Daily Papers

Retrieval-Augmented Generation (RAG) mitigates hallucination in large language models (LLMs) by incorporating external knowledge during generation. However, the effectiveness of RAG depends not only on the design of the retriever and the capacity of the underlying model, but also on how retrieved evidence is structured and aligned with the query. Existing RAG approaches typically retrieve and concatenate unstructured text fragments as context, which often introduces redundant or weakly relevant information. This practice leads to excessive context accumulation, reduced semantic alignment, and fragmented reasoning chains, thereby degrading generation quality while increasing token consumption. To address these challenges, we propose Tri-RAG, a structured triplet-based retrieval framework that improves retrieval efficiency through reasoning-aligned context construction. Tri-RAG automatically transforms external knowledge from natural language into standardized structured triplets consisting of Condition, Proof, and Conclusion, explicitly capturing logical relations among knowledge fragments using lightweight prompt-based adaptation with frozen model parameters. Building on this representation, the triplet head Condition is treated as an explicit semantic anchor for retrieval and matching, enabling precise identification of query-relevant knowledge units without directly concatenating lengthy raw texts. As a result, Tri-RAG achieves a favorable balance between retrieval accuracy and context token efficiency. Experimental results across multiple benchmark datasets demonstrate that Tri-RAG significantly improves retrieval quality and reasoning efficiency, while producing more stable generation behavior and more efficient resource utilization in complex reasoning scenarios.

Large language models (LLMs) and their multimodal variants can now process visual inputs, including images of text. This raises an intriguing question: can we compress textual inputs by feeding them as images to reduce token usage while preserving performance? In this paper, we show that visual text representations are a practical and surprisingly effective form of input compression for decoder LLMs. We exploit the idea of rendering long text inputs as a single image and provide it directly to the model. This leads to dramatically reduced number of decoder tokens required, offering a new form of input compression. Through experiments on two distinct benchmarks RULER (long-context retrieval) and CNN/DailyMail (document summarization) we demonstrate that this text-as-image method yields substantial token savings (often nearly half) without degrading task performance.

Large Language Models demonstrate outstanding performance in many language tasks but still face fundamental challenges in managing the non-linear flow of human conversation. The prevalent approach of treating dialogue history as a flat, linear sequence is misaligned with the intrinsically hierarchical and branching structure of natural discourse, leading to inefficient context utilization and a loss of coherence during extended interactions involving topic shifts or instruction refinements. To address this limitation, we introduce Context-Agent, a novel framework that models multi-turn dialogue history as a dynamic tree structure. This approach mirrors the inherent non-linearity of conversation, enabling the model to maintain and navigate multiple dialogue branches corresponding to different topics. Furthermore, to facilitate robust evaluation, we introduce the Non-linear Task Multi-turn Dialogue (NTM) benchmark, specifically designed to assess model performance in long-horizon, non-linear scenarios. Our experiments demonstrate that Context-Agent enhances task completion rates and improves token efficiency across various LLMs, underscoring the value of structured context management for complex, dynamic dialogues. The dataset and code is available at GitHub.

Large language models have recently evolved from fluent text generation to advanced reasoning across diverse domains, giving rise to reasoning language models. Among these domains, mathematical reasoning serves as a representative benchmark as it requires precise multi-step logic and abstract reasoning, which can be generalized to other tasks. While closed-source RLMs such as GPT-o3 demonstrate impressive reasoning capabilities, their proprietary nature limits transparency and reproducibility. Although many open-source projects aim to close this gap, most of them lack sufficient openness by omitting critical resources such as datasets and detailed training configurations, which hinders reproducibility. To contribute toward greater transparency in RLM development, we introduce the MiroMind-M1 series, a set of fully open-source RLMs built on the Qwen-2.5 backbone that match or exceed the performance of existing open-source RLMs. Specifically, our models are trained in two stages: SFT on a carefully curated corpus of 719K math-reasoning problems with verified CoT trajectories, followed by RLVR on 62K challenging and verifiable problems. To enhance the robustness and efficiency of the RLVR process, we introduce Context-Aware Multi-Stage Policy Optimization, an algorithm that integrates length-progressive training with an adaptive repetition penalty to encourage context-aware RL training. Our model achieves state-of-the-art or competitive performance and superior token efficiency among Qwen-2.5-based open-source 7B and 32B models on the AIME24, AIME25, and MATH benchmarks. To facilitate reproducibility, we release the complete stack: models (MiroMind-M1-SFT-7B, MiroMind-M1-RL-7B, MiroMind-M1-RL-32B); datasets (MiroMind-M1-SFT-719K, MiroMind-M1-RL-62K); and all training and evaluation configurations. We hope these resources will support further research and foster community advancement.

Despite the significant success of large language models (LLMs), their extensive memory requirements pose challenges for deploying them in long-context token generation. The substantial memory footprint of LLM decoders arises from the necessity to store all previous tokens in the attention module, a requirement imposed by key-value (KV) caching. In this work, our focus is on developing an efficient compression technique for the KV cache. Empirical evidence indicates a significant clustering tendency within key embeddings in the attention module. Building on this key insight, we have devised a novel caching method with sublinear complexity, employing online clustering on key tokens and online ell_2 sampling on values. The result is a provably accurate and efficient attention decoding algorithm, termed SubGen. Not only does this algorithm ensure a sublinear memory footprint and sublinear time complexity, but we also establish a tight error bound for our approach. Empirical evaluations on long-context question-answering tasks demonstrate that SubGen significantly outperforms existing and state-of-the-art KV cache compression methods in terms of performance and efficiency.

Question Answering (QA) in NLP is the task of finding answers to a query within a relevant context retrieved by a retrieval system. Yet, the mix of relevant and irrelevant information in these contexts can hinder performance enhancements in QA tasks. To address this, we introduce a context filtering approach that removes non-essential details, summarizing crucial content through Reward Modeling. This method emphasizes keeping vital data while omitting the extraneous during summarization model training. We offer a framework for developing efficient QA models by discerning useful information from dataset pairs, bypassing the need for costly human evaluation. Furthermore, we show that our approach can significantly outperform the baseline, as evidenced by a 6.8-fold increase in the EM Per Token (EPT) metric, which we propose as a measure of token efficiency, indicating a notable token-efficiency boost for low-resource settings.

Leveraging temporal context is crucial for success in partially observable robotic tasks. However, prior work in behavior cloning has demonstrated inconsistent performance gains when using multi-frame observations. In this paper, we introduce ContextVLA, a policy model that robustly improves robotic task performance by effectively leveraging multi-frame observations. Our approach is motivated by the key observation that Vision-Language-Action models (VLA), i.e., policy models built upon a Vision-Language Model (VLM), more effectively utilize multi-frame observations for action generation. This suggests that VLMs' inherent temporal understanding capability enables them to extract more meaningful context from multi-frame observations. However, the high dimensionality of video inputs introduces significant computational overhead, making VLA training and inference inefficient. To address this, ContextVLA compresses past observations into a single context token, allowing the policy to efficiently leverage temporal context for action generation. Our experiments show that ContextVLA consistently improves over single-frame VLAs and achieves the benefits of full multi-frame training but with reduced training and inference times.

Knowledge distillation improves large language model (LLM) reasoning by compressing the knowledge of a teacher LLM to train smaller LLMs. On-policy distillation advances this approach by having the student sample its own trajectories while a teacher LLM provides dense token-level supervision, addressing the distribution mismatch between training and inference in off-policy distillation methods. However, on-policy distillation typically requires a separate, often larger, teacher LLM and does not explicitly leverage ground-truth solutions available in reasoning datasets. Inspired by the intuition that a sufficiently capable LLM can rationalize external privileged reasoning traces and teach its weaker self (i.e., the version without access to privileged information), we introduce On-Policy Self-Distillation (OPSD), a framework where a single model acts as both teacher and student by conditioning on different contexts. The teacher policy conditions on privileged information (e.g., verified reasoning traces) while the student policy sees only the question; training minimizes the per-token divergence between these distributions over the student's own rollouts. We demonstrate the efficacy of our method on multiple mathematical reasoning benchmarks, achieving 4-8x token efficiency compared to reinforcement learning methods such as GRPO and superior performance over off-policy distillation methods.

In recent years, the long-range attention mechanism of vision transformers has driven significant performance breakthroughs across various computer vision tasks. However, the traditional self-attention mechanism, which processes both informative and non-informative tokens, suffers from inefficiency and inaccuracies. While sparse attention mechanisms have been introduced to mitigate these issues by pruning tokens involved in attention, they often lack context-awareness and intelligence. These mechanisms frequently apply a uniform token selection strategy across different inputs for batch training or optimize efficiency only for the inference stage. To overcome these challenges, we propose a novel algorithm: Select and Pack Attention (SPA). SPA dynamically selects informative tokens using a low-cost gating layer supervised by selection labels and packs these tokens into new batches, enabling a variable number of tokens to be used in parallelized GPU batch training and inference. Extensive experiments across diverse datasets and computer vision tasks demonstrate that SPA delivers superior performance and efficiency, including a 0.6 mAP improvement in object detection and a 16.4% reduction in computational costs.

Understanding how transformer components operate in LLMs is important, as it is at the core of recent technological advances in artificial intelligence. In this work, we revisit the challenges associated with interpretability of feed-forward modules (FFNs) and propose MemoryLLM, which aims to decouple FFNs from self-attention and enables us to study the decoupled FFNs as context-free token-wise neural retrieval memory. In detail, we investigate how input tokens access memory locations within FFN parameters and the importance of FFN memory across different downstream tasks. MemoryLLM achieves context-free FFNs by training them in isolation from self-attention directly using the token embeddings. This approach allows FFNs to be pre-computed as token-wise lookups (ToLs), enabling on-demand transfer between VRAM and storage, additionally enhancing inference efficiency. We also introduce Flex-MemoryLLM, positioning it between a conventional transformer design and MemoryLLM. This architecture bridges the performance gap caused by training FFNs with context-free token-wise embeddings.

There is growing demand for performing inference with hundreds of thousands of input tokens on trained transformer models. Inference at this extreme scale demands significant computational resources, hindering the application of transformers at long contexts on commodity (i.e not data center scale) hardware. To address the inference time costs associated with running self-attention based transformer language models on long contexts and enable their adoption on widely available hardware, we propose a tunable mechanism that reduces the cost of the forward pass by attending to only the most relevant tokens at every generation step using a top-k selection mechanism. We showcase the efficiency gains afforded by our method by performing inference on context windows up to 1M tokens using approximately 16GB of GPU RAM. Our experiments reveal that models are capable of handling the sparsity induced by the reduced number of keys and values. By attending to less than 2% of input tokens, we achieve over 95% of model performance on common benchmarks (RULER, AlpacaEval, and Open LLM Leaderboard).

Positional encoding is a vital component of Transformer architectures, enabling models to incorporate sequence order into self-attention mechanisms. Rotary Positional Embeddings (RoPE) have become a widely adopted solution due to their compatibility with relative position encoding and computational efficiency. However, RoPE relies on static, input-independent sinusoidal frequency patterns, limiting its ability to model context-sensitive relationships. In this work, we propose CARoPE (Context-Aware Rotary Positional Embedding), a novel generalization of RoPE that dynamically generates head-specific frequency patterns conditioned on token embeddings. This design introduces token- and context-sensitive positional representations while preserving RoPE efficiency and architectural simplicity. CARoPE computes input-dependent phase shifts using a bounded transformation of token embeddings and integrates them into the rotary mechanism across attention heads. We evaluate CARoPE on the FineWeb-Edu-10B dataset using GPT-2 variants trained on next-token prediction tasks. Experimental results show that CARoPE consistently outperforms RoPE and other common positional encoding baselines, achieving significantly lower perplexity, even at longer context lengths. Additionally, CARoPE enables faster training throughput without sacrificing model stability. These findings demonstrate that CARoPE offers a scalable, expressive, and efficient upgrade to existing positional encoding strategies in Transformer models.

The increasing context window size in Large Language Models (LLMs), such as the GPT and LLaMA series, has improved their ability to tackle complex, long-text tasks, but at the cost of inference efficiency, particularly regarding memory and computational complexity. Existing methods, including selective token retention and window-based attention, improve efficiency but risk discarding important tokens needed for future text generation. In this paper, we propose an approach that enhances LLM efficiency without token loss by reducing the memory and computational load of less important tokens, rather than discarding them.We address two challenges: 1) investigating the distribution of important tokens in the context, discovering recent tokens are more important than distant tokens in context, and 2) optimizing resources for distant tokens by sharing attention scores across layers. The experiments show that our method saves 35% KV cache without compromising the performance.

Modern large vision-language models (LVLMs) convert each input image into a large set of tokens, far outnumbering the text tokens. Although this improves visual perception, it introduces severe image token redundancy. Because image tokens carry sparse information, many add little to reasoning, yet greatly increase inference cost. The emerging image token pruning methods tackle this issue by identifying the most important tokens and discarding the rest. These methods can raise efficiency with only modest performance loss. However, most of them only consider single-image tasks and overlook multimodal in-context learning (ICL), where redundancy is greater and efficiency is more critical. Redundant tokens weaken the advantage of multimodal ICL for rapid domain adaptation and cause unstable performance. Applying existing pruning methods in this setting leads to large accuracy drops, exposing a clear gap and the need for new techniques. Thus, we propose Contextually Adaptive Token Pruning (CATP), a training-free pruning method targeted at multimodal ICL. CATP consists of two stages that perform progressive pruning to fully account for the complex cross-modal interactions in the input sequence. After removing 77.8\% of the image tokens, CATP produces an average performance gain of 0.6\% over the vanilla model on four LVLMs and eight benchmarks, exceeding all baselines remarkably. Meanwhile, it effectively improves efficiency by achieving an average reduction of 10.78\% in inference latency. CATP enhances the practical value of multimodal ICL and lays the groundwork for future progress in interleaved image-text scenarios.

We present context parallelism for long-context large language model inference, which achieves near-linear scaling for long-context prefill latency with up to 128 H100 GPUs across 16 nodes. Particularly, our method achieves 1M context prefill with Llama3 405B model in 77s (93% parallelization efficiency, 63% FLOPS utilization) and 128K context prefill in 3.8s. We develop two lossless exact ring attention variants: pass-KV and pass-Q to cover a wide range of use cases with the state-of-the-art performance: full prefill, persistent KV prefill and decode. Benchmarks on H100 GPU hosts inter-connected with RDMA and TCP both show similar scalability for long-context prefill, demonstrating that our method scales well using common commercial data center with medium-to-low inter-host bandwidth.

The escalating demand for long-context applications has intensified the necessity of extending the LLM context windows. Despite recent fine-tuning approaches successfully expanding context lengths, their high memory footprints, especially for activations, present a critical practical limitation. Current parameter-efficient fine-tuning methods prioritize reducing parameter update overhead over addressing activation memory constraints. Similarly, existing sparsity mechanisms improve computational efficiency but overlook activation memory optimization due to the phenomenon of Shadowy Activation. In this paper, we propose LeMo, the first LLM fine-tuning system that explores and exploits a new token-level sparsity mechanism inherent in long-context scenarios, termed Contextual Token Sparsity. LeMo minimizes redundant token involvement by assessing the informativeness of token embeddings while preserving model accuracy. Specifically, LeMo introduces three key techniques: (1) Token Elimination, dynamically identifying and excluding redundant tokens across varying inputs and layers. (2) Pattern Prediction, utilizing well-trained predictors to approximate token sparsity patterns with minimal overhead. (3) Kernel Optimization, employing permutation-free and segment-based strategies to boost system performance. We implement LeMo as an end-to-end fine-tuning system compatible with various LLM architectures and other optimization techniques. Comprehensive evaluations demonstrate that LeMo reduces memory consumption by up to 1.93x and achieves up to 1.36x speedups, outperforming state-of-the-art fine-tuning systems.

