Daily Papers

Autoregressive large language models (LLMs) have made remarkable progress in various natural language generation tasks. However, they incur high computation cost and latency resulting from the autoregressive token-by-token generation. To address this issue, several approaches have been proposed to reduce computational cost using early-exit strategies. These strategies enable faster text generation using reduced computation without applying the full computation graph to each token. While existing token-level early exit methods show promising results for online inference, they cannot be readily applied for batch inferencing and Key-Value caching. This is because they have to wait until the last token in a batch exits before they can stop computing. This severely limits the practical application of such techniques. In this paper, we propose a simple and effective token-level early exit method, SkipDecode, designed to work seamlessly with batch inferencing and KV caching. It overcomes prior constraints by setting up a singular exit point for every token in a batch at each sequence position. It also guarantees a monotonic decrease in exit points, thereby eliminating the need to recompute KV Caches for preceding tokens. Rather than terminating computation prematurely as in prior works, our approach bypasses lower to middle layers, devoting most of the computational resources to upper layers, allowing later tokens to benefit from the compute expenditure by earlier tokens. Our experimental results show that SkipDecode can obtain 2x to 5x inference speedups with negligible regression across a variety of tasks. This is achieved using OPT models of 1.3 billion and 6.7 billion parameters, all the while being directly compatible with batching and KV caching optimization techniques.

Residual transformations enhance the representational depth and expressive power of large language models (LLMs). However, applying static residual transformations across all tokens in auto-regressive generation leads to a suboptimal trade-off between inference efficiency and generation fidelity. Existing methods, including Early Exiting, Skip Decoding, and Mixture-of-Depth address this by modulating the residual transformation based on token-level complexity. Nevertheless, these approaches predominantly consider the distance traversed by tokens through the model layers, neglecting the underlying velocity of residual evolution. We introduce Mixture of Multi-rate Residuals (M2R2), a framework that dynamically modulates residual velocity to improve early alignment, enhancing inference efficiency. Evaluations on reasoning oriented tasks such as Koala, Self-Instruct, WizardLM, and MT-Bench show M2R2 surpasses state-of-the-art distance-based strategies, balancing generation quality and speedup. In self-speculative decoding setup, M2R2 achieves up to 2.8x speedups on MT-Bench, outperforming methods like 2-model speculative decoding, Medusa, LookAhead Decoding, and DEED. In Mixture-of-Experts (MoE) architectures, integrating early residual alignment with ahead-of-time expert loading into high-bandwidth memory (HBM) accelerates decoding, reduces expert-switching bottlenecks, and achieves a 2.9x speedup, making it highly effective in resource-constrained environments.

Speculative decoding has demonstrated its effectiveness in accelerating the inference of large language models while maintaining a consistent sampling distribution. However, the conventional approach of training a separate draft model to achieve a satisfactory token acceptance rate can be costly. Drawing inspiration from early exiting, we propose a novel self-speculative decoding framework Kangaroo, which uses a fixed shallow sub-network as a self-draft model, with the remaining layers serving as the larger target model. We train a lightweight and efficient adapter module on top of the sub-network to bridge the gap between the sub-network and the full model's representation ability. It is noteworthy that the inference latency of the self-draft model may no longer be negligible compared to the large model, necessitating strategies to increase the token acceptance rate while minimizing the drafting steps of the small model. To address this challenge, we introduce an additional early exiting mechanism for generating draft tokens. Specifically, we halt the small model's subsequent prediction during the drafting phase once the confidence level for the current token falls below a certain threshold. Extensive experiments on the Spec-Bench demonstrate the effectiveness of Kangaroo. Under single-sequence verification, Kangaroo achieves speedups up to 1.68times on Spec-Bench, outperforming Medusa-1 with 88.7\% fewer additional parameters (67M compared to 591M). The code for Kangaroo is available at https://github.com/Equationliu/Kangaroo.

