced this cost; however, this process still incurs hundreds of billions of tokens worth of training cost per compressed model. In this paper, we present Nemotron Elastic, a framework for building reasoning-oriented LLMs, including hybrid Mamba-Attention architectures, that embed multiple nested submodels within a single parent model, each optimized for different deployment configurations and budgets. Each of these submodels shares weights with the parent model and can be extracted zero-shot during deployment without additional training or fine-tuning. We enable this functionality through an end-to-end trained router, tightly coupled to a two-stage training curriculum designed specifically for reasoning models. We additionally introduce group-aware SSM elastification that preserves Mamba's structural constraints, heterogeneous MLP elastification, normalized MSE-based layer importance for improved depth selection, and knowledge distillation enabling simultaneous multi-budget optimization. We apply Nemotron Elastic to the Nemotron Nano V2 12B model, simultaneously producing a 9B and a 6B model using only 110B training tokens; this results in over 360x cost reduction compared to training model families from scratch, and around 7x compared to SoTA compression techniques. Each of the nested models performs on par or better than the SoTA in accuracy. Moreover, unlike other compression methods, the nested capability of our approach allows having a many-in-one reasoning model that has constant deployment memory against the number of models in the family.

This paper presents IronEngine, a general AI assistant platform organized around a unified orchestration core that connects a desktop user interface, REST and WebSocket APIs, Python clients, local and cloud model backends, persistent memory, task scheduling, reusable skills, 24-category tool execution, MCP-compatible extensibility, and hardware-facing integration. IronEngine introduces a three-phase pipeline -- Discussion (Planner--Reviewer collaboration), Model Switch (VRAM-aware transition), and Execution (tool-augmented action loop) -- that separates planning quality from execution capability. The system features a hierarchical memory architecture with multi-level consolidation, a vectorized skill repository backed by ChromaDB, an adaptive model management layer supporting 92 model profiles with VRAM-aware context budgeting, and an intelligent tool routing system with 130+ alias normalization and automatic error correction. We present experimental results on file operation benchmarks achieving 100\% task completion with a mean total time of 1541 seconds across four heterogeneous tasks, and provide detailed comparisons with representative AI assistant systems including ChatGPT, Claude Desktop, Cursor, Windsurf, and open-source agent frameworks. Without disclosing proprietary prompts or core algorithms, this paper analyzes the platform's architectural decomposition, subsystem design, experimental performance, safety boundaries, and comparative engineering advantages. The resulting study positions IronEngine as a system-oriented foundation for general-purpose personal assistants, automation frameworks, and future human-centered agent platforms.

Reconstructing 3D geometry from streaming video requires continuous inference under bounded resources. Recent geometric foundation models achieve impressive reconstruction quality through all-to-all attention, yet their quadratic cost confines them to short, offline sequences. Causal-attention variants such as StreamVGGT enable single-pass streaming but accumulate an ever-growing KV cache, exhausting GPU memory within hundreds of frames and precluding the long-horizon deployment that motivates streaming inference in the first place. We present OVGGT, a training-free framework that bounds both memory and compute to a fixed budget regardless of sequence length. Our approach combines Self-Selective Caching, which leverages FFN residual magnitudes to compress the KV cache while remaining fully compatible with FlashAttention, with Dynamic Anchor Protection, which shields coordinate-critical tokens from eviction to suppress geometric drift over extended trajectories. Extensive experiments on indoor, outdoor, and ultra-long sequence benchmarks demonstrate that OVGGT processes arbitrarily long videos within a constant VRAM envelope while achieving state-of-the-art 3D geometric accuracy.

We show that Transformer encoder architectures can be sped up, with limited accuracy costs, by replacing the self-attention sublayers with simple linear transformations that "mix" input tokens. These linear mixers, along with standard nonlinearities in feed-forward layers, prove competent at modeling semantic relationships in several text classification tasks. Most surprisingly, we find that replacing the self-attention sublayer in a Transformer encoder with a standard, unparameterized Fourier Transform achieves 92-97% of the accuracy of BERT counterparts on the GLUE benchmark, but trains 80% faster on GPUs and 70% faster on TPUs at standard 512 input lengths. At longer input lengths, our FNet model is significantly faster: when compared to the "efficient" Transformers on the Long Range Arena benchmark, FNet matches the accuracy of the most accurate models, while outpacing the fastest models across all sequence lengths on GPUs (and across relatively shorter lengths on TPUs). Finally, FNet has a light memory footprint and is particularly efficient at smaller model sizes; for a fixed speed and accuracy budget, small FNet models outperform Transformer counterparts.