Large language models (LLMs) have achieved notable progress. Despite their success, next-token prediction (NTP), the dominant method for LLM training and inference, is constrained in both contextual coverage and inference efficiency due to its inherently sequential process. To overcome these challenges, we propose leap multi-token prediction~(L-MTP), an innovative token prediction method that extends the capabilities of multi-token prediction (MTP) by introducing a leap-based mechanism. Unlike conventional MTP, which generates multiple tokens at adjacent positions, L-MTP strategically skips over intermediate tokens, predicting non-sequential ones in a single forward pass. This structured leap not only enhances the model's ability to capture long-range dependencies but also enables a decoding strategy specially optimized for non-sequential leap token generation, effectively accelerating inference. We theoretically demonstrate the benefit of L-MTP in improving inference efficiency. Experiments across diverse benchmarks validate its merit in boosting both LLM performance and inference speed. The source code is available at https://github.com/Xiaohao-Liu/L-MTP.

In Transformer architectures, tokens\textemdash discrete units derived from raw data\textemdash are formed by segmenting inputs into fixed-length chunks. Each token is then mapped to an embedding, enabling parallel attention computations while preserving the input's essential information. Due to the quadratic computational complexity of transformer self-attention mechanisms, token reduction has primarily been used as an efficiency strategy. This is especially true in single vision and language domains, where it helps balance computational costs, memory usage, and inference latency. Despite these advances, this paper argues that token reduction should transcend its traditional efficiency-oriented role in the era of large generative models. Instead, we position it as a fundamental principle in generative modeling, critically influencing both model architecture and broader applications. Specifically, we contend that across vision, language, and multimodal systems, token reduction can: (i) facilitate deeper multimodal integration and alignment, (ii) mitigate "overthinking" and hallucinations, (iii) maintain coherence over long inputs, and (iv) enhance training stability, etc. We reframe token reduction as more than an efficiency measure. By doing so, we outline promising future directions, including algorithm design, reinforcement learning-guided token reduction, token optimization for in-context learning, and broader ML and scientific domains. We highlight its potential to drive new model architectures and learning strategies that improve robustness, increase interpretability, and better align with the objectives of generative modeling.

Extending the functionality of the Transformer model to accommodate longer sequence lengths has become a critical challenge. This extension is crucial not only for improving tasks such as language translation and long-context processing but also for enabling novel applications like chatbots, code generation, and multimedia content creation. The primary obstacle is the self-attention mechanism, which scales quadratically with sequence length in terms of computation time and memory requirements. LongLoRA proposed shifted sparse attention (S\(^2\)-Attn), effectively enabling context extension and leading to non-trivial computation savings with similar performance to fine-tuning with vanilla attention. However, LongLoRA is still not as efficient as vanilla attention, reaching only 39\% of the perplexity improvement compared to full attention. This inefficiency is due to the cyclic shift applied within different attention head patterns, causing either chaos in the attention head structure or unnecessary information exchange between token groups. To address these issues, We propose SinkLoRA, which features better work partitioning. Specifically, (1) we developed SF-Attn with a segmentation and reassembly algorithm to proportionally return cyclically shifted groups of attention heads to their un-shifted state together with global attention of "sink attention tokens", achieving 92\% of the perplexity improvement compared to full attention after fine tuning, and (2) applied a SOTA KV cache compression algorithm H_2O to accelerate inference. Furthermore, We conducted supervised fine-tuning with SinkLoRA using a self collected LongAlpaca-plus dataset. All our code, models, datasets, and demos are available at https://github.com/Dexter-GT-86/SinkLoRA.

While Chain-of-thought (CoT) reasoning enables LLMs to solve challenging reasoning problems, as KV cache grows linearly with the number of generated tokens, CoT reasoning faces scaling issues in terms of speed and memory usage. In this work, we propose MemoSight (Memory-Foresight-based reasoning), a unified framework that integrates both context compression and multi-token prediction to mitigate the efficiency issues while maintaining CoT reasoning performance. Our framework adopts the same minimalist design for both context compression and multi-token prediction via special tokens and their corresponding position layout tailored to each token type. Comprehensive experiments on four reasoning benchmarks demonstrate that MemoSight reduces the KV cache footprint by up to 66% and accelerates inference by 1.56x, while outperforming existing CoT compression methods.

Processing long-form audio is a major challenge for Large Audio Language models (LALMs). These models struggle with the quadratic cost of attention (O(N^2)) and with modeling long-range temporal dependencies. Existing audio benchmarks are built mostly from short clips and do not evaluate models in realistic long context settings. To address this gap, we introduce AudioMarathon, a benchmark designed to evaluate both understanding and inference efficiency on long-form audio. AudioMarathon provides a diverse set of tasks built upon three pillars: long-context audio inputs with durations ranging from 90.0 to 300.0 seconds, which correspond to encoded sequences of 2,250 to 7,500 audio tokens, respectively, full domain coverage across speech, sound, and music, and complex reasoning that requires multi-hop inference. We evaluate state-of-the-art LALMs and observe clear performance drops as audio length grows. We also study acceleration techniques and analyze the trade-offs of token pruning and KV cache eviction. The results show large gaps across current LALMs and highlight the need for better temporal reasoning and memory-efficient architectures. We believe AudioMarathon will drive the audio and multimodal research community to develop more advanced audio understanding models capable of solving complex audio tasks.

Serving numerous users and requests concurrently requires good fairness in Large Language Models (LLMs) serving system. This ensures that, at the same cost, the system can meet the Service Level Objectives (SLOs) of more users , such as time to first token (TTFT) and time between tokens (TBT), rather than allowing a few users to experience performance far exceeding the SLOs. To achieve better fairness, the preemption-based scheduling policy dynamically adjusts the priority of each request to maintain balance during runtime. However, existing systems tend to overly prioritize throughput, overlooking the overhead caused by preemption-induced context switching, which is crucial for maintaining fairness through priority adjustments. In this work, we identify three main challenges that result in this overhead. 1) Inadequate I/O utilization. 2) GPU idleness. 3) Unnecessary I/O transmission during multi-turn conversations. Our key insight is that the block-based KV cache memory policy in existing systems, while achieving near-zero memory waste, leads to discontinuity and insufficient granularity in the KV cache memory. To respond, we introduce FastSwitch, a fairness-aware serving system that not only aligns with existing KV cache memory allocation policy but also mitigates context switching overhead. Our evaluation shows that FastSwitch outperforms the state-of-the-art LLM serving system vLLM with speedups of 1.4-11.2x across different tail TTFT and TBT.

Large reasoning models (LRMs) achieve state-of-the-art performance on challenging benchmarks by generating long chains of intermediate steps, but their inference cost is dominated by decoding, where each new token must attend to the entire growing sequence. Existing sparse attention methods reduce computation by pruning the key-value (KV) cache, yet they suffer from severe accuracy degradation on reasoning tasks due to cumulative selection errors and the dynamic importance of tokens over long derivations. We present DELTA, a training-free sparse attention mechanism that achieves computational efficiency without sacrificing model accuracy. DELTA partitions transformer layers into three groups: initial layers that use full attention, a small set of selection layers that identify salient tokens via aggregated head-level attention scores, and subsequent sparse-attention layers that attend only to the selected subset. This design preserves the full KV cache in GPU memory for accuracy, while avoiding expensive full-attention computation over many layers. On reasoning benchmarks such as AIME and GPQA-Diamond, DELTA matches or surpasses full attention in accuracy, while reducing the number of attended tokens by up to 5times and delivering 1.5times end-to-end speedup. Our results show that selective reuse of intermediate attention maps offers a robust path toward efficient long-context reasoning.

With the development of large language models (LLMs), the ability to handle longer contexts has become a key capability for Web applications such as cross-document understanding and LLM-powered search systems. However, this progress faces two major challenges: performance degradation due to sequence lengths out-of-distribution, and excessively long inference times caused by the quadratic computational complexity of attention. These issues hinder the application of LLMs in long-context scenarios. In this paper, we propose Dynamic Token-Level KV Cache Selection (TokenSelect), a model-agnostic, training-free method for efficient and accurate long-context inference. TokenSelect builds upon the observation of non-contiguous attention sparsity, using Query-Key dot products to measure per-head KV Cache criticality at token-level. By per-head soft voting mechanism, TokenSelect selectively involves a small number of critical KV cache tokens in the attention calculation without sacrificing accuracy. To further accelerate TokenSelect, we designed the Selection Cache based on observations of consecutive Query similarity and implemented efficient dot product kernel, significantly reducing the overhead of token selection. A comprehensive evaluation of TokenSelect demonstrates up to 23.84x speedup in attention computation and up to 2.28x acceleration in end-to-end latency, while providing superior performance compared to state-of-the-art long-context inference methods.

With the development of large language models (LLMs), there has been an increasing need for significant advancements in handling long contexts. To enhance long-context capabilities, constructing high-quality training data with long-range dependencies is crucial. Existing methods to select long-context data often rely on sentence-level analysis, which can be greatly optimized in both performance and efficiency. In this paper, we propose a novel token-level framework, LongAttn, which leverages the self-attention mechanism of LLMs to measure the long-range dependencies for the data. By calculating token-level dependency strength and distribution uniformity of token scores, LongAttn effectively quantifies long-range dependencies, enabling more accurate and efficient data selection. We filter LongABC-32K from open-source long-context datasets (ArXiv, Book, and Code). Through our comprehensive experiments, LongAttn has demonstrated its excellent effectiveness, scalability, and efficiency. To facilitate future research in long-context data, we released our code and the high-quality long-context training data LongABC-32K.

Long-context inference for Large Language Models (LLMs) is heavily limited by high computational demands. While several existing methods optimize attention computation, they still process the full set of hidden states at each layer, limiting overall efficiency. In this work, we propose SlimInfer, an innovative framework that aims to accelerate inference by directly pruning less critical prompt tokens during the forward pass. Our key insight is an information diffusion phenomenon: As information from critical tokens propagates through layers, it becomes distributed across the entire sequence. This diffusion process suggests that LLMs can maintain their semantic integrity when excessive tokens, even including these critical ones, are pruned in hidden states. Motivated by this, SlimInfer introduces a dynamic fine-grained pruning mechanism that accurately removes redundant tokens of hidden state at intermediate layers. This layer-wise pruning naturally enables an asynchronous KV cache manager that prefetches required token blocks without complex predictors, reducing both memory usage and I/O costs. Extensive experiments show that SlimInfer can achieve up to 2.53times time-to-first-token (TTFT) speedup and 1.88times end-to-end latency reduction for LLaMA3.1-8B-Instruct on a single RTX 4090, without sacrificing performance on LongBench. Our code will be released upon acceptance.

The Model Context Protocol (MCP) introduces a standard specification that defines how Foundation Model (FM)-based agents should interact with external systems by invoking tools. However, to understand a tool's purpose and features, FMs rely on natural-language tool descriptions, making these descriptions a critical component in guiding FMs to select the optimal tool for a given (sub)task and to pass the right arguments to the tool. While defects or smells in these descriptions can misguide FM-based agents, their prevalence and consequences in the MCP ecosystem remain unclear. Hence, we examine 856 tools spread across 103 MCP servers empirically, assess their description quality, and their impact on agent performance. We identify six components of tool descriptions from the literature, develop a scoring rubric utilizing these components, and then formalize tool description smells based on this rubric. By operationalizing this rubric through an FM-based scanner, we find that 97.1% of the analyzed tool descriptions contain at least one smell, with 56% failing to state their purpose clearly. While augmenting these descriptions for all components improves task success rates by a median of 5.85 percentage points and improves partial goal completion by 15.12%, it also increases the number of execution steps by 67.46% and regresses performance in 16.67% of cases. These results indicate that achieving performance gains is not straightforward; while execution cost can act as a trade-off, execution context can also impact. Furthermore, component ablations show that compact variants of different component combinations often preserve behavioral reliability while reducing unnecessary token overhead, enabling more efficient use of the FM context window and lower execution costs.

Large Language Models (LLMs) have demonstrated remarkable capabilities in handling long context inputs, but this comes at the cost of increased computational resources and latency. Our research introduces a novel approach for the long context bottleneck to accelerate LLM inference and reduce GPU memory consumption. Our research demonstrates that LLMs can identify relevant tokens in the early layers before generating answers to a query. Leveraging this insight, we propose an algorithm that uses early layers of an LLM as filters to select and compress input tokens, significantly reducing the context length for subsequent processing. Our method, GemFilter, demonstrates substantial improvements in both speed and memory efficiency compared to existing techniques, such as standard attention and SnapKV/H2O. Notably, it achieves a 2.4times speedup and 30\% reduction in GPU memory usage compared to SOTA methods. Evaluation on the Needle in a Haystack task shows that GemFilter significantly outperforms standard attention, SnapKV and demonstrates comparable performance on the LongBench challenge. GemFilter is simple, training-free, and broadly applicable across different LLMs. Crucially, it provides interpretability by allowing humans to inspect the selected input sequence. These findings not only offer practical benefits for LLM deployment, but also enhance our understanding of LLM internal mechanisms, paving the way for further optimizations in LLM design and inference. Our code is available at https://github.com/SalesforceAIResearch/GemFilter.

Large language models (LLMs) have shown remarkable performance in various natural language processing tasks. However, a primary constraint they face is the context limit, i.e., the maximum number of tokens they can process. Previous works have explored architectural changes and modifications in positional encoding to relax the constraint, but they often require expensive training or do not address the computational demands of self-attention. In this paper, we present Hierarchical cOntext MERging (HOMER), a new training-free scheme designed to overcome the limitations. HOMER uses a divide-and-conquer algorithm, dividing long inputs into manageable chunks. Each chunk is then processed collectively, employing a hierarchical strategy that merges adjacent chunks at progressive transformer layers. A token reduction technique precedes each merging, ensuring memory usage efficiency. We also propose an optimized computational order reducing the memory requirement to logarithmically scale with respect to input length, making it especially favorable for environments with tight memory restrictions. Our experiments demonstrate the proposed method's superior performance and memory efficiency, enabling the broader use of LLMs in contexts requiring extended context. Code is available at https://github.com/alinlab/HOMER.

Transformers achieve state-of-the-art results across many tasks, but their uniform application of quadratic self-attention to every token at every layer makes them computationally expensive. We introduce DTRNet (Dynamic Token Routing Network), an improved Transformer architecture that allows tokens to dynamically skip the quadratic cost of cross-token mixing while still receiving lightweight linear updates. By preserving the MLP module and reducing the attention cost for most tokens to linear, DTRNet ensures that every token is explicitly updated while significantly lowering overall computation. This design offers an efficient and effective alternative to standard dense attention. Once trained, DTRNet blocks routes only ~10% of tokens through attention at each layer while maintaining performance comparable to a full Transformer. It consistently outperforms routing-based layer skipping methods such as MoD and D-LLM in both accuracy and memory at matched FLOPs, while routing fewer tokens to full attention. Its efficiency gains, scales with sequence length, offering significant reduction in FLOPs for long-context inputs. By decoupling token updates from attention mixing, DTRNet substantially reduces the quadratic share of computation, providing a simple, efficient, and scalable alternative to Transformers.