Recent advances in Transformer-based large language models (LLMs) have led to significant performance improvements across many tasks. These gains come with a drastic increase in the models' size, potentially leading to slow and costly use at inference time. In practice, however, the series of generations made by LLMs is composed of varying levels of difficulty. While certain predictions truly benefit from the models' full capacity, other continuations are more trivial and can be solved with reduced compute. In this work, we introduce Confident Adaptive Language Modeling (CALM), a framework for dynamically allocating different amounts of compute per input and generation timestep. Early exit decoding involves several challenges that we address here, such as: (1) what confidence measure to use; (2) connecting sequence-level constraints to local per-token exit decisions; and (3) attending back to missing hidden representations due to early exits in previous tokens. Through theoretical analysis and empirical experiments on three diverse text generation tasks, we demonstrate the efficacy of our framework in reducing compute -- potential speedup of up to times 3 -- while provably maintaining high performance.

Automatic speech recognition (ASR) systems based on large language models (LLMs) achieve superior performance by leveraging pretrained LLMs as decoders, but their token-by-token generation mechanism leads to inference latency that grows linearly with sequence length. Meanwhile, discrete diffusion large language models (dLLMs) offer a promising alternative, enabling high-quality parallel sequence generation with pretrained decoders. However, directly applying native text-oriented dLLMs to ASR leads to a fundamental mismatch between open-ended text generation and the acoustically conditioned transcription paradigm required by ASR. As a result, it introduces unnecessary difficulty and computational redundancy, such as denoising from pure noise, inflexible generation lengths, and fixed denoising steps. We propose dLLM-ASR, an efficient dLLM-based ASR framework that formulates dLLM's decoding as a prior-guided and adaptive denoising process. It leverages an ASR prior to initialize the denoising process and provide an anchor for sequence length. Building upon this prior, length-adaptive pruning dynamically removes redundant tokens, while confidence-based denoising allows converged tokens to exit the denoising loop early, enabling token-level adaptive computation. Experiments demonstrate that dLLM-ASR achieves recognition accuracy comparable to autoregressive LLM-based ASR systems and delivers a 4.44times inference speedup, establishing a practical and efficient paradigm for ASR.

Early exiting has recently emerged as a promising technique for accelerating large language models (LLMs) by effectively reducing the hardware computation and memory access. In this paper, we present SpecEE, a fast LLM inference engine with speculative early exiting. (1) At the algorithm level, we propose the speculation-based lightweight predictor design by exploiting the probabilistic correlation between the speculative tokens and the correct results and high parallelism of GPUs. (2) At the system level, we point out that not all layers need a predictor and design the two-level heuristic predictor scheduling engine based on skewed distribution and contextual similarity. (3) At the mapping level, we point out that different decoding methods share the same essential characteristics, and propose the context-aware merged mapping for predictor with efficient GPU implementations to support speculative decoding, and form a framework for various existing orthogonal acceleration techniques (e.g., quantization and sparse activation) on cloud and personal computer (PC) scenarios, successfully pushing the Pareto frontier of accuracy and speedup. It is worth noting that SpecEE can be applied to any LLM by negligible training overhead in advance without affecting the model original parameters. Extensive experiments show that SpecEE achieves 2.25x and 2.43x speedup with Llama2-7B on cloud and PC scenarios respectively.

Increasing the size of large language models (LLMs) has been shown to lead to better performance. However, this comes at the cost of slower and more expensive inference. Early-exiting is a promising approach for improving the efficiency of LLM inference by enabling next token prediction at intermediate layers. Yet, the large vocabulary size in modern LLMs makes the confidence estimation required for exit decisions computationally expensive, diminishing the efficiency gains. To address this, we propose dynamically pruning the vocabulary at test time for each token. Specifically, the vocabulary is pruned at one of the initial layers, and the smaller vocabulary is then used throughout the rest of the forward pass. Our experiments demonstrate that such post-hoc dynamic vocabulary pruning improves the efficiency of confidence estimation in early-exit LLMs while maintaining competitive performance.