Spiking Neural Networks (SNNs) are energy efficient alternatives to commonly used deep neural networks (DNNs). Through event-driven information processing, SNNs can reduce the expensive compute requirements of DNNs considerably, while achieving comparable performance. However, high inference latency is a significant hindrance to the edge deployment of deep SNNs. Computation over multiple timesteps not only increases latency as well as overall energy budget due to higher number of operations, but also incurs memory access overhead of fetching membrane potentials, both of which lessen the energy benefits of SNNs. To overcome this bottleneck and leverage the full potential of SNNs, we propose an Iterative Initialization and Retraining method for SNNs (IIR-SNN) to perform single shot inference in the temporal axis. The method starts with an SNN trained with T timesteps (T>1). Then at each stage of latency reduction, the network trained at previous stage with higher timestep is utilized as initialization for subsequent training with lower timestep. This acts as a compression method, as the network is gradually shrunk in the temporal domain. In this paper, we use direct input encoding and choose T=5, since as per literature, it is the minimum required latency to achieve satisfactory performance on ImageNet. The proposed scheme allows us to obtain SNNs with up to unit latency, requiring a single forward pass during inference. We achieve top-1 accuracy of 93.05%, 70.15% and 67.71% on CIFAR-10, CIFAR-100 and ImageNet, respectively using VGG16, with just 1 timestep. In addition, IIR-SNNs perform inference with 5-2500X reduced latency compared to other state-of-the-art SNNs, maintaining comparable or even better accuracy. Furthermore, in comparison with standard DNNs, the proposed IIR-SNNs provide25-33X higher energy efficiency, while being comparable to them in classification performance.

The costly self-attention layers in modern Transformers require memory and compute quadratic in sequence length. Existing approximation methods usually underperform and fail to obtain significant speedups in practice. Here we present SwitchHead - a novel method that reduces both compute and memory requirements and achieves wall-clock speedup, while matching the language modeling performance of baseline Transformers with the same parameter budget. SwitchHead uses Mixture-of-Experts (MoE) layers for the value and output projections and requires 4 to 8 times fewer attention matrices than standard Transformers. Our novel attention can also be combined with MoE MLP layers, resulting in an efficient fully-MoE "SwitchAll" Transformer model. Our code is public.

Existing key-value (KV) cache compression methods typically rely on heuristics, such as uniform cache allocation across layers or static eviction policies, however, they ignore the critical interplays among layer-specific feature patterns and task performance, which can lead to degraded generalization. In this paper, we propose EvolKV, an adaptive framework for layer-wise, task-driven KV cache compression that jointly optimizes the memory efficiency and task performance. By reformulating cache allocation as a multi-objective optimization problem, EvolKV leverages evolutionary search to dynamically configure layer budgets while directly maximizing downstream performance. Extensive experiments on 11 tasks demonstrate that our approach outperforms all baseline methods across a wide range of KV cache budgets on long-context tasks and surpasses heuristic baselines by up to 7 percentage points on GSM8K. Notably, EvolKV achieves superior performance over the full KV cache setting on code completion while utilizing only 1.5% of the original budget, suggesting the untapped potential in learned compression strategies for KV cache budget allocation.

Mobile displays refresh at 90-120 Hz, yet most video is encoded at 24-30 frames per second; real-time frame-rate doubling requires each synthesized frame within 33.3 ms on mobile neural processing units. We show that mainstream flow-based video frame interpolation faces three structural deployment barriers on mobile accelerators: spatial sampling operators exceed the frame budget or lack hardware support, iterative flow refinement collapses under 8-bit post-training quantization, and memory-bound operators dominate the inference graph. ANVIL addresses these barriers by reusing motion vectors already computed by the H.264 decoder to prealign input frames, removing learned optical flow, spatial sampling, and iterative accumulation from the accelerator graph. The remaining residual is refined by a convolution-dominated network whose inference graph is composed almost entirely of compute-bound operators. On a Snapdragon 8 Gen 3 device, ANVIL achieves 12.8 ms 1080p network inference in 8-bit integer precision; an open-source Android player sustains 28.4 ms median end-to-end latency per interpolated frame pair over 54,623 consecutively logged samples during 30-minute continuous playback. Per-operator causal analysis identifies quantized accumulation on recurrent flow states as a key mechanism behind integer quantization failure in iterative methods. The current design targets H.264 playback scenarios with decoder-exposed motion vectors.