Vision transformer has achieved impressive performance for many vision tasks. However, it may suffer from high redundancy in capturing local features for shallow layers. Local self-attention or early-stage convolutions are thus utilized, which sacrifice the capacity to capture long-range dependency. A challenge then arises: can we access efficient and effective global context modeling at the early stages of a neural network? To address this issue, we draw inspiration from the design of superpixels, which reduces the number of image primitives in subsequent processing, and introduce super tokens into vision transformer. Super tokens attempt to provide a semantically meaningful tessellation of visual content, thus reducing the token number in self-attention as well as preserving global modeling. Specifically, we propose a simple yet strong super token attention (STA) mechanism with three steps: the first samples super tokens from visual tokens via sparse association learning, the second performs self-attention on super tokens, and the last maps them back to the original token space. STA decomposes vanilla global attention into multiplications of a sparse association map and a low-dimensional attention, leading to high efficiency in capturing global dependencies. Based on STA, we develop a hierarchical vision transformer. Extensive experiments demonstrate its strong performance on various vision tasks. In particular, without any extra training data or label, it achieves 86.4% top-1 accuracy on ImageNet-1K with less than 100M parameters. It also achieves 53.9 box AP and 46.8 mask AP on the COCO detection task, and 51.9 mIOU on the ADE20K semantic segmentation task. Code will be released at https://github.com/hhb072/SViT.

We present core attention disaggregation (CAD), a technique that improves long-context large language model training by decoupling the core attention computation, softmax(QK^T)V, from the rest of the model and executing it on a separate pool of devices. In existing systems, core attention is colocated with other layers; at long context lengths, its quadratic compute growth compared to the near-linear growth of other components causes load imbalance and stragglers across data and pipeline parallel groups. CAD is enabled by two observations. First, core attention is stateless: it has no trainable parameters and only minimal transient data, so balancing reduces to scheduling compute-bound tasks. Second, it is composable: modern attention kernels retain high efficiency when processing fused batches of token-level shards with arbitrary lengths. CAD partitions core attention into token-level tasks and dispatches them to dedicated attention servers, which dynamically rebatch tasks to equalize compute without sacrificing kernel efficiency. We implement CAD in a system called DistCA, which uses a ping-pong execution scheme to fully overlap communication with computation and in-place execution on attention servers to reduce memory use. On 512 H200 GPUs and context lengths up to 512k tokens, DistCA improves end-to-end training throughput by up to 1.35x, eliminates data and pipeline parallel stragglers, and achieves near-perfect compute and memory balance.

Training Large Language Models (LLMs) on long contexts is severely constrained by prohibitive GPU memory overhead, not training time. The primary culprits are the activations, whose memory footprints scale linearly with sequence length. We introduce OOMB, a highly memory-efficient training system that directly confronts this barrier. Our approach employs a chunk-recurrent training framework with on-the-fly activation recomputation, which maintains a constant activation memory footprint (O(1)) and shifts the primary bottleneck to the growing KV cache. To manage the KV cache, OOMB integrates a suite of synergistic optimizations: a paged memory manager for both the KV cache and its gradients to eliminate fragmentation, asynchronous CPU offloading to hide data transfer latency, and page-level sparse attention to reduce both computational complexity and communication overhead. The synergy of these techniques yields exceptional efficiency. Our empirical results show that for every additional 10K tokens of context, the end-to-end training memory overhead increases by a mere 10MB for Qwen2.5-7B. This allows training Qwen2.5-7B with a 4M-token context on a single H200 GPU, a feat that would otherwise require a large cluster using context parallelism. This work represents a substantial advance in resource efficiency for long-context LLM training. The source code is available at https://github.com/wenhaoli-xmu/OOMB.

Large Language Model agents increasingly operate external systems through programmatic interfaces, yet practitioners lack empirical guidance on how to structure the context these agents consume. Using SQL generation as a proxy for programmatic agent operations, we present a systematic study of context engineering for structured data, comprising 9,649 experiments across 11 models, 4 formats (YAML, Markdown, JSON, Token-Oriented Object Notation [TOON]), and schemas ranging from 10 to 10,000 tables. Our findings challenge common assumptions. First, architecture choice is model-dependent: file-based context retrieval improves accuracy for frontier-tier models (Claude, GPT, Gemini; +2.7%, p=0.029) but shows mixed results for open source models (aggregate -7.7%, p<0.001), with deficits varying substantially by model. Second, format does not significantly affect aggregate accuracy (chi-squared=2.45, p=0.484), though individual models, particularly open source, exhibit format-specific sensitivities. Third, model capability is the dominant factor, with a 21 percentage point accuracy gap between frontier and open source tiers that dwarfs any format or architecture effect. Fourth, file-native agents scale to 10,000 tables through domain-partitioned schemas while maintaining high navigation accuracy. Fifth, file size does not predict runtime efficiency: compact or novel formats can incur a token overhead driven by grep output density and pattern unfamiliarity, with the magnitude depending on model capability. These findings provide practitioners with evidence-based guidance for deploying LLM agents on structured systems, demonstrating that architectural decisions should be tailored to model capability rather than assuming universal best practices.

Large language models (LLMs) process entire input contexts indiscriminately, which is inefficient in cases where the information required to answer a query is localized within the context. We present dynamic context cutoff, a human-inspired method enabling LLMs to self-terminate processing upon acquiring sufficient task-relevant information. Through analysis of model internals, we discover that specific attention heads inherently encode "sufficiency signals" - detectable through lightweight classifiers - that predict when critical information has been processed. This reveals a new efficiency paradigm: models' internal understanding naturally dictates processing needs rather than external compression heuristics. Comprehensive experiments across six QA datasets (up to 40K tokens) with three model families (LLaMA/Qwen/Mistral, 1B0-70B) demonstrate 1.33x average token reduction while improving accuracy by 1.3%. Furthermore, our method demonstrates better performance with the same rate of token reduction compared to other context efficiency methods. Additionally, we observe an emergent scaling phenomenon: while smaller models require require probing for sufficiency detection, larger models exhibit intrinsic self-assessment capabilities through prompting.

Most current popular subword tokenizers are trained based on word frequency statistics over a corpus, without considering information about co-occurrence or context. Nevertheless, the resulting vocabularies are used in language models' highly contextualized settings. We present SaGe, a tokenizer that tailors subwords for their downstream use by baking in the contextualized signal at the vocabulary creation phase. We show that SaGe does a better job than current widespread tokenizers in keeping token contexts cohesive, while not incurring a large price in terms of encoding efficiency or domain robustness. SaGe improves performance on English GLUE classification tasks as well as on NER, and on Inference and NER in Turkish, demonstrating its robustness to language properties such as morphological exponence and agglutination.

Despite their powerful capabilities, Multimodal Large Language Models (MLLMs) suffer from considerable computational overhead due to their reliance on massive visual tokens. Recent studies have explored token pruning to alleviate this problem, which typically uses text-vision cross-attention or [CLS] attention to assess and discard redundant visual tokens. In this work, we identify a critical limitation of such attention-first pruning approaches, i.e., they tend to preserve semantically similar tokens, resulting in pronounced performance drops under high pruning ratios. To this end, we propose {HoloV}, a simple yet effective, plug-and-play visual token pruning framework for efficient inference. Distinct from previous attention-first schemes, HoloV rethinks token retention from a holistic perspective. By adaptively distributing the pruning budget across different spatial crops, HoloV ensures that the retained tokens capture the global visual context rather than isolated salient features. This strategy minimizes representational collapse and maintains task-relevant information even under aggressive pruning. Experimental results demonstrate that our HoloV achieves superior performance across various tasks, MLLM architectures, and pruning ratios compared to SOTA methods. For instance, LLaVA1.5 equipped with HoloV preserves 95.8\% of the original performance after pruning 88.9\% of visual tokens, achieving superior efficiency-accuracy trade-offs.

The quadratic computational complexity of softmax transformers has become a bottleneck in long-context scenarios. In contrast, linear attention model families provide a promising direction towards a more efficient sequential model. These linear attention models compress past KV values into a single hidden state, thereby efficiently reducing complexity during both training and inference. However, their expressivity remains limited by the size of their hidden state. Previous work proposed interleaving softmax and linear attention layers to reduce computational complexity while preserving expressivity. Nevertheless, the efficiency of these models remains bottlenecked by their softmax attention layers. In this paper, we propose Neural Attention Search Linear (NAtS-L), a framework that applies both linear attention and softmax attention operations within the same layer on different tokens. NAtS-L automatically determines whether a token can be handled by a linear attention model, i.e., tokens that have only short-term impact and can be encoded into fixed-size hidden states, or require softmax attention, i.e., tokens that contain information related to long-term retrieval and need to be preserved for future queries. By searching for optimal Gated DeltaNet and softmax attention combinations across tokens, we show that NAtS-L provides a strong yet efficient token-level hybrid architecture.

In-Context Operator Networks (ICONs) have demonstrated the ability to learn operators across diverse partial differential equations using few-shot, in-context learning. However, existing ICONs process each spatial point as an individual token, severely limiting computational efficiency when handling dense data in higher spatial dimensions. We propose Vision In-Context Operator Networks (VICON), which integrates vision transformer architectures to efficiently process 2D data through patch-wise operations while preserving ICON's adaptability to multiphysics systems and varying timesteps. Evaluated across three fluid dynamics benchmarks, VICON significantly outperforms state-of-the-art baselines: DPOT and MPP, reducing the averaged last-step rollout error by 37.9% compared to DPOT and 44.7% compared to MPP, while requiring only 72.5% and 34.8% of their respective inference times. VICON naturally supports flexible rollout strategies with varying timestep strides, enabling immediate deployment in imperfect measurement systems where sampling frequencies may differ or frames might be dropped - common challenges in real-world settings - without requiring retraining or interpolation. In these realistic scenarios, VICON exhibits remarkable robustness, experiencing only 24.41% relative performance degradation compared to 71.37%-74.49% degradation in baseline methods, demonstrating its versatility for deploying in realistic applications. Our scripts for processing datasets and code are publicly available at https://github.com/Eydcao/VICON.

Recent advancements in Transformer-based large language models (LLMs) have set new standards in natural language processing. However, the classical softmax attention incurs significant computational costs, leading to a O(T) complexity for per-token generation, where T represents the context length. This work explores reducing LLMs' complexity while maintaining performance by introducing Rodimus and its enhanced version, Rodimus+. Rodimus employs an innovative data-dependent tempered selection (DDTS) mechanism within a linear attention-based, purely recurrent framework, achieving significant accuracy while drastically reducing the memory usage typically associated with recurrent models. This method exemplifies semantic compression by maintaining essential input information with fixed-size hidden states. Building on this, Rodimus+ combines Rodimus with the innovative Sliding Window Shared-Key Attention (SW-SKA) in a hybrid approach, effectively leveraging the complementary semantic, token, and head compression techniques. Our experiments demonstrate that Rodimus+-1.6B, trained on 1 trillion tokens, achieves superior downstream performance against models trained on more tokens, including Qwen2-1.5B and RWKV6-1.6B, underscoring its potential to redefine the accuracy-efficiency balance in LLMs. Model code and pre-trained checkpoints will be available soon.

Improving text representation has attracted much attention to achieve expressive text-to-speech (TTS). However, existing works only implicitly learn the prosody with masked token reconstruction tasks, which leads to low training efficiency and difficulty in prosody modeling. We propose CLAPSpeech, a cross-modal contrastive pre-training framework that explicitly learns the prosody variance of the same text token under different contexts. Specifically, 1) We encourage the model to connect the text context with its corresponding prosody pattern in the joint multi-modal space with the elaborate design of the encoder inputs and contrastive loss; 2) We introduce a multi-scale pre-training pipeline to capture prosody patterns in multiple levels. We show how to incorporate CLAPSpeech into existing TTS models for better prosody. Experiments on three datasets not only show that CLAPSpeech could improve the prosody prediction for existing TTS methods, but also demonstrate its generalization ability to adapt to multiple languages and multi-speaker TTS. We also deeply analyze the principle behind the performance of CLAPSpeech. Ablation studies demonstrate the necessity of each component in our method. Source code and audio samples are available at https://clapspeech.github.io.

Vision-language models (VLMs) often generate massive visual tokens that greatly increase inference latency and memory footprint; while training-free token pruning offers a practical remedy, existing methods still struggle to balance local evidence and global context under aggressive compression. We propose Focus-Scan-Refine (FSR), a human-inspired, plug-and-play pruning framework that mimics how humans answer visual questions: focus on key evidence, then scan globally if needed, and refine the scanned context by aggregating relevant details. FSR first focuses on key evidence by combining visual importance with instruction relevance, avoiding the bias toward visually salient but query-irrelevant regions. It then scans for complementary context conditioned on the focused set, selecting tokens that are most different from the focused evidence. Finally, FSR refines the scanned context by aggregating nearby informative tokens into the scan anchors via similarity-based assignment and score-weighted merging, without increasing the token budget. Extensive experiments across multiple VLM backbones and vision-language benchmarks show that FSR consistently improves the accuracy-efficiency trade-off over existing state-of-the-art pruning methods. The source codes can be found at https://github.com/ILOT-code/FSR.

Tokenization plays a critical role in language modeling, yet existing approaches such as Byte-Pair Encoding (BPE) or WordPiece operate purely on frequency statistics, ignoring the underlying semantic structure of text. This leads to over-tokenization of semantically redundant spans and underutilization of contextual coherence, particularly in long-context scenarios. In this work, we propose SemToken, a semantic-aware tokenization framework that jointly reduces token redundancy and improves computation efficiency. SemToken first extracts contextual semantic embeddings via lightweight encoders and performs local semantic clustering to merge semantically equivalent tokens. Then, it allocates heterogeneous token granularity based on semantic density, allowing finer-grained tokenization in content-rich regions and coarser compression in repetitive or low-entropy spans. SemToken can be seamlessly integrated with modern language models and attention acceleration methods. Experiments on long-context language modeling benchmarks such as WikiText-103 and LongBench show that SemToken achieves up to 2.4times reduction in token count and 1.9times speedup, with negligible or no degradation in perplexity and downstream accuracy. Our findings suggest that semantic structure offers a promising new axis for optimizing tokenization and computation in large language models.

We present Nacrith, a lossless compression system that combines a 135M-parameter transformer language model (SmolLM2-135M) with an ensemble of lightweight online predictors and a 32-bit arithmetic coder. Beyond the base LLM-plus-arithmetic-coding paradigm, Nacrith introduces several contributions: (1) a CDF precision upgrade from 2^16 to 2^24 that eliminates ~75% of quantization overhead caused by minimum-probability floors in large vocabularies; (2) a token-level N-gram model for fast local predictions; (3) an adaptive log-space bias head correcting per-document LLM errors via online gradient descent; (4) confidence-based LLM skip for accelerating highly predictable tokens; (5) a hybrid binary format (NC06) extending neural compression to arbitrary binary files--to our knowledge a first among LLM-based compressors; (6) a llama.cpp inference backend achieving ~7x faster single-token decode than PyTorch; (7) parallel multi-GPU compression across up to 8 workers; and (8) native KV cache sliding window reducing per-slide cost by ~37x. The system requires only ~500 MB of GGUF weights and ~1.2 GB VRAM per worker, running on consumer GPUs. On alice29.txt (Canterbury Corpus, 152 KB), Nacrith achieves 0.918 bits per byte (bpb)--outperforming gzip by 3.1x, bzip2 by 2.5x, CMIX v21 by 44%, and ts_zip by 20%, while compressing below the 0th-, 1st-, and 2nd-order byte-level Shannon entropy bounds. On enwik8 (100 MB), Nacrith achieves 0.9389 bpb (11.74%), surpassing ts_zip (~1.11 bpb) by 15% and FineZip (1.024 bpb) by 8% despite using a 60x smaller model with no fine-tuning. An out-of-distribution evaluation on a document published after the model's training cutoff confirms these gains are not memorization artifacts, achieving 0.723 bpb on unseen text.