To tackle the high inference latency exhibited by autoregressive language models, previous studies have proposed an early-exiting framework that allocates adaptive computation paths for each token based on the complexity of generating the subsequent token. However, we observed several shortcomings, including performance degradation caused by a state copying mechanism or numerous exit paths, and sensitivity to exit confidence thresholds. Consequently, we propose a Fast and Robust Early-Exiting (FREE) framework, which incorporates a shallow-deep module and a synchronized parallel decoding. Our framework enables faster inference by synchronizing the decoding process of the current token with previously stacked early-exited tokens. Furthermore, as parallel decoding allows us to observe predictions from both shallow and deep models, we present a novel adaptive threshold estimator that exploits a Beta mixture model to determine suitable confidence thresholds. We empirically demonstrated the superiority of our proposed framework on extensive generation tasks.

The recent advancements in large language models (LLMs) have been extraordinary, yet the escalating inference costs associated with them present challenges in real-world applications. To address these challenges, we propose a novel approach called Early-exiting Speculative Decoding (EESD) with lossless acceleration. Specifically, EESD utilizes a segment of the LLM to generate draft tokens, incorporating Early-exiting structures after the first N layers. To enhance the quality of draft tokens, a self-distillation method is integrated. This early-exiting design not only reduces deployment and training costs but also significantly accelerates the token generation speed. Moreover, we introduce a novel sampling mechanism that leverages Thompson Sampling to regulate the generation processes, automatically determining the quantity of draft tokens in each round. The original LLM is then employed to validate these draft tokens through a single forward pass, and thus guarantees that the final output text maintains a distribution consistent with vanilla auto-regressive decoding. The experimental results on both 13B and 70B models demonstrate that our approach decodes tokens at a markedly accelerated rate compared to prior methods, showing the effectiveness of our approach.

This paper presents a novel technique for accelerating inference in large, pre-trained language models (LLMs) by introducing early exits during inference. The computational demands of these models, used across a wide range of applications, can be substantial. By capitalizing on the inherent variability in token complexity, our approach enables selective acceleration of the inference process. Specifically, we propose the integration of early exit ''heads'' atop existing transformer layers, which facilitate conditional terminations based on a confidence metric. These heads are trained in a self-supervised manner using the model's own predictions as training data, thereby eliminating the need for additional annotated data. The confidence metric, established using a calibration set, ensures a desired level of accuracy while enabling early termination when confidence exceeds a predetermined threshold. Notably, our method preserves the original accuracy and reduces computational time on certain tasks, leveraging the existing knowledge of pre-trained LLMs without requiring extensive retraining. This lightweight, modular modification has the potential to greatly enhance the practical usability of LLMs, particularly in applications like real-time language processing in resource-constrained environments.

Overparameterized large-scale language models have impressive generalization performance of in-context few-shot learning. However, most language models allocate the same amount of parameters or computation to each token, disregarding the complexity or importance of the input data. We argue that in language model pretraining, a variable amount of computation should be assigned to different tokens, and this can be efficiently achieved via a simple routing mechanism. Different from conventional early stopping techniques where tokens can early exit at only early layers, we propose a more general method that dynamically skips the execution of a layer (or module) for any input token with a binary router. In our extensive evaluation across 24 NLP tasks, we demonstrate that the proposed method can significantly improve the 1-shot performance compared to other competitive baselines only at mild extra cost for inference.

Large reasoning models achieve strong performance on complex tasks by generating extended chains of thought, but they often "overthink": continuing to reason long after they have enough information to answer correctly. This wastes inference-time compute and can hurt accuracy. Existing attempts to stop early either manipulate decoding with extra sampling and heuristics, rely on auxiliary verifier models, or operate only as post-hoc analysis pipelines without formal guarantees. We introduce LYNX, an online early-exit mechanism that turns a model's own hidden-state awareness into confidence-controlled stopping decisions. LYNX attaches exit decisions to naturally occurring reasoning cues (e.g., "hmm", "wait") during generation, trains a lightweight probe on hidden states at those cue tokens using supervision from forced exits, and wraps the resulting scores in split conformal prediction to obtain distribution-free control over premature exits. Crucially, we train and calibrate this probe once on a generic mathematical corpus and reuse it unchanged across benchmarks, decoding temperatures, and even non-mathematical tasks. Across three model families spanning 1.5B to 32B parameters, a single mathematically trained probe per base model yields strong accuracy--efficiency tradeoffs. On GSM8K, LYNX matches or improves baseline accuracy while reducing tokens by 40--65\%; on MATH-500 it improves accuracy by up to 12 points with roughly 35--60\% fewer tokens; on AIME 2024 it recovers baseline accuracy with more than 50\% token savings; and on CommonsenseQA, a non-math benchmark, it transfers zero-shot with modest accuracy gains and up to 70\% fewer tokens. Compared to state-of-the-art early-exit methods, LYNX offers competitive or superior Pareto frontiers while remaining fully online, requiring no proxy models at inference, and providing explicit, user-tunable confidence guarantees.