We introduce ReALLM, a novel approach for compression and memory-efficient adaptation of pre-trained language models that encompasses most of the post-training quantization and fine-tuning methods for a budget of <4 bits. Pre-trained matrices are decomposed into a high-precision low-rank component and a vector-quantized latent representation (using an autoencoder). During the fine-tuning step, only the low-rank components are updated. Our results show that pre-trained matrices exhibit different patterns. ReALLM adapts the shape of the encoder (small/large embedding, high/low bit VQ, etc.) to each matrix. ReALLM proposes to represent each matrix with a small embedding on b bits and a neural decoder model D_phi with its weights on b_phi bits. The decompression of a matrix requires only one embedding and a single forward pass with the decoder. Our weight-only quantization algorithm yields the best results on language generation tasks (C4 and WikiText-2) for a budget of 3 bits without any training. With a budget of 2 bits, ReALLM achieves state-of-the art performance after fine-tuning on a small calibration dataset.

Keyword spotting (KWS) is a critical component for enabling speech based user interactions on smart devices. It requires real-time response and high accuracy for good user experience. Recently, neural networks have become an attractive choice for KWS architecture because of their superior accuracy compared to traditional speech processing algorithms. Due to its always-on nature, KWS application has highly constrained power budget and typically runs on tiny microcontrollers with limited memory and compute capability. The design of neural network architecture for KWS must consider these constraints. In this work, we perform neural network architecture evaluation and exploration for running KWS on resource-constrained microcontrollers. We train various neural network architectures for keyword spotting published in literature to compare their accuracy and memory/compute requirements. We show that it is possible to optimize these neural network architectures to fit within the memory and compute constraints of microcontrollers without sacrificing accuracy. We further explore the depthwise separable convolutional neural network (DS-CNN) and compare it against other neural network architectures. DS-CNN achieves an accuracy of 95.4%, which is ~10% higher than the DNN model with similar number of parameters.

Visual Autoregressive (VAR) models have emerged as a powerful approach for multi-modal content creation, offering high efficiency and quality across diverse multimedia applications. However, they face significant memory bottlenecks due to extensive KV cache accumulation during inference. Existing KV cache compression techniques for large language models are suboptimal for VAR models due to, as we identify in this paper, two distinct categories of attention heads in VAR models: Structural Heads, which preserve spatial coherence through diagonal attention patterns, and Contextual Heads, which maintain semantic consistency through vertical attention patterns. These differences render single-strategy KV compression techniques ineffective for VAR models. To address this, we propose HACK, a training-free Head-Aware Compression method for KV cache. HACK allocates asymmetric cache budgets and employs pattern-specific compression strategies tailored to the essential characteristics of each head category. Experiments on Infinity-2B, Infinity-8B, and VAR-d30 demonstrate its effectiveness in text-to-image and class-conditional generation tasks. HACK can hack down up to 50\% and 70\% of cache with minimal performance degradation for VAR-d30 and Infinity-8B, respectively. Even with 70\% and 90\% KV cache compression in VAR-d30 and Infinity-8B, HACK still maintains high-quality generation while reducing memory usage by 44.2\% and 58.9\%, respectively.

Data-free meta-learning (DFML) aims to enable efficient learning of new tasks by meta-learning from a collection of pre-trained models without access to the training data. Existing DFML work can only meta-learn from (i) white-box and (ii) small-scale pre-trained models (iii) with the same architecture, neglecting the more practical setting where the users only have inference access to the APIs with arbitrary model architectures and model scale inside. To solve this issue, we propose a Bi-level Data-free Meta Knowledge Distillation (BiDf-MKD) framework to transfer more general meta knowledge from a collection of black-box APIs to one single meta model. Specifically, by just querying APIs, we inverse each API to recover its training data via a zero-order gradient estimator and then perform meta-learning via a novel bi-level meta knowledge distillation structure, in which we design a boundary query set recovery technique to recover a more informative query set near the decision boundary. In addition, to encourage better generalization within the setting of limited API budgets, we propose task memory replay to diversify the underlying task distribution by covering more interpolated tasks. Extensive experiments in various real-world scenarios show the superior performance of our BiDf-MKD framework.