Token attribution methods provide intuitive explanations for language model outputs by identifying causally important input tokens. However, as modern LLMs increasingly rely on extended reasoning chains, existing schemes face two critical challenges: (1) efficiency bottleneck, where attributing a target span of M tokens within a context of length N requires O(M*N) operations, making long-context attribution prohibitively slow; and (2) faithfulness drop, where intermediate reasoning tokens absorb attribution mass, preventing importance from propagating back to the original input. To address these, we introduce FlashTrace, an efficient multi-token attribution method that employs span-wise aggregation to compute attribution over multi-token targets in a single pass, while maintaining faithfulness. Moreover, we design a recursive attribution mechanism that traces importance through intermediate reasoning chains back to source inputs. Extensive experiments on long-context retrieval (RULER) and multi-step reasoning (MATH, MorehopQA) tasks demonstrate that FlashTrace achieves over 130x speedup over existing baselines while maintaining superior faithfulness. We further analyze the dynamics of recursive attribution, showing that even a single recursive hop improves faithfulness by tracing importance through the reasoning chain.

Extending context window sizes allows large language models (LLMs) to process longer sequences and handle more complex tasks. Rotary Positional Embedding (RoPE) has become the de facto standard due to its relative positional encoding properties that benefit long-context training. However, we observe that using RoPE with BFloat16 format results in numerical issues, causing it to deviate from its intended relative positional encoding, especially in long-context scenarios. This issue arises from BFloat16's limited precision and accumulates as context length increases, with the first token contributing significantly to this problem. To address this, we develop AnchorAttention, a plug-and-play attention method that alleviates numerical issues caused by BFloat16, improves long-context capabilities, and speeds up training. AnchorAttention reduces unnecessary attention computations, maintains semantic coherence, and boosts computational efficiency by treating the first token as a shared anchor with a consistent position ID, making it visible to all documents within the training context. Experiments on three types of LLMs demonstrate that AnchorAttention significantly improves long-context performance and reduces training time by over 50\% compared to standard full attention mechanisms, while preserving the original LLM's capabilities on general tasks. Our code is available at https://github.com/haonan3/AnchorContext.

Training and serving long-context large language models (LLMs) incurs substantial overhead. To address this, two critical steps are often required: a pretrained LLM typically undergoes a separate stage for context length extension by training on long-context data, followed by architectural modifications to reduce the overhead of KV cache during serving. This paper argues that integrating length extension with a GPU-friendly KV cache reduction architecture not only reduces training overhead during length extension, but also achieves better long-context performance. This leads to our proposed LongGen, which finetunes a pretrained LLM into an efficient architecture during length extension. LongGen builds on three key insights: (1) Sparse attention patterns, such as window attention (attending to recent tokens), attention sink (initial ones), and blockwise sparse attention (strided token blocks) are well-suited for building efficient long-context models, primarily due to their GPU-friendly memory access patterns, enabling efficiency gains not just theoretically but in practice as well. (2) It is essential for the model to have direct access to all tokens. A hybrid architecture with 1/3 full attention layers and 2/3 efficient ones achieves a balanced trade-off between efficiency and long-context performance. (3) Lightweight training on 5B long-context data is sufficient to extend the hybrid model's context length from 4K to 128K. We evaluate LongGen on both Llama-2 7B and Llama-2 70B, demonstrating its effectiveness across different scales. During training with 128K-long contexts, LongGen achieves 1.55x training speedup and reduces wall-clock time by 36%, compared to a full-attention baseline. During inference, LongGen reduces KV cache memory by 62%, achieving 1.67x prefilling speedup and 1.41x decoding speedup.

Long context capability is a crucial competency for large language models (LLMs) as it mitigates the human struggle to digest long-form texts. This capability enables complex task-solving scenarios such as book summarization, code assistance, and many more tasks that are traditionally manpower-intensive. However, transformer-based LLMs face significant challenges with long context input due to the growing size of the KV cache and the intrinsic complexity of attending to extended inputs; where multiple schools of efficiency-driven approaches -- such as KV cache quantization, token dropping, prompt compression, linear-time sequence models, and hybrid architectures -- have been proposed to produce efficient yet long context-capable models. Despite these advancements, no existing work has comprehensively benchmarked these methods in a reasonably aligned environment. In this work, we fill this gap by providing a taxonomy of current methods and evaluating 10+ state-of-the-art approaches across seven categories of long context tasks. Our work reveals numerous previously unknown phenomena and offers insights -- as well as a friendly workbench -- for the future development of long context-capable LLMs. The source code will be available at https://github.com/henryzhongsc/longctx_bench

We present Nautile-370M, a 371-million-parameter small language model designed for efficient reasoning under strict parameter and inference budgets. Nautile-370M uses a hybrid backbone in which two SeqCond Attention (SCA) layers, a linear-time spectral sequence operator inspired by SeqCondenser, alternate with one transformer layer. This design aims to retain the long-context efficiency and state-tracking benefits of structured sequential models while preserving the expressive token-to-token routing of attention. The model was trained on a single Cloud TPU v4-64 pod slice provided through the Google TPU Research Cloud (TRC) program; the subsequent reinforcement learning stage was carried out on a single NVIDIA DGX Spark. We prove that the SCA readout mechanism can exactly retrieve any individual token from the prefix summary and can reproduce any output of softmax attention as a special case, establishing that SCA is at least as expressive as full self-attention in the continuous limit. We also describe the training data pipeline and outline a reinforcement learning stage specialized for reasoning, verification, and response quality.

The rapid advancement of Large Language Models (LLMs) has driven the need for more efficient serving strategies. In this context, efficiency refers to the proportion of requests that meet their Service Level Objectives (SLOs), particularly for Time To First Token (TTFT) and Time Between Tokens (TBT). However, existing systems often prioritize one metric at the cost of the other. We present Ascendra, an LLM serving system designed to meet both TTFT and TBT SLOs simultaneously. The core insight behind Ascendra is that a request's urgency evolves as it approaches its deadline. To leverage this, Ascendra partitions GPU resources into two types of instances: low-priority and high-priority. Low-priority instances maximize throughput by processing requests out of arrival order, but at the risk of request starvation. To address this, Ascendra employs a performance model to predict requests at risk of missing their SLOs and proactively offloads them to high-priority instances. High-priority instances are optimized for low-latency execution and handle urgent requests nearing their deadlines. This partitioned architecture enables Ascendra to effectively balance high throughput and low latency. Extensive evaluation shows that Ascendra improves system throughput by up to 1.7x compared to vLLM and Sarathi-Serve while meeting both TTFT and TBT SLOs.

Retrieval Augmented Generation (RAG) systems remain vulnerable to hallucinated answers despite incorporating external knowledge sources. We present LettuceDetect a framework that addresses two critical limitations in existing hallucination detection methods: (1) the context window constraints of traditional encoder-based methods, and (2) the computational inefficiency of LLM based approaches. Building on ModernBERT's extended context capabilities (up to 8k tokens) and trained on the RAGTruth benchmark dataset, our approach outperforms all previous encoder-based models and most prompt-based models, while being approximately 30 times smaller than the best models. LettuceDetect is a token-classification model that processes context-question-answer triples, allowing for the identification of unsupported claims at the token level. Evaluations on the RAGTruth corpus demonstrate an F1 score of 79.22% for example-level detection, which is a 14.8% improvement over Luna, the previous state-of-the-art encoder-based architecture. Additionally, the system can process 30 to 60 examples per second on a single GPU, making it more practical for real-world RAG applications.

Despite remarkable progress in driving world models, their potential for autonomous systems remains largely untapped: the world models are mostly learned for world simulation and decoupled from trajectory planning. While recent efforts aim to unify world modeling and planning in a single framework, the synergistic facilitation mechanism of world modeling for planning still requires further exploration. In this work, we introduce a new driving paradigm named Policy World Model (PWM), which not only integrates world modeling and trajectory planning within a unified architecture, but is also able to benefit planning using the learned world knowledge through the proposed action-free future state forecasting scheme. Through collaborative state-action prediction, PWM can mimic the human-like anticipatory perception, yielding more reliable planning performance. To facilitate the efficiency of video forecasting, we further introduce a dynamically enhanced parallel token generation mechanism, equipped with a context-guided tokenizer and an adaptive dynamic focal loss. Despite utilizing only front camera input, our method matches or exceeds state-of-the-art approaches that rely on multi-view and multi-modal inputs. Code and model weights will be released at https://github.com/6550Zhao/Policy-World-Model.

Structured width pruning of GLU-MLP layers, guided by the Maximum Absolute Weight (MAW) criterion, reveals a systematic dichotomy in how reducing the expansion ratio affects different model capabilities. While performance on tasks relying on parametric knowledge (e.g., MMLU, GSM8K) and perplexity metrics degrades predictably, instruction-following capabilities improve substantially (+46% to +75% in IFEval for Llama-3.2-1B and 3B models), and multi-step reasoning remains robust (MUSR). This pattern challenges the prevailing assumption that pruning induces uniform degradation. We evaluated seven expansion ratio configurations using comprehensive benchmarks assessing factual knowledge, mathematical reasoning, language comprehension, instruction-following, and truthfulness. Our analysis identifies the expansion ratio as a critical architectural parameter that selectively modulates cognitive capabilities, rather than merely serving as a compression metric. We provide the first systematic characterization of this selective preservation phenomenon. Notably, we document a robust inverse correlation (r = -0.864, p = 0.012 in Llama-3B) between factual knowledge capacity (MMLU) and truthfulness metrics (TruthfulQA-MC2): as knowledge degrades, the model's ability to discriminate misconceptions improves consistently. This connects two previously distinct research areas, demonstrating that MAW-guided width pruning acts as a selective filter, reducing parametric knowledge while preserving or enhancing behavioral alignment. Additionally, we quantify context-dependent efficiency trade-offs: pruned configurations achieve up to 23% reduction in energy consumption (J/token) but incur penalties in single-request latency, whereas batch processing workloads benefit uniformly.

Long-context inference in LLMs faces the dual challenges of quadratic attention complexity and prohibitive KV cache memory. While token-level sparse attention offers superior accuracy, its indexing overhead is costly; block-level methods improve efficiency but sacrifice precision. We propose AsyncTLS, a hierarchical sparse attention system that combines coarse-grained block filtering with fine-grained token selection to balance accuracy and efficiency, coupled with an asynchronous offloading engine that overlaps KV cache transfers with computation via temporal locality exploitation. Evaluated on Qwen3 and GLM-4.7-Flash across GQA, and MLA architectures, AsyncTLS achieves accuracy comparable to full attention while delivering 1.2x - 10.0x operator speedups and 1.3x - 4.7x end-to-end throughput improvements on 48k - 96k contexts.

The long-context capability of recent large transformer models can be surmised to rely on techniques such as attention/model parallelism, as well as hardware-level optimizations. While these strategies allow input lengths to scale to millions of tokens, they do not fundamentally mitigate the quadratic computational and memory complexity of the core attention mechanism. In this paper, we address the challenge of long-context processing on a single device using generic implementations by reducing the quadratic memory footprint and inference cost. Existing approaches to extend the context length for generic single device implementations -- such as token merging and modified attentions -- are often modality specific and attain a suboptimal tradeoff between accuracy and efficiency. To overcome these limitations, we propose MAELRE (Modality Agnostic Efficient Long Range Encoder), a unified and efficient transformer architecture designed for long-range encoding across diverse modalities. MAELRE integrates token merging with attention approximation, progressively merging tokens at different stages of internal computational blocks. It employs a lightweight attention approximation when the number of tokens is large, and switches to standard dot-product attention as the sequence becomes shorter through successive aggregation. We demonstrate that MAELRE achieves superior accuracy while reducing computational cost compared to existing long-context models on classification tasks spanning multiple modalities, including text, time series, audio, and vision.

Code auditing is the process of reviewing code with the aim of identifying bugs. Large Language Models (LLMs) have demonstrated promising capabilities for this task without requiring compilation, while also supporting user-friendly customization. However, auditing a code repository with LLMs poses significant challenges: limited context windows and hallucinations can degrade the quality of bug reports, and analyzing large-scale repositories incurs substantial time and token costs, hindering efficiency and scalability. This work introduces an LLM-based agent, RepoAudit, designed to perform autonomous repository-level code auditing. Equipped with agent memory, RepoAudit explores the codebase on demand by analyzing data-flow facts along feasible program paths within individual functions. It further incorporates a validator module to mitigate hallucinations by verifying data-flow facts and checking the satisfiability of path conditions associated with potential bugs, thereby reducing false positives. RepoAudit detects 40 true bugs across 15 real-world benchmark projects with a precision of 78.43%, requiring on average only 0.44 hours and $2.54 per project. Also, it detects 185 new bugs in high-profile projects, among which 174 have been confirmed or fixed. We have open-sourced RepoAudit at https://github.com/PurCL/RepoAudit.

Most language model pre-training frameworks concatenate multiple documents into fixed-length sequences and use causal masking to compute the likelihood of each token given its context; this strategy is widely adopted due to its simplicity and efficiency. However, to this day, the influence of the pre-training sequence composition strategy on the generalisation properties of the model remains under-explored. In this work, we find that applying causal masking can lead to the inclusion of distracting information from previous documents during pre-training, which negatively impacts the performance of the models on language modelling and downstream tasks. In intra-document causal masking, the likelihood of each token is only conditioned on the previous tokens in the same document, eliminating potential distracting information from previous documents and significantly improving performance. Furthermore, we find that concatenating related documents can reduce some potential distractions during pre-training, and our proposed efficient retrieval-based sequence construction method, BM25Chunk, can improve in-context learning (+11.6\%), knowledge memorisation (+9.8\%), and context utilisation (+7.2\%) abilities of language models without sacrificing efficiency.

Drug discovery is a complex process that involves multiple scenarios and stages, such as fragment-constrained molecule generation, hit generation and lead optimization. However, existing molecular generative models can only tackle one or two of these scenarios and lack the flexibility to address various aspects of the drug discovery pipeline. In this paper, we present Generalist Molecular generative model (GenMol), a versatile framework that addresses these limitations by applying discrete diffusion to the Sequential Attachment-based Fragment Embedding (SAFE) molecular representation. GenMol generates SAFE sequences through non-autoregressive bidirectional parallel decoding, thereby allowing utilization of a molecular context that does not rely on the specific token ordering and enhanced computational efficiency. Moreover, under the discrete diffusion framework, we introduce fragment remasking, a strategy that optimizes molecules by replacing fragments with masked tokens and regenerating them, enabling effective exploration of chemical space. GenMol significantly outperforms the previous GPT-based model trained on SAFE representations in de novo generation and fragment-constrained generation, and achieves state-of-the-art performance in goal-directed hit generation and lead optimization. These experimental results demonstrate that GenMol can tackle a wide range of drug discovery tasks, providing a unified and versatile approach for molecular design.

In this report, we introduce Falcon-H1, a new series of large language models (LLMs) featuring hybrid architecture designs optimized for both high performance and efficiency across diverse use cases. Unlike earlier Falcon models built solely on Transformer or Mamba architectures, Falcon-H1 adopts a parallel hybrid approach that combines Transformer-based attention with State Space Models (SSMs), known for superior long-context memory and computational efficiency. We systematically revisited model design, data strategy, and training dynamics, challenging conventional practices in the field. Falcon-H1 is released in multiple configurations, including base and instruction-tuned variants at 0.5B, 1.5B, 1.5B-deep, 3B, 7B, and 34B parameters. Quantized instruction-tuned models are also available, totaling over 30 checkpoints on Hugging Face Hub. Falcon-H1 models demonstrate state-of-the-art performance and exceptional parameter and training efficiency. The flagship Falcon-H1-34B matches or outperforms models up to 70B scale, such as Qwen3-32B, Qwen2.5-72B, and Llama3.3-70B, while using fewer parameters and less data. Smaller models show similar trends: the Falcon-H1-1.5B-Deep rivals current leading 7B-10B models, and Falcon-H1-0.5B performs comparably to typical 7B models from 2024. These models excel across reasoning, mathematics, multilingual tasks, instruction following, and scientific knowledge. With support for up to 256K context tokens and 18 languages, Falcon-H1 is suitable for a wide range of applications. All models are released under a permissive open-source license, underscoring our commitment to accessible and impactful AI research.