In this paper, we propose Patience-based Early Exit, a straightforward yet effective inference method that can be used as a plug-and-play technique to simultaneously improve the efficiency and robustness of a pretrained language model (PLM). To achieve this, our approach couples an internal-classifier with each layer of a PLM and dynamically stops inference when the intermediate predictions of the internal classifiers remain unchanged for a pre-defined number of steps. Our approach improves inference efficiency as it allows the model to make a prediction with fewer layers. Meanwhile, experimental results with an ALBERT model show that our method can improve the accuracy and robustness of the model by preventing it from overthinking and exploiting multiple classifiers for prediction, yielding a better accuracy-speed trade-off compared to existing early exit methods.

Early exits (EEs) offer a promising approach to reducing computational costs and latency by dynamically terminating inference once a satisfactory prediction confidence on a data sample is achieved. Although many works integrate EEs into encoder-only Transformers, their application to decoder-only architectures and, more importantly, Mamba models, a novel family of state-space architectures in the LLM realm, remains insufficiently explored. This work introduces DYNAMAX, the first framework to exploit the unique properties of Mamba architectures for early exit mechanisms. We not only integrate EEs into Mamba but also repurpose Mamba as an efficient EE classifier for both Mamba-based and transformer-based LLMs, showcasing its versatility. Our experiments employ the Mistral 7B transformer compared to the Codestral 7B Mamba model, using data sets such as TruthfulQA, CoQA, and TriviaQA to evaluate computational savings, accuracy, and consistency. The results highlight the adaptability of Mamba as a powerful EE classifier and its efficiency in balancing computational cost and performance quality across NLP tasks. By leveraging Mamba's inherent design for dynamic processing, we open pathways for scalable and efficient inference in embedded applications and resource-constrained environments. This study underscores the transformative potential of Mamba in redefining dynamic computing paradigms for LLMs.

The possibility of dynamically modifying the computational load of neural models at inference time is crucial for on-device processing, where computational power is limited and time-varying. Established approaches for neural model compression exist, but they provide architecturally static models. In this paper, we investigate the use of early-exit architectures, that rely on intermediate exit branches, applied to large-vocabulary speech recognition. This allows for the development of dynamic models that adjust their computational cost to the available resources and recognition performance. Unlike previous works, besides using pre-trained backbones we also train the model from scratch with an early-exit architecture. Experiments on public datasets show that early-exit architectures from scratch not only preserve performance levels when using fewer encoder layers, but also improve task accuracy as compared to using single-exit models or using pre-trained models. Additionally, we investigate an exit selection strategy based on posterior probabilities as an alternative to frame-based entropy.

Token-level code completion is one of the most critical features in modern Integrated Development Environments (IDEs). It assists developers by suggesting relevant identifiers and APIs during coding. While completions are typically derived from static analysis, their usefulness depends heavily on how they are ranked, as correct predictions buried deep in the list are rarely seen by users. Most current systems rely on hand-crafted heuristics or lightweight machine learning models trained on user logs, which can be further improved to capture context information and generalize across projects and coding styles. In this work, we propose a new scoring approach to ranking static completions using language models in a lightweight and model-agnostic way. Our method organizes all valid completions into a prefix tree and performs a single greedy decoding pass to collect token-level scores across the tree. This enables a precise token-aware ranking without needing beam search, prompt engineering, or model adaptations. The approach is fast, architecture-agnostic, and compatible with already deployed models for code completion. These findings highlight a practical and effective pathway for integrating language models into already existing tools within IDEs, and ultimately providing smarter and more responsive developer assistance.