The Key-Value (KV) cache is crucial for efficient Large Language Models (LLMs) inference, but excessively long contexts drastically increase KV cache memory footprint. Existing KV cache compression methods typically rely on input-side attention patterns within a prompt observation window to estimate token importance during the prefill stage. They fail to preserve critical tokens for future generation since these assessments are not derived from the decoding process. Intuitively, an effective observation window should mirror the decoding-stage queries to accurately reflect which tokens the generation process will attend to. However, ground-truth decoding queries are inherently unavailable during inference. For constructing pseudo queries to approximate them, we find that positional information plays a more critical role than semantic content. Motivated by this insight, we propose decoding-aligned KV cache compression via position-aware pseudo queries (DapQ), a novel and lightweight eviction framework that leverages position-aware pseudo queries to simulate the output tokens, thereby establishing an effective observation window for importance assessment. It aligns closely with the actual generation context and enables precise token eviction. Extensive evaluations across multiple benchmarks and LLMs demonstrate that DapQ achieves superior performance, particularly under strict memory constraints (e.g., up to nearly lossless performance 99.5% on NIAH with 3% KV cache budgets).

Head detection and tracking are essential for downstream tasks, but current methods often require large computational budgets, which increase latencies and ties up resources (e.g., processors, memory, and bandwidth). To address this, we propose a framework to enhance tiny head detection and tracking by optimizing the balance between performance and efficiency. Our framework integrates (1) a cross-domain detection loss, (2) a multi-scale module, and (3) a small receptive field detection mechanism. These innovations enhance detection by bridging the gap between large and small detectors, capturing high-frequency details at multiple scales during training, and using filters with small receptive fields to detect tiny heads. Evaluations on the CroHD and CrowdHuman datasets show improved Multiple Object Tracking Accuracy (MOTA) and mean Average Precision (mAP), demonstrating the effectiveness of our approach in crowded scenes.

Large language models(LLMs) have sparked a new wave of exciting AI applications. Hosting these models at scale requires significant memory resources. One crucial memory bottleneck for the deployment stems from the context window. It is commonly recognized that model weights are memory hungry; however, the size of key-value embedding stored during the generation process (KV cache) can easily surpass the model size. The enormous size of the KV cache puts constraints on the inference batch size, which is crucial for high throughput inference workload. Inspired by an interesting observation of the attention scores, we hypothesize the persistence of importance: only pivotal tokens, which had a substantial influence at one step, will significantly influence future generations. Based on our empirical verification and theoretical analysis around this hypothesis, we propose Scissorhands, a system that maintains the memory usage of the KV cache at a fixed budget without finetuning the model. In essence, Scissorhands manages the KV cache by storing the pivotal tokens with a higher probability. We validate that Scissorhands reduces the inference memory usage of the KV cache by up to 5X without compromising model quality. We further demonstrate that Scissorhands can be combined with 4-bit quantization, traditionally used to compress model weights, to achieve up to 20X compression.

Large Language Models (LLMs) have demonstrated strong capabilities across various domains, with recent advancements in challenging reasoning tasks such as mathematics and programming. However, solving reasoning tasks often requires long decoding chains (of thoughts), which incur O(N) time and memory consumption, where N is the chain length. To mitigate O(N) time and memory consumption, existing sparsity-based algorithms propose retaining only the most critical token's intermediate data (i.e., key-value cache) and discarding the rest. However, these existing algorithms struggle with the ``impossible trinity'' of accuracy, time, and memory. For example, the state-of-the-art algorithm, Quest, achieves high accuracy with O(L) time but O(N) memory (L is the cache budget, L ll N). To address this issue, in this paper, we identify a new attention pattern during the decode stage of reasoning tasks, where milestone tokens (analogous to lemmas in mathematical proofs) emerge, are utilized, and then become unimportant afterward. Based on this pattern, we propose a new algorithm named RaaS that identifies and retains milestone tokens only until they are no longer needed, achieving high accuracy with O(L) time and O(L) memory complexity.