Human-agent dialogues often exhibit topic continuity-a stable thematic frame that evolves through temporally adjacent exchanges-yet most large language model (LLM) agent memory systems fail to preserve it. Existing designs follow a fragmentation-compensation paradigm: they first break dialogue streams into isolated utterances for storage, then attempt to restore coherence via embedding-based retrieval. This process irreversibly damages narrative and causal flow, while biasing retrieval towards lexical similarity. We introduce membox, a hierarchical memory architecture centered on a Topic Loom that continuously monitors dialogue in a sliding-window fashion, grouping consecutive same-topic turns into coherent "memory boxes" at storage time. Sealed boxes are then linked by a Trace Weaver into long-range event-timeline traces, recovering macro-topic recurrences across discontinuities. Experiments on LoCoMo demonstrate that Membox achieves up to 68% F1 improvement on temporal reasoning tasks, outperforming competitive baselines (e.g., Mem0, A-MEM). Notably, Membox attains these gains while using only a fraction of the context tokens required by existing methods, highlighting a superior balance between efficiency and effectiveness. By explicitly modeling topic continuity, Membox offers a cognitively motivated mechanism for enhancing both coherence and efficiency in LLM agents.

Large-scale fine-grained image retrieval (FGIR) aims to retrieve images belonging to the same subcategory as a given query by capturing subtle differences in a large-scale setting. Recently, Vision Transformers (ViT) have been employed in FGIR due to their powerful self-attention mechanism for modeling long-range dependencies. However, most Transformer-based methods focus primarily on leveraging self-attention to distinguish fine-grained details, while overlooking the high computational complexity and redundant dependencies inherent to these models, limiting their scalability and effectiveness in large-scale FGIR. In this paper, we propose an Efficient and Effective ViT-based framework, termed EET, which integrates token pruning module with a discriminative transfer strategy to address these limitations. Specifically, we introduce a content-based token pruning scheme to enhance the efficiency of the vanilla ViT, progressively removing background or low-discriminative tokens at different stages by exploiting feature responses and self-attention mechanism. To ensure the resulting efficient ViT retains strong discriminative power, we further present a discriminative transfer strategy comprising both discriminative knowledge transfer and discriminative region guidance. Using a distillation paradigm, these components transfer knowledge from a larger ``teacher'' ViT to a more efficient ``student'' model, guiding the latter to focus on subtle yet crucial regions in a cost-free manner. Extensive experiments on two widely-used fine-grained datasets and four large-scale fine-grained datasets demonstrate the effectiveness of our method. Specifically, EET reduces the inference latency of ViT-Small by 42.7\% and boosts the retrieval performance of 16-bit hash codes by 5.15\% on the challenging NABirds dataset.

We propose a novel end-to-end document understanding model called SeRum (SElective Region Understanding Model) for extracting meaningful information from document images, including document analysis, retrieval, and office automation. Unlike state-of-the-art approaches that rely on multi-stage technical schemes and are computationally expensive, SeRum converts document image understanding and recognition tasks into a local decoding process of the visual tokens of interest, using a content-aware token merge module. This mechanism enables the model to pay more attention to regions of interest generated by the query decoder, improving the model's effectiveness and speeding up the decoding speed of the generative scheme. We also designed several pre-training tasks to enhance the understanding and local awareness of the model. Experimental results demonstrate that SeRum achieves state-of-the-art performance on document understanding tasks and competitive results on text spotting tasks. SeRum represents a substantial advancement towards enabling efficient and effective end-to-end document understanding.

This paper studies contextual biasing with Large Language Models (LLMs), where during second-pass rescoring additional contextual information is provided to a LLM to boost Automatic Speech Recognition (ASR) performance. We propose to leverage prompts for a LLM without fine tuning during rescoring which incorporate a biasing list and few-shot examples to serve as additional information when calculating the score for the hypothesis. In addition to few-shot prompt learning, we propose multi-task training of the LLM to predict both the entity class and the next token. To improve the efficiency for contextual biasing and to avoid exceeding LLMs' maximum sequence lengths, we propose dynamic prompting, where we select the most likely class using the class tag prediction, and only use entities in this class as contexts for next token prediction. Word Error Rate (WER) evaluation is performed on i) an internal calling, messaging, and dictation dataset, and ii) the SLUE-Voxpopuli dataset. Results indicate that biasing lists and few-shot examples can achieve 17.8% and 9.6% relative improvement compared to first pass ASR, and that multi-task training and dynamic prompting can achieve 20.0% and 11.3% relative WER improvement, respectively.

Processing and reasoning over long contexts is crucial for many practical applications of Large Language Models (LLMs), such as document comprehension and agent construction. Despite recent strides in making LLMs process contexts with more than 100K tokens, there is currently a lack of a standardized benchmark to evaluate this long-context capability. Existing public benchmarks typically focus on contexts around 10K tokens, limiting the assessment and comparison of LLMs in processing longer contexts. In this paper, we propose inftyBench, the first LLM benchmark featuring an average data length surpassing 100K tokens. inftyBench comprises synthetic and realistic tasks spanning diverse domains, presented in both English and Chinese. The tasks in inftyBench are designed to require well understanding of long dependencies in contexts, and make simply retrieving a limited number of passages from contexts not sufficient for these tasks. In our experiments, based on inftyBench, we evaluate the state-of-the-art proprietary and open-source LLMs tailored for processing long contexts. The results indicate that existing long context LLMs still require significant advancements to effectively process 100K+ context. We further present three intriguing analyses regarding the behavior of LLMs processing long context.

Video Large Language Models (VLLMs) demonstrate strong video understanding but suffer from inefficiency due to redundant visual tokens. Existing pruning primary targets intra-frame spatial redundancy or prunes inside the LLM with shallow-layer overhead, yielding suboptimal spatiotemporal reduction and underutilizing long-context compressibility. All of them often discard subtle yet informative context from merged or pruned tokens. In this paper, we propose a new perspective that elaborates token Anchors within intra-frame and inter-frame to comprehensively aggregate the informative contexts via local-global Optimal Transport (AOT). Specifically, we first establish local- and global-aware token anchors within each frame under the attention guidance, which then optimal transport aggregates the informative contexts from pruned tokens, constructing intra-frame token anchors. Then, building on the temporal frame clips, the first frame within each clip will be considered as the keyframe anchors to ensemble similar information from consecutive frames through optimal transport, while keeping distinct tokens to represent temporal dynamics, leading to efficient token reduction in a training-free manner. Extensive evaluations show that our proposed AOT obtains competitive performances across various short- and long-video benchmarks on leading video LLMs, obtaining substantial computational efficiency while preserving temporal and visual fidelity. Project webpage: https://tyroneli.github.io/AOT{AOT}.

Large Language Models (LLMs) based autonomous agents demonstrate multifaceted capabilities to contribute substantially to economic production. However, existing benchmarks remain focused on single agentic capability, failing to capture long-horizon real-world scenarios. Moreover, the reliance on human-in-the-loop feedback for realistic tasks creates a scalability bottleneck, hindering automated rollout collection and evaluation. To bridge this gap, we introduce AgencyBench, a comprehensive benchmark derived from daily AI usage, evaluating 6 core agentic capabilities across 32 real-world scenarios, comprising 138 tasks with specific queries, deliverables, and rubrics. These scenarios require an average of 90 tool calls, 1 million tokens, and hours of execution time to resolve. To enable automated evaluation, we employ a user simulation agent to provide iterative feedback, and a Docker sandbox to conduct visual and functional rubric-based assessment. Experiments reveal that closed-source models significantly outperform open-source models (48.4% vs 32.1%). Further analysis reveals significant disparities across models in resource efficiency, feedback-driven self-correction, and specific tool-use preferences. Finally, we investigate the impact of agentic scaffolds, observing that proprietary models demonstrate superior performance within their native ecosystems (e.g., Claude-4.5-Opus via Claude-Agent-SDK), while open-source models exhibit distinct performance peaks, suggesting potential optimization for specific execution frameworks. AgencyBench serves as a critical testbed for next-generation agents, highlighting the necessity of co-optimizing model architecture with agentic frameworks. We believe this work sheds light on the future direction of autonomous agents, and we release the full benchmark and evaluation toolkit at https://github.com/GAIR-NLP/AgencyBench.

State Space Models (SSMs) have emerged as a promising alternative to Transformers for long-context sequence modeling, offering linear O(N) computational complexity compared to the Transformer's quadratic O(N^2) scaling. This paper presents a comprehensive benchmarking study comparing the Mamba SSM against the LLaMA Transformer on long-context sequences, using dyadic therapy sessions as a representative test case. We evaluate both architectures across two dimensions: (1) computational efficiency, where we measure memory usage and inference speed from 512 to 8,192 tokens, and (2) representational efficiency, where we analyze hidden state dynamics and attention patterns. Our findings provide actionable insights for practitioners working with long-context applications, establishing precise conditions under which SSMs offer advantages over Transformers.

Large Language Models (LLMs) have revolutionized natural language processing, excelling in handling longer sequences. However, the inefficiency and redundancy in processing extended in-context tokens remain a challenge. Many attempts to address this rely on compressing tokens with smaller text encoders, yet we question whether text encoders are truly indispensable. Our journey leads to an unexpected discovery-a much smaller vision encoder, applied directly to sequences of text tokens, can rival text encoders on text tasks. When pre-trained on large amounts of data and transferred to multiple mid-sized or small text understanding benchmarks, VIST leads to comparable results with 16% fewer FLOPs and 50% less memory usage. We further uncover significant token redundancy and devise a frequency-based masking strategy to guide the focus of the visual encoder toward the most critical tokens. Interestingly, we observe the trained visual encoder performs like a summarizer, selectively ignoring less important words such as prepositions and conjunctions. This approach delivers remarkable results, outperforming traditional text encoder-based methods by 5.7% on average over benchmarks like TriviaQA, NQ, PopQA, TREF, SST2, and SST5, setting a new standard for token efficiency in LLMs.

Long context inference presents challenges at the system level with increased compute and memory requirements, as well as from an accuracy perspective in being able to reason over long contexts. Recently, several methods have been proposed to compress the prompt to reduce the context length. However, there has been little work on comparing the different proposed methods across different tasks through a standardized analysis. This has led to conflicting results. To address this, here we perform a comprehensive characterization and evaluation of different prompt compression methods. In particular, we analyze extractive compression, summarization-based abstractive compression, and token pruning methods. Surprisingly, we find that extractive compression often outperforms all the other approaches, and enables up to 10x compression with minimal accuracy degradation. Interestingly, we also find that despite several recent claims, token pruning methods often lag behind extractive compression. We only found marginal improvements on summarization tasks.

The proliferation of long-context large language models (LLMs) exposes a key bottleneck: the rapidly expanding key-value cache during decoding, which imposes heavy memory and latency costs. While recent approaches attempt to alleviate this by sharing a single set of crucial tokens across layers, such coarse-grained sharing undermines model performance by neglecting the functional diversity of attention heads. To address this, we propose LycheeDecode, an efficient decoding method centered on a fine-grained hybrid-head attention mechanism that employs a hardware-efficient top-k selection strategy. Specifically, the novel HardKuma-based mechanism partitions attention heads into a small subset of retrieval heads that dynamically identify crucial tokens and a majority of sparse heads that reuse them for efficient computation. Through extensive experiments on leading models like Llama3 and Qwen3 across diverse benchmarks for long-context understanding (e.g., LongBench, RULER) and complex reasoning (e.g., AIME24, OlympiadBench), we demonstrate that LycheeDecode achieves generative quality comparable to, and at times surpassing even the full-attention baseline. Crucially, this is accomplished with up to a 2.7x speedup at a 128K context length. By preserving the functional diversity of attention heads, our fine-grained strategy overcomes the performance bottlenecks of existing methods, providing a powerful and validated pathway to both efficient and high-quality long-context LLM inference.

High-quality main content extraction from web pages is a critical prerequisite for constructing large-scale training corpora. While traditional heuristic extractors are efficient, they lack the semantic reasoning required to handle the structural heterogeneity of the modern web. Conversely, well-pretrained generative Large Language Models (LLMs) offer superior document comprehension but are prohibited by excessive computational costs, limited context windows, and hallucination risks when applied at web scale. We present Dripper, a lightweight framework that resolves these bottlenecks through four contributions: (1) We reformulate extraction as a constrained sequence labeling task using SLMs (Small Language Models). This paradigm eliminates generative hallucinations and achieves exceptional efficiency, reaching a throughput of 3.08 pages per second on a single A100 GPU. (2) We construct WebMainBench, a rigorous benchmark of 7,809 human-annotated pages covering 5,434 unique domains and multiple languages. Evaluations show our Dripper-0.6B model outperforms heuristics like Trafilatura and rivals massive models like DeepSeek-V3.2(685B), GPT-5 and Gemini-2.5-Pro, offering an optimal efficiency-accuracy trade-off. (3) We demonstrate infrastructural value by pre-training a 1B model on a Dripper-curated corpus (63B tokens). This model significantly outperforms baselines in downstream tasks, proving the critical role of extraction quality and the effectiveness of our framework. (4) We open-source the Dripper-0.6B weights and codebase to facilitate the construction of high-quality datasets.

Molecular understanding is central to advancing areas such as scientific discovery, yet Large Language Models (LLMs) struggle to understand molecular graphs effectively. Existing graph-LLM bridges often adapt the Q-Former-style connector with fixed-length static tokens, which is originally designed for vision tasks. These designs overlook stereochemistry and substructural context and typically require costly LLM-backbone fine-tuning, limiting efficiency and generalization. We introduce EDT-Former, an Entropy-guided Dynamic Token Transformer that generates tokens aligned with informative molecular patches, thereby preserving both local and global structural features for molecular graph understanding. Beyond prior approaches, EDT-Former enables alignment between frozen graph encoders and LLMs without tuning the LLM backbone (excluding the embedding layer), resulting in computationally efficient finetuning, and achieves stateof-the-art results on MoleculeQA, Molecule-oriented Mol-Instructions, and property prediction benchmarks (TDC, MoleculeNet), underscoring its effectiveness for scalable and generalizable multimodal molecular understanding

Retrieval-Augmented Generation (RAG) utilizes external knowledge to augment Large Language Models' (LLMs) reliability. For flexibility, agentic RAG employs autonomous, multi-round retrieval and reasoning to resolve queries. Although recent agentic RAG has improved via reinforcement learning, they often incur substantial token overhead from search and reasoning processes. This trade-off prioritizes accuracy over efficiency. To address this issue, this work proposes TeaRAG, a token-efficient agentic RAG framework capable of compressing both retrieval content and reasoning steps. 1) First, the retrieved content is compressed by augmenting chunk-based semantic retrieval with a graph retrieval using concise triplets. A knowledge association graph is then built from semantic similarity and co-occurrence. Finally, Personalized PageRank is leveraged to highlight key knowledge within this graph, reducing the number of tokens per retrieval. 2) Besides, to reduce reasoning steps, Iterative Process-aware Direct Preference Optimization (IP-DPO) is proposed. Specifically, our reward function evaluates the knowledge sufficiency by a knowledge matching mechanism, while penalizing excessive reasoning steps. This design can produce high-quality preference-pair datasets, supporting iterative DPO to improve reasoning conciseness. Across six datasets, TeaRAG improves the average Exact Match by 4% and 2% while reducing output tokens by 61% and 59% on Llama3-8B-Instruct and Qwen2.5-14B-Instruct, respectively. Code is available at https://github.com/Applied-Machine-Learning-Lab/TeaRAG.