The ability to dynamically adjust the computational load of neural models during inference in a resource aware manner is crucial for on-device processing scenarios, characterised by limited and time-varying computational resources. Early-exit architectures represent an elegant and effective solution, since they can process the input with a subset of their layers, exiting at intermediate branches (the upmost layers are hence removed from the model). From a different perspective, for automatic speech recognition applications there are memory-efficient neural architectures that apply variable frame rate analysis, through downsampling/upsampling operations in the middle layers, reducing the overall number of operations and improving significantly the performance on well established benchmarks. One example is the Zipformer. However, these architectures lack the modularity necessary to inject early-exit branches. With the aim of improving the performance in early-exit models, we propose introducing parallel layers in the architecture that process downsampled versions of their inputs. % in conjunction with standard processing layers. We show that in this way the speech recognition performance on standard benchmarks significantly improve, at the cost of a small increase in the overall number of model parameters but without affecting the inference time.

Large-scale Transformer models bring significant improvements for various downstream vision language tasks with a unified architecture. The performance improvements come with increasing model size, resulting in slow inference speed and increased cost for severing. While some certain predictions benefit from the full complexity of the large-scale model, not all of inputs need the same amount of computation to conduct, potentially leading to computation resource waste. To handle this challenge, early exiting is proposed to adaptively allocate computational power in term of input complexity to improve inference efficiency. The existing early exiting strategies usually adopt output confidence based on intermediate layers as a proxy of input complexity to incur the decision of skipping following layers. However, such strategies cannot apply to encoder in the widely-used unified architecture with both encoder and decoder due to difficulty of output confidence estimation in the encoder. It is suboptimal in term of saving computation power to ignore the early exiting in encoder component. To handle this challenge, we propose a novel early exiting strategy for unified visual language models, which allows dynamically skip the layers in encoder and decoder simultaneously in term of input layer-wise similarities with multiple times of early exiting, namely MuE. By decomposing the image and text modalities in the encoder, MuE is flexible and can skip different layers in term of modalities, advancing the inference efficiency while minimizing performance drop. Experiments on the SNLI-VE and MS COCO datasets show that the proposed approach MuE can reduce expected inference time by up to 50\% and 40\% while maintaining 99\% and 96\% performance respectively.

Despite their strong performance on reasoning tasks, large reasoning models (LRMs) often suffer from overthinking, producing unnecessarily long outputs and incurring high end-to-end latency, a significant limitation to their real-world deployment. To address overthinking, early-exit mechanisms have been proposed to terminate reasoning before typical completion, showing that this approach can effectively shorten generation length with minimal impact on accuracy. However, their reliance on probing mechanisms introduces a detection overhead that limits their end-to-end latency gains and compromises their generalizability across diverse problems. Inspired by the use of hidden states in speculative decoding, we propose SpecExit, a novel framework that predicts both future tokens and an early-exit signal directly from a lightweight draft model without probing overhead. Our method offers significant improvements, reducing average generation length by 66\% and achieving a 2.5x speedup in end-to-end latency compared to the speculative decoding baseline, without compromising accuracy. Our method leverages the inherent signals from hidden states to provide effective early-exit signals, suggesting broader use of hidden states for efficient reasoning. Our code is available at https://github.com/Tencent/AngelSlim.

Currently, pre-trained models can be considered the default choice for a wide range of NLP tasks. Despite their SoTA results, there is practical evidence that these models may require a different number of computing layers for different input sequences, since evaluating all layers leads to overconfidence in wrong predictions (namely overthinking). This problem can potentially be solved by implementing adaptive computation time approaches, which were first designed to improve inference speed. Recently proposed PonderNet may be a promising solution for performing an early exit by treating the exit layer's index as a latent variable. However, the originally proposed exit criterion, relying on sampling from trained posterior distribution on the probability of exiting from the i-th layer, introduces major variance in exit layer indices, significantly reducing the resulting model's performance. In this paper, we propose improving PonderNet with a novel deterministic Q-exit criterion and a revisited model architecture. We adapted the proposed mechanism to ALBERT and RoBERTa and compared it with recent methods for performing an early exit. We observed that the proposed changes can be considered significant improvements on the original PonderNet architecture and outperform PABEE on a wide range of GLUE tasks. In addition, we also performed an in-depth ablation study of the proposed architecture to further understand Lambda layers and their performance.