Recent advances in large language models (LLMs) have given rise to powerful coding agents, making it possible for code assistants to evolve into code engineers. However, existing methods still face significant challenges in achieving high-fidelity document-to-codebase synthesis--such as scientific papers to code--primarily due to a fundamental conflict between information overload and the context bottlenecks of LLMs. In this work, we introduce DeepCode, a fully autonomous framework that fundamentally addresses this challenge through principled information-flow management. By treating repository synthesis as a channel optimization problem, DeepCode seamlessly orchestrates four information operations to maximize task-relevant signals under finite context budgets: source compression via blueprint distillation, structured indexing using stateful code memory, conditional knowledge injection via retrieval-augmented generation, and closed-loop error correction. Extensive evaluations on the PaperBench benchmark demonstrate that DeepCode achieves state-of-the-art performance, decisively outperforming leading commercial agents such as Cursor and Claude Code, and crucially, surpassing PhD-level human experts from top institutes on key reproduction metrics. By systematically transforming paper specifications into production-grade implementations comparable to human expert quality, this work establishes new foundations for autonomous scientific reproduction that can accelerate research evaluation and discovery.

Transformer-based large language models (LLMs) have demonstrated impressive capabilities in a variety of natural language processing (NLP) tasks. Nonetheless, it is challenging to deploy and fine-tune LLMs on mobile edge devices with limited computing, memory, and energy budgets. In this paper, we propose Confidant, a multi-backend collaborative training framework for customizing state-of-the-art LLMs on commodity mobile devices like smartphones. Confidant partitions an LLM into several sub-models so that each fits into a mobile device's memory. A pipeline parallel training mechanism is further developed to ensure fast and efficient distributed training. In addition, we propose a novel backend scheduler to allocate different attention heads to heterogeneous compute hardware, including mobile CPU and GPUs, to maximize the compute resource utilization on each edge device. Our preliminary experimental results show that Confidant achieves at most 45.3% memory reduction and 8.03x inference speedup in practical settings.

Large Language Model (LLM) inference in production must meet stringent service-level objectives for both time-to-first-token (TTFT) and time-between-token (TBT) while maximizing throughput under fixed compute, memory, and interconnect budgets. Modern serving systems adopt stall-free scheduling techniques such as chunked prefill, which splits long prompt processing along the token dimension and interleaves prefill with ongoing decode iterations. While effective at stabilizing TBT, chunked prefill incurs substantial overhead in Mixture-of-Experts (MoE) models: redundant expert weight loads increase memory traffic by up to 39% and inflate energy consumption. We propose layered prefill, a new scheduling paradigm that treats transformer layer groups as the primary scheduling unit. By vertically partitioning the model into contiguous layer groups and interleaving prefill and decode across the groups, layered prefill sustains stall-free decoding while eliminating chunk-induced MoE weight reloads. It reduces off-chip bandwidth demand, lowering TTFT by up to 70%, End-to-End latency by 41% and per-token energy by up to 22%. Evaluations show that layered prefill consistently improves the TTFT--TBT Pareto frontier over chunked prefill, reducing expert-load traffic and energy cost while maintaining stall-free decoding. Overall, shifting the scheduling axis from tokens to layers unlocks a new operating regime for high-efficiency, energy-aware LLM serving in co-located environments.

Quantization of Deep Neural Network (DNN) activations is a commonly used technique to reduce compute and memory demands during DNN inference, which can be particularly beneficial on resource-constrained devices. To achieve high accuracy, existing methods for quantizing activations rely on complex mathematical computations or perform extensive searches for the best hyper-parameters. However, these expensive operations are impractical on devices with limited computation capabilities, memory capacities, and energy budgets. Furthermore, many existing methods do not focus on sub-6-bit (or deep) quantization. To fill these gaps, in this paper we propose DQA (Deep Quantization of DNN Activations), a new method that focuses on sub-6-bit quantization of activations and leverages simple shifting-based operations and Huffman coding to be efficient and achieve high accuracy. We evaluate DQA with 3, 4, and 5-bit quantization levels and three different DNN models for two different tasks, image classification and image segmentation, on two different datasets. DQA shows significantly better accuracy (up to 29.28%) compared to the direct quantization method and the state-of-the-art NoisyQuant for sub-6-bit quantization.