While significant advancements have been made in compressed representations for text embeddings in large language models (LLMs), the compression of visual tokens in large multi-modal models (LMMs) has remained a largely overlooked area. In this work, we present the study on the analysis of redundancy concerning visual tokens and efficient training within these models. Our initial experiments show that eliminating up to 70% of visual tokens at the testing stage by simply average pooling only leads to a minimal 3% reduction in visual question answering accuracy on the GQA benchmark, indicating significant redundancy in visual context. Addressing this, we introduce Visual Context Compressor, which reduces the number of visual tokens during training to enhance training efficiency without sacrificing performance. To minimize information loss caused by the compression on visual tokens while maintaining training efficiency, we develop LLaVolta as a lite training scheme. LLaVolta incorporates stage-wise visual context compression to progressively compress the visual tokens from heavily to lightly, and finally no compression at the end of training, yielding no loss of information when testing. Extensive experiments demonstrate that our approach enhances the performance of MLLMs in both image-language and video-language understanding, while also significantly cutting training costs. Code is available at https://github.com/Beckschen/LLaVolta

Large Language Models (LLMs) face significant computational challenges when processing long contexts due to the quadratic complexity of self-attention. While soft context compression methods, which map input text to smaller latent representations, have shown promise, their real-world adoption is limited. Existing techniques typically compress the context as a single unit, which leads to quadratic compression complexity and an inability to reuse computations across queries with overlapping contexts. In this work, we introduce CompLLM, a soft compression technique designed for practical deployment. Instead of processing the context holistically, CompLLM divides it into segments and compresses each one independently. This simple design choice yields three critical properties: efficiency, as the compression step scales linearly with the context length; scalability, enabling models trained on short sequences (e.g., 1k tokens) to generalize to contexts of 100k tokens; and reusability, allowing compressed segments to be cached and reused across different queries. Our experiments show that with a 2x compression rate, at high context lengths CompLLM speeds up Time To First Token (TTFT) by up to 4x and reduces the KV cache size by 50%. Furthermore, CompLLM achieves performance comparable to that obtained with the uncompressed context, and even surpasses it on very long sequences, demonstrating its effectiveness and practical utility.

Model-based reinforcement learning agents utilizing transformers have shown improved sample efficiency due to their ability to model extended context, resulting in more accurate world models. However, for complex reasoning and planning tasks, these methods primarily rely on continuous representations. This complicates modeling of discrete properties of the real world such as disjoint object classes between which interpolation is not plausible. In this work, we introduce discrete abstract representations for transformer-based learning (DART), a sample-efficient method utilizing discrete representations for modeling both the world and learning behavior. We incorporate a transformer-decoder for auto-regressive world modeling and a transformer-encoder for learning behavior by attending to task-relevant cues in the discrete representation of the world model. For handling partial observability, we aggregate information from past time steps as memory tokens. DART outperforms previous state-of-the-art methods that do not use look-ahead search on the Atari 100k sample efficiency benchmark with a median human-normalized score of 0.790 and beats humans in 9 out of 26 games. We release our code at https://pranaval.github.io/DART/.

Existing approaches for improving the efficiency of Large Vision-Language Models (LVLMs) are largely based on the concept of visual token reduction. This approach, however, creates an information bottleneck that impairs performance, especially on challenging tasks that require fine-grained understanding and reasoning. In this work, we challenge this paradigm by introducing VISion On Request (VISOR), a method that reduces inference cost without discarding visual information. Instead of compressing the image, VISOR improves efficiency by sparsifying the interaction between image and text tokens. Specifically, the language model attends to the full set of high-resolution visual tokens through a small, strategically placed set of attention layers: general visual context is provided by efficient cross-attention between text-image, while a few well-placed and dynamically selected self-attention layers refine the visual representations themselves, enabling complex, high-resolution reasoning when needed. Based on this principle, we first train a single universal network on a range of computational budgets by varying the number of self-attention layers, and then introduce a lightweight policy mechanism that dynamically allocates visual computation based on per-sample complexity. Extensive experiments show that VISOR drastically reduces computational cost while matching or exceeding state-of-the-art results across a diverse suite of benchmarks, and excels in challenging tasks that require detailed visual understanding.

Retrieval-augmented generation (RAG) for long-context question answering is bottlenecked by inference-time prefilling over large retrieved contexts. A common strategy is to precompute key-value (KV) caches for individual documents and selectively recompute a small subset of tokens to restore global causal dependencies, but existing methods rely on heuristics or representation discrepancies without modeling whether selected tokens can effectively influence generation. We cast selective KV recomputation as an information flow problem and show that a simple attention-norm signal from the query reliably identifies tokens that are both semantically relevant and structurally positioned to propagate information, when computed under an inference-consistent RoPE geometry. We therefore reconstruct global positional assignments for retrieved chunks and introduce an information-flow-guided chunk reordering strategy. Experiments on LLM and VLM benchmarks demonstrate consistent gains over prior methods under comparable efficiency budgets.

Large Language Models (LLMs) face significant challenges in long-context processing, including quadratic computational costs, information forgetting, and the context fragmentation inherent in retrieval-augmented generation (RAG). We propose a cognitively inspired framework for efficient long-context inference based on chunk-wise compression and selective memory recall, rather than processing all raw tokens. The framework segments long inputs into chunks and encodes each chunk into compressed memory representations using a learned compressor. A gating module dynamically selects relevant memory blocks, which are then iteratively processed by a reasoning module with an evolving working memory to solve downstream tasks. The compressor and reasoner are jointly optimized via end-to-end reinforcement learning, while the gating module is trained separately as a classifier. Experimental results show that the proposed method achieves competitive accuracy on multi-hop reasoning benchmarks such as RULER-HQA, extrapolates context length from 7K to 1.75M tokens, and offers a favorable accuracy-efficiency trade-off compared to strong long-context baselines. In particular, it achieves up to a 2 times reduction in peak GPU memory usage and a 6 times inference speedup over MemAgent.

Recent advances in large language models (LLMs) have showcased exceptional performance in long-context tasks, while facing significant inference efficiency challenges with limited GPU memory. Existing solutions first proposed the sliding-window approach to accumulate a set of historical key-value (KV) pairs for reuse, then further improvements selectively retain its subsets at each step. However, due to the sparse attention distribution across a long context, it is hard to identify and recall relevant KV pairs, as the attention is distracted by massive candidate pairs. Additionally, we found it promising to select representative tokens as probe-Query in each sliding window to effectively represent the entire context, which is an approach overlooked by existing methods. Thus, we propose ActQKV, a training-free, Activation-aware approach that dynamically determines probe-Query and leverages it to retrieve the relevant KV pairs for inference. Specifically, ActQKV monitors a token-level indicator, Activation Bias, within each context window, enabling the proper construction of probe-Query for retrieval at pre-filling stage. To accurately recall the relevant KV pairs and minimize the irrelevant ones, we design a dynamic KV cut-off mechanism guided by information density across layers at the decoding stage. Experiments on the Long-Bench and infty Benchmarks demonstrate its state-of-the-art performance with competitive inference quality and resource efficiency.

Large language models (LLMs) owe much of their stellar performance to expansive input contexts, yet such verbosity inflates monetary costs, carbon footprint, and inference-time latency. Much of this overhead manifests from the redundant low-utility tokens present in typical prompts, as only a fraction of tokens typically carries the majority of the semantic weight. We address this inefficiency by introducing FrugalPrompt, a novel prompt compression framework for LLMs, which retains only the most semantically significant tokens. Leveraging two state-of-the-art token attribution methods, GlobEnc and DecompX, we assign salience scores to every token in an input sequence, rank them to preserve the top-k% tokens in their original order, and obtain a sparse frugalized prompt. We evaluate the approach across four NLP tasks: Sentiment Analysis, Commonsense QA, Summarization, and Mathematical Reasoning, using a suite of frontier LLMs. For the first three tasks, a 20% prompt reduction incurs only a marginal loss in task performance, demonstrating that contemporary LLMs can reconstruct elided context from high-salience cues. In contrast, performance on mathematical reasoning deteriorates sharply, reflecting a stronger dependence on complete token continuity. Further analysis with bottom-k% and random-k% tokens reveals asymmetric performance patterns that may suggest potential task contamination effects, wherein models may resort to shallow memorized patterns from pretraining exposure for conventional NLP tasks. We posit that our work contributes to a more nuanced understanding of LLM behavior in performance-efficiency trade-offs, and delineate the boundary between tasks tolerant to contextual sparsity and those requiring exhaustive context. Our source code and models are available at: https://github.com/Starscream-11813/Frugal-ICL.

As web agents (e.g., Deep Research) routinely consume massive volumes of web pages to gather and analyze information, LLM context management -- under large token budgets and low signal density -- emerges as a foundational, high-importance, and technically challenging problem for agentic and RAG pipelines. Existing solutions for extracting relevant content are inadequate: generative extraction models suffer from high latency, rule-based heuristics lack adaptability, and chunk-and-rerank methods are blind to webpage structure. To overcome these issues, we introduce Index-based Web Content Extraction to reframe the extraction process from slow, token-by-token generation into a highly efficient, discriminative task of index prediction, achieving both effectiveness and efficiency. We partition HTML into structure-aware, addressable segments, and extract only the positional indices of content relevant to a given query. This method decouples extraction latency from content length, enabling rapid, query-relevant extraction. We first evaluate our method as a post-retrieval processing component within an RAG QA system and find that it improves QA accuracy. Then we directly measure its match rate with the target content in two scenarios: main content extraction (ME) and query-relevant extraction (QE). Experimental results show that our method outperforms existing works in both accuracy and speed, effectively bridging the gap between LLMs and the vast webpages.

"Compression Tells Intelligence", is supported by research in artificial intelligence, particularly concerning (multimodal) large language models (LLMs/MLLMs), where compression efficiency often correlates with improved model performance and capabilities. For compression, classical visual coding based on traditional information theory has developed over decades, achieving great success with numerous international industrial standards widely applied in multimedia (e.g., image/video) systems. Except that, the recent emergingvisual token technology of generative multi-modal large models also shares a similar fundamental objective like visual coding: maximizing semantic information fidelity during the representation learning while minimizing computational cost. Therefore, this paper provides a comprehensive overview of two dominant technique families first -- Visual Coding and Vision Token Technology -- then we further unify them from the aspect of optimization, discussing the essence of compression efficiency and model performance trade-off behind. Next, based on the proposed unified formulation bridging visual coding andvisual token technology, we synthesize bidirectional insights of themselves and forecast the next-gen visual codec and token techniques. Last but not least, we experimentally show a large potential of the task-oriented token developments in the more practical tasks like multimodal LLMs (MLLMs), AI-generated content (AIGC), and embodied AI, as well as shedding light on the future possibility of standardizing a general token technology like the traditional codecs (e.g., H.264/265) with high efficiency for a wide range of intelligent tasks in a unified and effective manner.

A unique aspect of ColBERT is its use of [MASK] tokens in queries to score documents (query augmentation). Prior work shows [MASK] tokens weighting non-[MASK] query terms, emphasizing certain tokens over others , rather than introducing whole new terms as initially proposed. We begin by demonstrating that a term weighting behavior previously reported for [MASK] tokens in ColBERTv1 holds for ColBERTv2. We then examine the effect of changing the number of [MASK] tokens from zero to up to four times past the query input length used in training, both for first stage retrieval, and for scoring candidates, observing an initial decrease in performance with few [MASK]s, a large increase when enough [MASK]s are added to pad queries to an average length of 32, then a plateau in performance afterwards. Additionally, we compare baseline performance to performance when the query length is extended to 128 tokens, and find that differences are small (e.g., within 1% on various metrics) and generally statistically insignificant, indicating performance does not collapse if ColBERT is presented with more [MASK] tokens than expected.

Large language model (LLM) providers boast big numbers for maximum context window sizes. To test the real world use of context windows, we 1) define a concept of maximum effective context window, 2) formulate a testing method of a context window's effectiveness over various sizes and problem types, and 3) create a standardized way to compare model efficacy for increasingly larger context window sizes to find the point of failure. We collected hundreds of thousands of data points across several models and found significant differences between reported Maximum Context Window (MCW) size and Maximum Effective Context Window (MECW) size. Our findings show that the MECW is, not only, drastically different from the MCW but also shifts based on the problem type. A few top of the line models in our test group failed with as little as 100 tokens in context; most had severe degradation in accuracy by 1000 tokens in context. All models fell far short of their Maximum Context Window by as much as 99 percent. Our data reveals the Maximum Effective Context Window shifts based on the type of problem provided, offering clear and actionable insights into how to improve model accuracy and decrease model hallucination rates.

As the context limits of Large Language Models (LLMs) increase, the range of possible applications and downstream functions broadens. In many real-world tasks, decisions depend on details scattered across collections of often disparate documents containing mostly irrelevant information. Long-context LLMs appear well-suited to this form of complex information retrieval and reasoning, which has traditionally proven costly and time-consuming. However, although the development of longer context models has seen rapid gains in recent years, our understanding of how effectively LLMs use their context has not kept pace. To address this, we conduct a set of retrieval experiments designed to evaluate the capabilities of 17 leading LLMs, such as their ability to follow threads of information through the context window. Strikingly, we find that many models are remarkably threadsafe: capable of simultaneously following multiple threads without significant loss in performance. Still, for many models, we find the effective context limit is significantly shorter than the supported context length, with accuracy decreasing as the context window grows. Our study also highlights the important point that token counts from different tokenizers should not be directly compared -- they often correspond to substantially different numbers of written characters. We release our code and long-context experimental data.

The versatility of self-attention mechanism earned transformers great success in almost all data modalities, with limitations on the quadratic complexity and difficulty of training. To apply transformers across different data modalities, practitioners have to make specific clever data-modality-dependent constructions. In this paper, we propose Sampling Foundational Transformer (SFT) that can work on multiple data modalities (e.g., point cloud, graph, and sequence) and constraints (e.g., rotational-invariant). The existence of such model is important as contemporary foundational modeling requires operability on multiple data sources. For efficiency on large number of tokens, our model relies on our context aware sampling-without-replacement mechanism for both linear asymptotic computational complexity and real inference time gain. For efficiency, we rely on our newly discovered pseudoconvex formulation of transformer layer to increase model's convergence rate. As a model working on multiple data modalities, SFT has achieved competitive results on many benchmarks, while being faster in inference, compared to other very specialized models.

Diffusion Transformers process images as fixed-length sequences of tokens produced by a static patchify operation. While effective, this design spends uniform compute on low- and high-information regions alike, ignoring that images contain regions of varying detail and that the denoising process progresses from coarse structure at early timesteps to fine detail at late timesteps. We introduce the Dynamic Chunking Diffusion Transformer (DC-DiT), which augments the DiT backbone with a learned encoder-router-decoder scaffold that adaptively compresses the 2D input into a shorter token sequence in a data-dependent manner using a chunking mechanism learned end-to-end with diffusion training. The mechanism learns to compress uniform background regions into fewer tokens and detail-rich regions into more tokens, with meaningful visual segmentations emerging without explicit supervision. Furthermore, it also learns to adapt its compression across diffusion timesteps, using fewer tokens at noisy stages and more tokens as fine details emerge. On class-conditional ImageNet 256{times}256, DC-DiT consistently improves FID and Inception Score over both parameter-matched and FLOP-matched DiT baselines across 4{times} and 16{times} compression, showing this is a promising technique with potential further applications to pixel-space, video and 3D generation. Beyond accuracy, DC-DiT is practical: it can be upcycled from pretrained DiT checkpoints with minimal post-training compute (up to 8{times} fewer training steps) and composes with other dynamic computation methods to further reduce generation FLOPs.