The increasing complexity of advanced machine learning models requires innovative approaches to manage computational resources effectively. One such method is the Early Exit strategy, which allows for adaptive computation by providing a mechanism to shorten the processing path for simpler data instances. In this paper, we propose EERO, a new methodology to translate the problem of early exiting to a problem of using multiple classifiers with reject option in order to better select the exiting head for each instance. We calibrate the probabilities of exiting at the different heads using aggregation with exponential weights to guarantee a fixed budget .We consider factors such as Bayesian risk, budget constraints, and head-specific budget consumption. Experimental results, conducted using a ResNet-18 model and a ConvNext architecture on Cifar and ImageNet datasets, demonstrate that our method not only effectively manages budget allocation but also enhances accuracy in overthinking scenarios.

Large Reasoning Language Models (LRLMs) demonstrate impressive capabilities on complex tasks by utilizing long Chain-of-Thought reasoning. However, they are prone to overthinking, which generates redundant reasoning steps that degrade both performance and efficiency. Recently, early-exit strategies are proposed to mitigate overthinking by dynamically and adaptively terminating redundant reasoning. However, current early-exit methods either introduce extra training overhead by relying on proxy models or limit inference throughput due to the frequent content switching between reasoning and generating probing answers. Moreover, most early-exit methods harm LRLMs performance due to over-truncation. Our insight stems from an observation: overthinking often causes LRLMs to deviate from the correct reasoning path, which is frequently accompanied by high-entropy transition tokens. Given this, we propose an early-exit method deeply coupled with the native reasoning process, which leverages the path deviation index as a dedicated monitoring metric for the frequent occurrence of high-entropy transition tokens to dynamically detect and terminate overthinking trajectories. We conduct experiments across multiple benchmarks using LRLMs of different types and scales, and the results indicate that our method delivers the largest performance improvement over vanilla CoT compared to existing early-exit methods.

We study the text generation task under the approach of pre-trained language models (PLMs). Typically, an auto-regressive (AR) method is adopted for generating texts in a token-by-token manner. Despite many advantages of AR generation, it usually suffers from inefficient inference. Therefore, non-autoregressive (NAR) models are proposed to generate all target tokens simultaneously. However, NAR models usually generate texts of lower quality due to the absence of token dependency in the output text. In this paper, we propose ELMER: an efficient and effective PLM for NAR text generation to explicitly model the token dependency during NAR generation. By leveraging the early exit technique, ELMER enables the token generations at different layers, according to their prediction confidence (a more confident token will exit at a lower layer). Besides, we propose a novel pre-training objective, Layer Permutation Language Modeling, to pre-train ELMER by permuting the exit layer for each token in sequences. Experiments on three text generation tasks show that ELMER significantly outperforms NAR models and further narrows the performance gap with AR PLMs (\eg ELMER (29.92) vs BART (30.61) ROUGE-L in XSUM) while achieving over 10 times inference speedup.

High-resolution Vision-Language Models (VLMs) have been widely used in multimodal tasks to enhance accuracy by preserving detailed image information. However, these models often generate excessive visual tokens due to encoding multiple partitions of the input image. Processing these excessive visual tokens is computationally challenging, especially in resource-constrained environments with commodity GPUs. To support high-resolution images while meeting resource constraints, we propose High-Resolution Early Dropping (HiRED), a token-dropping scheme that operates within a fixed token budget before the Large Language Model (LLM) stage. HiRED can be integrated with existing high-resolution VLMs in a plug-and-play manner, as it requires no additional training while still maintaining superior accuracy. We strategically use the vision encoder's attention in the initial layers to assess the visual content of each image partition and allocate the token budget accordingly. Then, using the attention in the final layer, we select the most important visual tokens from each partition within the allocated budget, dropping the rest. Empirically, when applied to LLaVA-Next-7B on NVIDIA TESLA P40 GPU, HiRED with a 20% token budget increases token generation throughput by 4.7, reduces first-token generation latency by 15 seconds, and saves 2.3 GB of GPU memory for a single inference.

Early exiting has become a promising approach to improving the inference efficiency of deep networks. By structuring models with multiple classifiers (exits), predictions for ``easy'' samples can be generated at earlier exits, negating the need for executing deeper layers. Current multi-exit networks typically implement linear classifiers at intermediate layers, compelling low-level features to encapsulate high-level semantics. This sub-optimal design invariably undermines the performance of later exits. In this paper, we propose Dynamic Perceiver (Dyn-Perceiver) to decouple the feature extraction procedure and the early classification task with a novel dual-branch architecture. A feature branch serves to extract image features, while a classification branch processes a latent code assigned for classification tasks. Bi-directional cross-attention layers are established to progressively fuse the information of both branches. Early exits are placed exclusively within the classification branch, thus eliminating the need for linear separability in low-level features. Dyn-Perceiver constitutes a versatile and adaptable framework that can be built upon various architectures. Experiments on image classification, action recognition, and object detection demonstrate that our method significantly improves the inference efficiency of different backbones, outperforming numerous competitive approaches across a broad range of computational budgets. Evaluation on both CPU and GPU platforms substantiate the superior practical efficiency of Dyn-Perceiver. Code is available at https://www.github.com/LeapLabTHU/Dynamic_Perceiver.

Large Reasoning Models (LRMs) often suffer from overthinking, a phenomenon in which redundant reasoning steps are generated after a correct solution has already been reached. Existing early reasoning exit methods primarily rely on output-level heuristics or trained probing models to skip redundant reasoning steps, thereby mitigating overthinking. However, these approaches typically require additional rollout computation or externally labeled datasets. In this paper, we propose NEAT, a Neuron-based Early reAsoning exiT framework that monitors neuron-level activation dynamics to enable training-free early exits, without introducing additional test-time computation. NEAT identifies exit-associated neurons and tracks their activation patterns during reasoning to dynamically trigger early exit or suppress reflection, thereby reducing unnecessary reasoning while preserving solution quality. Experiments on four reasoning benchmarks across six models with different scales and architectures show that, for each model, NEAT achieves an average token reduction of 22\% to 28\% when averaged over the four benchmarks, while maintaining accuracy.

We introduce EXIT, an extractive context compression framework that enhances both the effectiveness and efficiency of retrieval-augmented generation (RAG) in question answering (QA). Current RAG systems often struggle when retrieval models fail to rank the most relevant documents, leading to the inclusion of more context at the expense of latency and accuracy. While abstractive compression methods can drastically reduce token counts, their token-by-token generation process significantly increases end-to-end latency. Conversely, existing extractive methods reduce latency but rely on independent, non-adaptive sentence selection, failing to fully utilize contextual information. EXIT addresses these limitations by classifying sentences from retrieved documents - while preserving their contextual dependencies - enabling parallelizable, context-aware extraction that adapts to query complexity and retrieval quality. Our evaluations on both single-hop and multi-hop QA tasks show that EXIT consistently surpasses existing compression methods and even uncompressed baselines in QA accuracy, while also delivering substantial reductions in inference time and token count. By improving both effectiveness and efficiency, EXIT provides a promising direction for developing scalable, high-quality QA solutions in RAG pipelines. Our code is available at https://github.com/ThisIsHwang/EXIT

Recent advances have demonstrated that integrating reinforcement learning with rule-based rewards can significantly enhance the reasoning capabilities of large language models, even without supervised fine-tuning. However, prevalent reinforcement learning algorithms such as GRPO and its variants like DAPO, suffer from a coarse granularity issue when computing the advantage. Specifically, they compute rollout-level advantages that assign identical values to every token within a sequence, failing to capture token-specific contributions and hindering effective learning. To address this limitation, we propose Key-token Advantage Estimation (KTAE) - a novel algorithm that estimates fine-grained, token-level advantages without introducing additional models. KTAE leverages the correctness of sampled rollouts and applies statistical analysis to quantify the importance of individual tokens within a sequence to the final outcome. This quantified token-level importance is then combined with the rollout-level advantage to obtain a more fine-grained token-level advantage estimation. Empirical results show that models trained with GRPO+KTAE and DAPO+KTAE outperform baseline methods across five mathematical reasoning benchmarks. Notably, they achieve higher accuracy with shorter responses and even surpass R1-Distill-Qwen-1.5B using the same base model.