Large Language Models for code often entail significant computational complexity, which grows significantly with the length of the input code sequence. We propose LeanCode for code simplification to reduce training and prediction time, leveraging code contexts in utilizing attention scores to represent the tokens' importance. We advocate for the selective removal of tokens based on the average context-aware attention scores rather than average scores across all inputs. LeanCode uses the attention scores of `CLS' tokens within the encoder for classification tasks, such as code search. It also employs the encoder-decoder attention scores to determine token significance for sequence-to-sequence tasks like code summarization. Our evaluation shows LeanCode's superiority over the SOTAs DietCode and Slimcode, with improvements of 60% and 16% for code search, and 29% and 27% for code summarization, respectively.

The Model Context Protocol (MCP) has become a common interface for connecting large language model (LLM) agents to external tools, but its reliance on stateless, eager schema injection imposes a hidden per-turn overhead the MCP Tax or Tools Tax that practitioner reports place between roughly 10k and 60k tokens in typical multi-server deployments. This payload inflates the key-value cache, is associated with reasoning degradation as context utilization approaches published fracture points around 70%, and turns token budgets into a recurring operational cost. We introduce Tool Attention, a middleware-layer mechanism that generalizes the "Attention Is All You Need" paradigm from self-attention over tokens to gated attention over tools. Tool Attention combines (i) an Intent Schema Overlap (ISO) score from sentence embeddings, (ii) a state-aware gating function enforcing preconditions and access scopes, and (iii) a two-phase lazy schema loader that keeps a compact summary pool in context and promotes full JSON schemas only for top-k gated tools. We evaluate on a simulated 120-tool, six-server benchmark whose per-server token counts are calibrated to public audits of real MCP deployments. In this simulation, Tool Attention directly reduces measured per-turn tool tokens by 95.0% (47.3k -> 2.4k) and raises effective context utilization (a token-ratio quantity) from 24% to 91%. End-to-end figures for task success, latency, cost, and reasoning quality are reported as projections derived from the measured token counts combined with published deployment telemetry; they are not measured on live LLM agents, and we mark projected values explicitly throughout. Taken together, the results support a simple thesis: protocol-level efficiency, not raw context length, is a binding constraint on scalable gentic systems. The code for this work is accessible at https://github.com/asadani/tool-attention

We describe KVLink, an approach for efficient key-value (KV) cache reuse in large language models (LLMs). In many LLM applications, different inputs can share overlapping context, such as the same retrieved document appearing in multiple queries. However, the LLMs still need to encode the entire context for each query, leading to redundant computation. In this paper, we investigate a new strategy to eliminate such inefficiency, where the KV cache of each document is precomputed independently. During inference, the KV caches of retrieved documents are concatenated, allowing the model to reuse cached representations instead of recomputing them. To mitigate the performance degradation when using KV caches computed independently for each document, KVLink introduces two key techniques: adjusting positional embeddings of the KV cache at inference to match the global position after concatenation, and using trainable special tokens to restore self-attention across independently encoded documents. Experiments across 7 datasets demonstrate that KVLink improves question answering accuracy by an average of 4% over state-of-the-art methods. Furthermore, by leveraging precomputed KV caches, our approach reduces time-to-first-token by up to 96% compared to standard LLM inference, making it a scalable and efficient solution for context reuse. Additionally, KVLink can be combined with KV cache compression to further save cache loading and storage overhead while outperforming the baselines.

Large Language Models (LLMs) have ignited an innovative surge of AI applications, marking a new era of exciting possibilities equipped with extended context windows. However, hosting these models is cost-prohibitive mainly due to the extensive memory consumption of KV Cache involving long-context modeling. Despite several works proposing to evict unnecessary tokens from the KV Cache, most of them rely on the biased local statistics of accumulated attention scores and report performance using unconvincing metric like perplexity on inadequate short-text evaluation. In this paper, we propose NACL, a general framework for long-context KV cache eviction that achieves more optimal and efficient eviction in a single operation during the encoding phase. Due to NACL's efficiency, we combine more accurate attention score statistics in PROXY TOKENS EVICTION with the diversified random eviction strategy of RANDOM EVICTION, aiming to alleviate the issue of attention bias and enhance the robustness in maintaining pivotal tokens for long-context modeling tasks. Notably, our method significantly improves the performance on short- and long-text tasks by 80% and 76% respectively, reducing KV Cache by up to 50% with over 95% performance maintenance. The code is available at https://github.com/PaddlePaddle/Research/tree/master/NLP/ACL2024-NACL.

We introduce methods to quantify how Large Language Models (LLMs) encode and store contextual information, revealing that tokens often seen as minor (e.g., determiners, punctuation) carry surprisingly high context. Notably, removing these tokens -- especially stopwords, articles, and commas -- consistently degrades performance on MMLU and BABILong-4k, even if removing only irrelevant tokens. Our analysis also shows a strong correlation between contextualization and linearity, where linearity measures how closely the transformation from one layer's embeddings to the next can be approximated by a single linear mapping. These findings underscore the hidden importance of filler tokens in maintaining context. For further exploration, we present LLM-Microscope, an open-source toolkit that assesses token-level nonlinearity, evaluates contextual memory, visualizes intermediate layer contributions (via an adapted Logit Lens), and measures the intrinsic dimensionality of representations. This toolkit illuminates how seemingly trivial tokens can be critical for long-range understanding.

Autoregressive (AR) video generative models rely on video tokenizers that compress pixels into discrete token sequences. The length of these token sequences is crucial for balancing reconstruction quality against downstream generation computational cost. Traditional video tokenizers apply a uniform token assignment across temporal blocks of different videos, often wasting tokens on simple, static, or repetitive segments while underserving dynamic or complex ones. To address this inefficiency, we introduce EVATok, a framework to produce Efficient Video Adaptive Tokenizers. Our framework estimates optimal token assignments for each video to achieve the best quality-cost trade-off, develops lightweight routers for fast prediction of these optimal assignments, and trains adaptive tokenizers that encode videos based on the assignments predicted by routers. We demonstrate that EVATok delivers substantial improvements in efficiency and overall quality for video reconstruction and downstream AR generation. Enhanced by our advanced training recipe that integrates video semantic encoders, EVATok achieves superior reconstruction and state-of-the-art class-to-video generation on UCF-101, with at least 24.4% savings in average token usage compared to the prior state-of-the-art LARP and our fixed-length baseline.

The input tokens to Vision Transformers carry little semantic meaning as they are defined as regular equal-sized patches of the input image, regardless of its content. However, processing uniform background areas of an image should not necessitate as much compute as dense, cluttered areas. To address this issue, we propose a dynamic mixed-scale tokenization scheme for ViT, MSViT. Our method introduces a conditional gating mechanism that selects the optimal token scale for every image region, such that the number of tokens is dynamically determined per input. The proposed gating module is lightweight, agnostic to the choice of transformer backbone, and trained within a few epochs (e.g., 20 epochs on ImageNet) with little training overhead. In addition, to enhance the conditional behavior of the gate during training, we introduce a novel generalization of the batch-shaping loss. We show that our gating module is able to learn meaningful semantics despite operating locally at the coarse patch-level. We validate MSViT on the tasks of classification and segmentation where it leads to improved accuracy-complexity trade-off.

Large language models (LLMs) have demonstrated exceptional capabilities in generating text, images, and video content. However, as context length grows, the computational cost of attention increases quadratically with the number of tokens, presenting significant efficiency challenges. This paper presents an analysis of various Key-Value (KV) cache compression strategies, offering a comprehensive taxonomy that categorizes these methods by their underlying principles and implementation techniques. Furthermore, we evaluate their impact on performance and inference latency, providing critical insights into their effectiveness. Our findings highlight the trade-offs involved in KV cache compression and its influence on handling long-context scenarios, paving the way for more efficient LLM implementations.

Retrieval-augmented generation (RAG) has demonstrated significant proficiency in conducting question-answering (QA) tasks within a specified corpus. Nonetheless, numerous failure instances of RAG in QA still exist. These failures are not solely attributable to the limitations of Large Language Models (LLMs); instead, they predominantly arise from the retrieval of inaccurate information for LLMs due to two limitations: (1) Current RAG methods segment the corpus without considering semantics, making it difficult to find relevant context due to impaired correlation between questions and the segments. (2) There is a trade-off between missing essential context with fewer context retrieved and getting irrelevant context with more context retrieved. In this paper, we introduce a RAG framework (SAGE), to overcome these limitations. First, to address the segmentation issue without considering semantics, we propose to train a semantic segmentation model. This model is trained to segment the corpus into semantically complete chunks. Second, to ensure that only the most relevant chunks are retrieved while the irrelevant ones are ignored, we design a chunk selection algorithm to dynamically select chunks based on the decreasing speed of the relevance score, leading to a more relevant selection. Third, to further ensure the precision of the retrieved chunks, we propose letting LLMs assess whether retrieved chunks are excessive or lacking and then adjust the amount of context accordingly. Experiments show that SAGE outperforms baselines by 61.25% in the quality of QA on average. Moreover, by avoiding retrieving noisy context, SAGE lowers the cost of the tokens consumed in LLM inference and achieves a 49.41% enhancement in cost efficiency on average. Additionally, our work offers valuable insights for boosting RAG.

This paper presents a novel approach that enables autoregressive video generation with high efficiency. We propose to reformulate the video generation problem as a non-quantized autoregressive modeling of temporal frame-by-frame prediction and spatial set-by-set prediction. Unlike raster-scan prediction in prior autoregressive models or joint distribution modeling of fixed-length tokens in diffusion models, our approach maintains the causal property of GPT-style models for flexible in-context capabilities, while leveraging bidirectional modeling within individual frames for efficiency. With the proposed approach, we train a novel video autoregressive model without vector quantization, termed NOVA. Our results demonstrate that NOVA surpasses prior autoregressive video models in data efficiency, inference speed, visual fidelity, and video fluency, even with a much smaller model capacity, i.e., 0.6B parameters. NOVA also outperforms state-of-the-art image diffusion models in text-to-image generation tasks, with a significantly lower training cost. Additionally, NOVA generalizes well across extended video durations and enables diverse zero-shot applications in one unified model. Code and models are publicly available at https://github.com/baaivision/NOVA.

Multimodal Large Language Models (MLLMs) have shown impressive capabilities in reasoning, yet come with substantial computational cost, limiting their deployment in resource-constrained settings. Despite recent efforts on improving the efficiency of MLLMs, prior solutions fall short in responding to varying runtime conditions, in particular changing resource availability (e.g., contention due to the execution of other programs on the device). To bridge this gap, we introduce AdaLLaVA, an adaptive inference framework that learns to dynamically reconfigure operations in an MLLM during inference, accounting for the input data and a latency budget. We conduct extensive experiments across benchmarks involving question-answering, reasoning, and hallucination. Our results show that AdaLLaVA effectively adheres to input latency budget, achieving varying accuracy and latency tradeoffs at runtime. Further, we demonstrate that AdaLLaVA adapts to both input latency and content, can be integrated with token selection for enhanced efficiency, and generalizes across MLLMs. Our project webpage with code release is at https://zhuoyan-xu.github.io/ada-llava/.

Designing a unified neural network to efficiently and inherently process sequential data with arbitrary lengths is a central and challenging problem in sequence modeling. The design choices in Transformer, including quadratic complexity and weak length extrapolation, have limited their ability to scale to long sequences. In this work, we propose Gecko, a neural architecture that inherits the design of Mega and Megalodon (exponential moving average with gated attention), and further introduces multiple technical components to improve its capability to capture long range dependencies, including timestep decay normalization, sliding chunk attention mechanism, and adaptive working memory. In a controlled pretraining comparison with Llama2 and Megalodon in the scale of 7 billion parameters and 2 trillion training tokens, Gecko achieves better efficiency and long-context scalability. Gecko reaches a training loss of 1.68, significantly outperforming Llama2-7B (1.75) and Megalodon-7B (1.70), and landing close to Llama2-13B (1.67). Notably, without relying on any context-extension techniques, Gecko exhibits inherent long-context processing and retrieval capabilities, stably handling sequences of up to 4 million tokens and retrieving information from contexts up to 4times longer than its attention window. Code: https://github.com/XuezheMax/gecko-llm

Efficient long-context language modeling remains a significant challenge in Natural Language Processing (NLP). While Transformers dominate language tasks, they struggle with long sequences due to quadratic computational complexity in training and linearly scaling memory costs during inference. Recent State Space Models (SSMs) such as Mamba offer alternatives with constant memory usage, but they underperform in tasks requiring extensive in-context retrieval. We introduce Taipan, a novel hybrid architecture that combines Mamba-2 with Selective Attention Layers (SALs). These SALs identify tokens requiring long-range interactions, remove less important features, and then augment their representations using the attention module. This approach balances Mamba's efficiency with Transformer-like performance in memory-intensive tasks. By constraining the attention budget, Taipan extends accurate predictions to context lengths of up to 1 million tokens while preserving computational efficiency. Our experiments demonstrate Taipan's superior performance across various scales and tasks, offering a promising solution for efficient long-context language modeling.

Recent advances in large language models (LLMs) have driven impressive progress in omni-modal understanding and generation. However, training omni-modal LLMs remains a significant challenge due to the heterogeneous model architectures required to process diverse modalities, necessitating sophisticated system design for efficient large-scale training. Existing frameworks typically entangle model definition with parallel logic, incurring limited scalability and substantial engineering overhead for end-to-end omni-modal training. % We present \veomni, a modular and efficient training framework to accelerate the development of omni-modal LLMs. \veomni introduces model-centric distributed recipes that decouples communication from computation, enabling efficient 3D parallelism on omni-modal LLMs. \veomni also features a flexible configuration interface supporting seamless integration of new modalities with minimal code change. % Using \veomni, a omni-modal mixture-of-experts (MoE) model with 30B parameters can be trained with over 2,800 tokens/sec/GPU throughput and scale to 160K context lengths via 3D parallelism on 128 GPUs, showcasing its superior efficiency and scalability for training large omni-modal LLMs.

In Multimodal Large Language Models (MLLMs), a visual projector plays a crucial role in bridging pre-trained vision encoders with LLMs, enabling profound visual understanding while harnessing the LLMs' robust capabilities. Despite the importance of the visual projector, it has been relatively less explored. In this study, we first identify two essential projector properties: (i) flexibility in managing the number of visual tokens, crucial for MLLMs' overall efficiency, and (ii) preservation of local context from visual features, vital for spatial understanding. Based on these findings, we propose a novel projector design that is both flexible and locality-enhanced, effectively satisfying the two desirable properties. Additionally, we present comprehensive strategies to effectively utilize multiple and multifaceted instruction datasets. Through extensive experiments, we examine the impact of individual design choices. Finally, our proposed MLLM, Honeybee, remarkably outperforms previous state-of-the-art methods across various benchmarks, including MME, MMBench, SEED-Bench, and LLaVA-Bench, achieving significantly higher efficiency. Code and models are available at https://github.com/kakaobrain/honeybee.

Large language models are often used to answer queries grounded in large text corpora (e.g. codebases, legal documents, or chat histories) by placing the entire corpus in the context window and leveraging in-context learning (ICL). Although current models support contexts of 100K-1M tokens, this setup is costly to serve because the memory consumption of the KV cache scales with input length. We explore an alternative: training a smaller KV cache offline on each corpus. At inference time, we load this trained KV cache, which we call a Cartridge, and decode a response. Critically, the cost of training a Cartridge can be amortized across all the queries referencing the same corpus. However, we find that the naive approach of training the Cartridge with next-token prediction on the corpus is not competitive with ICL. Instead, we propose self-study, a training recipe in which we generate synthetic conversations about the corpus and train the Cartridge with a context-distillation objective. We find that Cartridges trained with self-study replicate the functionality of ICL, while being significantly cheaper to serve. On challenging long-context benchmarks, Cartridges trained with self-study match ICL performance while using 38.6x less memory and enabling 26.4x higher throughput. Self-study also extends the model's effective context length (e.g. from 128k to 484k tokens on MTOB) and surprisingly, leads to Cartridges that can be composed at inference time without retraining.

Efficient attentions have greatly improved the computational efficiency of Transformers. However, most existing linear attention mechanisms suffer from an efficiency degradation problem, leading to inefficiencies in causal language modeling and hindering their application in long-range language models. This problem is more pronounced under language modeling with unbounded contexts. In this paper, we propose Linear Attention Via Orthogonal memory~(\shortname) to address these limitations, achieving strong performance while maintaining linear complexity. \shortname employs orthogonal decomposition to compress a context into a fixed-size orthogonal memory while effectively minimizing redundancy within the context. Given that orthogonal memory compresses global information, we further dissect the context to amplify fine-grained local information. Additionally, we embed the relative position encoding into \shortname to improve the extrapolation ability. Experimental results show that \shortname greatly improves the efficiency of the causal language model with the best extrapolation performance and outperforms other efficient baselines. Further, we endeavor to employ \shortname for unbounded language modeling and successfully scale the context length to 128K.

Scaling laws enable the optimal selection of data amount and language model size, yet the impact of the data unit, the token, on this relationship remains underexplored. In this work, we systematically investigate how the information granularity of tokens, controlled by the compression rate (i.e., average bytes of text per token), affects scaling trends. We train 988 latent tokenized models (BLT) ranging from 50M to 7B parameters that enable setting the desired compression rate. This flexibility allows us to study the role of compression rate well beyond 4.57 bytes per token obtained with a popular BPE tokenizer. Our experiments reveal that in compute-optimal configurations, model parameter counts scale proportionally to data size measured in bytes, not in tokens as commonly perceived (Kaplan et al., 2020; Hoffmann et al., 2022). Furthermore, we discover that the optimal compression rate differs from the one obtained with BPE and decreases with compute. These findings generalize to both latent and subword tokenization, as well as to languages other than English, guiding language model developers on tokenization scheme selection for maximal compute efficiency.

Question Answering (QA) tasks, which involve extracting answers from a given context, are relatively straightforward for modern Large Language Models (LLMs) when the context is short. However, long contexts pose challenges due to the quadratic complexity of the self-attention mechanism. This challenge is compounded in Indic languages, which are often low-resource. This study explores context-shortening techniques, including Open Information Extraction (OIE), coreference resolution, Answer Paragraph Selection (APS), and their combinations, to improve QA performance. Compared to the baseline of unshortened (long) contexts, our experiments on four Indic languages (Hindi, Tamil, Telugu, and Urdu) demonstrate that context-shortening techniques yield an average improvement of 4\% in semantic scores and 47\% in token-level scores when evaluated on three popular LLMs without fine-tuning. Furthermore, with fine-tuning, we achieve an average increase of 2\% in both semantic and token-level scores. Additionally, context-shortening reduces computational overhead. Explainability techniques like LIME and SHAP reveal that when the APS model confidently identifies the paragraph containing the answer, nearly all tokens within the selected text receive high relevance scores. However, the study also highlights the limitations of LLM-based QA systems in addressing non-factoid questions, particularly those requiring reasoning or debate. Moreover, verbalizing OIE-generated triples does not enhance system performance. These findings emphasize the potential of context-shortening techniques to improve the efficiency and effectiveness of LLM-based QA systems, especially for low-resource languages. The source code and resources are available at https://github.com/ritwikmishra/IndicGenQA.

We classify and re-examine some of the current approaches to improve the performance-computes trade-off of language models, including (1) non-causal models (such as masked language models), (2) extension of batch length with efficient attention, (3) recurrence, (4) conditional computation and (5) retrieval. We identify some limitations (1) - (4) suffer from. For example, (1) currently struggles with open-ended text generation with the output loosely constrained by the input as well as performing general textual tasks like GPT-2/3 due to its need for a specific fine-tuning dataset. (2) and (3) do not improve the prediction of the first sim 10^3 tokens. Scaling up a model size (e.g. efficiently with (4)) still results in poor performance scaling for some tasks. We argue (5) would resolve many of these limitations, and it can (a) reduce the amount of supervision and (b) efficiently extend the context over the entire training dataset and the entire past of the current sample. We speculate how to modify MARGE to perform unsupervised causal modeling that achieves (b) with the retriever jointly trained.

The ability to process long contexts is crucial for many natural language processing tasks, yet it remains a significant challenge. While substantial progress has been made in enhancing the efficiency of attention mechanisms, there is still a gap in understanding how attention heads function in long-context settings. In this paper, we observe that while certain heads consistently attend to local information only, others swing between attending to local and long-context information depending on the query. This raises the question: can we identify which heads require long-context information to predict the next token accurately? We demonstrate that it's possible to predict which heads are crucial for long-context processing using only local keys. The core idea here is to exploit a simple model for the long-context scores via second moment approximations. These findings unveil simple properties of attention in the context of long sequences, and open the door to potentially significant gains in efficiency.

Recent years have witnessed the rapid advancements of large language models (LLMs) and their expanding applications, leading to soaring demands for computational resources. The widespread adoption of test-time scaling further aggravates the tension between model capability and resource consumption, highlighting the importance of inference efficiency. However, a unified metric that accurately reflects an LLM's efficiency across different model sizes and architectures remains absent. Motivated by the correlation between compression and intelligence, we introduce information capacity, a measure of model efficiency based on text compression performance relative to computational complexity. Larger models can predict the next token more accurately, achieving greater compression gains but at higher computational costs. Empirical evaluations on mainstream open-source models show that models of varying sizes within a series exhibit consistent information capacity. This metric enables a fair efficiency comparison across model series and accurate performance prediction within a model series. A distinctive feature of information capacity is that it incorporates tokenizer efficiency, which affects both input and output token counts but is often neglected in LLM evaluations. We assess the information capacity of 49 models on 5 heterogeneous datasets and observe consistent results on the influences of tokenizer efficiency, pretraining data, and the mixture-of-experts architecture.

As large language models increasingly gain popularity in real-world applications, processing extremely long contexts, often exceeding the model's pre-trained context limits, has emerged as a critical challenge. While existing approaches to efficient long-context processing show promise, recurrent compression-based methods struggle with information preservation, whereas random access approaches require substantial memory resources. We introduce REFORM, a novel inference framework that efficiently handles long contexts through a two-phase approach. First, it incrementally processes input chunks while maintaining a compressed KV cache, constructs cross-layer context embeddings, and utilizes early exit strategy for improved efficiency. Second, it identifies and gathers essential tokens via similarity matching and selectively recomputes the KV cache. Compared to baselines, REFORM achieves over 50% and 27% performance gains on RULER and BABILong respectively at 1M context length. It also outperforms baselines on Infinite-Bench and MM-NIAH, demonstrating flexibility across diverse tasks and domains. Additionally, REFORM reduces inference time by 30% and peak memory usage by 5%, achieving both efficiency and superior performance.

Solving complex or long-horizon problems often requires large language models (LLMs) to use external tools and operate over a significantly longer context window. New LLMs enable longer context windows and support tool calling capabilities. Prior works have focused mainly on evaluation of LLMs on long-context prompts, leaving agentic setup relatively unexplored, both from capability and safety perspectives. Our work addresses this gap. We find that LLM agents could be sensitive to length, type, and placement of the context, exhibiting unexpected and inconsistent shifts in task performance and in refusals to execute harmful requests. Models with 1M-2M token context windows show severe degradation already at 100K tokens, with performance drops exceeding 50\% for both benign and harmful tasks. Refusal rates shift unpredictably: GPT-4.1-nano increases from sim5\% to sim40\% while Grok 4 Fast decreases from sim80\% to sim10\% at 200K tokens. Our work shows potential safety issues with agents operating on longer context and opens additional questions on the current metrics and paradigm for evaluating LLM agent safety on long multi-step tasks. In particular, our results on LLM agents reveal a notable divergence in both capability and safety performance compared to prior evaluations of LLMs on similar criteria.

This paper details the work of the University of Groningen for the BabyLM Challenge. We follow the idea that, like babies, language models should be introduced to simpler concepts first and build off of that knowledge to understand more complex concepts. We examine this strategy of simple-then-complex through a variety of lenses, namely context size, vocabulary, and overall linguistic complexity of the data. We find that only one, context size, is truly beneficial to training a language model. However this simple change to context size gives us improvements of 2 points on average on (Super)GLUE tasks, 1 point on MSGS tasks, and 12\% on average on BLiMP tasks. Our context-limited model outperforms the baseline that was trained on 10times the amount of data.

Recent advancements in large language models (LLM) capable of processing extremely long texts highlight the need for a dedicated evaluation benchmark to assess their long-context capabilities. However, existing methods, like the needle-in-a-haystack test, do not effectively assess whether these models fully utilize contextual information, raising concerns about the reliability of current evaluation techniques. To thoroughly examine the effectiveness of existing benchmarks, we introduce a new metric called information coverage (IC), which quantifies the proportion of the input context necessary for answering queries. Our findings indicate that current benchmarks exhibit low IC; although the input context may be extensive, the actual usable context is often limited. To address this, we present ETHIC, a novel benchmark designed to assess LLMs' ability to leverage the entire context. Our benchmark comprises 2,648 test instances spanning four long-context tasks with high IC scores in the domains of books, debates, medicine, and law. Our evaluations reveal significant performance drops in contemporary LLMs, highlighting a critical challenge in managing long contexts. Our benchmark is available at https://github.com/dmis-lab/ETHIC.

While recent language models have the ability to take long contexts as input, relatively little is known about how well the language models use longer context. We analyze language model performance on two tasks that require identifying relevant information within their input contexts: multi-document question answering and key-value retrieval. We find that performance is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle of long contexts. Furthermore, performance substantially decreases as the input context grows longer, even for explicitly long-context models. Our analysis provides a better understanding of how language models use their input context and provides new evaluation protocols for future long-context models.

While long context support of large language models has extended their abilities, it also incurs challenges in memory and compute which becomes crucial bottlenecks in resource-restricted devices. Token eviction, a widely adopted post-training methodology designed to alleviate the bottlenecks by evicting less important tokens from the cache, typically uses attention scores as proxy metrics for token importance. However, one major limitation of attention score as a token-wise importance metrics is that it lacks the information about contribution of tokens to the attention output. In this paper, we propose a simple eviction criterion based on the contribution of cached tokens to attention outputs. Our method, CAOTE, optimizes for eviction error due to token eviction, by seamlessly integrating attention scores and value vectors. This is the first method which uses value vector information on top of attention-based eviction scores. Additionally, CAOTE can act as a meta-heuristic method with flexible usage with any token eviction method. We show that CAOTE, when combined with the state-of-the-art attention score-based methods, always improves accuracies on the downstream task, indicating the importance of leveraging information from values during token eviction process.

Fine tuning has been regarded as a de facto approach for adapting large language models (LLMs) to downstream tasks, but the high training memory consumption inherited from LLMs makes this process inefficient. Among existing memory efficient approaches, activation-related optimization has proven particularly effective, as activations consistently dominate overall memory consumption. Although prior arts offer various activation optimization strategies, their data-agnostic nature ultimately results in ineffective and unstable fine tuning. In this paper, we propose TokenSeek, a universal plugin solution for various transformer-based models through instance-aware token seeking and ditching, achieving significant fine-tuning memory savings (e.g., requiring only 14.8% of the memory on Llama3.2 1B) with on-par or even better performance. Furthermore, our interpretable token seeking process reveals the underlying reasons for its effectiveness, offering valuable insights for future research on token efficiency. Homepage: https://runjia.tech/iclr_tokenseek/

Efficiency is a key property to foster inclusiveness and reduce environmental costs, especially in an era of LLMs. In this work, we provide a comprehensive evaluation of efficiency for MT evaluation metrics. Our approach involves replacing computation-intensive transformers with lighter alternatives and employing linear and quadratic approximations for alignment algorithms on top of LLM representations. We evaluate six (reference-free and reference-based) metrics across three MT datasets and examine 16 lightweight transformers. In addition, we look into the training efficiency of metrics like COMET by utilizing adapters. Our results indicate that (a) TinyBERT provides the optimal balance between quality and efficiency, (b) CPU speed-ups are more substantial than those on GPU; (c) WMD approximations yield no efficiency gains while reducing quality and (d) adapters enhance training efficiency (regarding backward pass speed and memory requirements) as well as, in some cases, metric quality. These findings can help to strike a balance between evaluation speed and quality, which is essential for effective NLG systems. Furthermore, our research contributes to the ongoing efforts to optimize NLG evaluation metrics with minimal impact on performance. To our knowledge, ours is the most comprehensive analysis of different aspects of efficiency for MT metrics conducted so far.

In-context learning (ICL) is a new learning paradigm that has gained popularity along with the development of large language models. In this work, we adapt a recently proposed hardness metric, pointwise V-usable information (PVI), to an in-context version (in-context PVI). Compared to the original PVI, in-context PVI is more efficient in that it requires only a few exemplars and does not require fine-tuning. We conducted a comprehensive empirical analysis to evaluate the reliability of in-context PVI. Our findings indicate that in-context PVI estimates exhibit similar characteristics to the original PVI. Specific to the in-context setting, we show that in-context PVI estimates remain consistent across different exemplar selections and numbers of shots. The variance of in-context PVI estimates across different exemplar selections is insignificant, which suggests that in-context PVI are stable. Furthermore, we demonstrate how in-context PVI can be employed to identify challenging instances. Our work highlights the potential of in-context PVI and provides new insights into the capabilities of ICL.

Large Language Models (LLMs) rely on the Key-Value (KV) Cache to store token history, enabling efficient decoding of tokens. As the KV-Cache grows, it becomes a major memory and computation bottleneck, however, there is an opportunity to alleviate this bottleneck, especially because prior research has shown that only a small subset of tokens contribute meaningfully to each decoding step. A key challenge in finding these critical tokens is that they are dynamic, and heavily input query-dependent. Existing methods either risk quality by evicting tokens permanently, or retain the full KV-Cache but rely on retrieving chunks (pages) of tokens at generation, failing at dense, context-rich tasks. Additionally, many existing KV-Cache sparsity methods rely on inaccurate proxies for token importance. To address these limitations, we introduce TokenButler, a high-granularity, query-aware predictor that learns to identify these critical tokens. By training a light-weight predictor with less than 1.2% parameter overhead, TokenButler prioritizes tokens based on their contextual, predicted importance. This improves perplexity & downstream accuracy by over 8% relative to SoTA methods for estimating token importance. We evaluate TokenButler on a novel synthetic small-context co-referential retrieval task, demonstrating near-oracle accuracy. Code, models and benchmarks: https://github.com/abdelfattah-lab/TokenButler