We present EE-LLM, a framework for large-scale training and inference of early-exit large language models (LLMs). While recent works have shown preliminary evidence for the efficacy of early exiting in accelerating LLM inference, EE-LLM makes a foundational step towards scaling up early-exit LLMs by supporting their training and inference with massive 3D parallelism. Built upon Megatron-LM, EE-LLM implements a variety of algorithmic innovations and performance optimizations tailored to early exiting, including a lightweight method that facilitates backpropagation for the early-exit training objective with pipeline parallelism, techniques of leveraging idle resources in the original pipeline schedule for computation related to early-exit layers, and two approaches of early-exit inference that are compatible with KV caching for autoregressive generation. Our analytical and empirical study shows that EE-LLM achieves great training efficiency with negligible computational overhead compared to standard LLM training, as well as outstanding inference speedup without compromising output quality. To facilitate further research and adoption, we release EE-LLM at https://github.com/pan-x-c/EE-LLM.

Mobile devices such as smartphones and autonomous vehicles increasingly rely on deep neural networks (DNNs) to execute complex inference tasks such as image classification and speech recognition, among others. However, continuously executing the entire DNN on mobile devices can quickly deplete their battery. Although task offloading to cloud/edge servers may decrease the mobile device's computational burden, erratic patterns in channel quality, network, and edge server load can lead to a significant delay in task execution. Recently, approaches based on split computing (SC) have been proposed, where the DNN is split into a head and a tail model, executed respectively on the mobile device and on the edge server. Ultimately, this may reduce bandwidth usage as well as energy consumption. Another approach, called early exiting (EE), trains models to embed multiple "exits" earlier in the architecture, each providing increasingly higher target accuracy. Therefore, the trade-off between accuracy and delay can be tuned according to the current conditions or application demands. In this paper, we provide a comprehensive survey of the state of the art in SC and EE strategies by presenting a comparison of the most relevant approaches. We conclude the paper by providing a set of compelling research challenges.

Diffusion large language models (dLLMs) have emerged as a promising alternative to autoregressive models, offering flexible generation orders and strong performance on complex reasoning tasks. However, instruction-tuned dLLMs exhibit a critical vulnerability we term <eos> overflow: as allocated sequence length increases, responses paradoxically become shorter, collapsing into early termination or degenerating into streams of <eos> tokens. Although noticed in practice, this issue has not been systematically analyzed. We trace its root cause to the dual role of <eos> as both termination and padding, which concentrates probability mass on <eos> at later positions and propagates backward to trigger early termination. To address this, we introduce Rainbow Padding, a simple remedy that replaces repeated <eos> placeholders with a repeating cycle of distinct padding tokens, distributing probability mass and breaking <eos> dominance. Experiments show that Rainbow Padding substantially improves length robustness and output quality, with as few as seven padding tokens sufficient to prevent early termination. Moreover, the method integrates efficiently into existing instruction-tuned models: LoRA fine-tuning for a single epoch on minimal data yields significant improvements, making this solution highly practical. The code is publicly available at https://github.com/quasar529/rainbow-padding.

Speculative Decoding (SD) is a widely used approach to accelerate the inference of large language models (LLMs) without reducing generation quality. It operates by first using a compact model to draft multiple tokens efficiently, followed by parallel verification using the target LLM. This approach leads to faster inference compared to auto-regressive decoding. While there are multiple approaches to create a draft model, one promising approach is to use early-exit methods. These methods draft candidate tokens by using a subset of layers of the primary model and applying the remaining layers for verification, allowing a single model to handle both drafting and verification. While this technique reduces memory usage and computational cost, its performance relies on the choice of the exit layer for drafting and the number of tokens drafted (speculation length) in each SD round. Prior works use hyperparameter exploration to statically select these values. However, our evaluations show that these hyperparameter values are task-specific, and even within a task they are dependent on the current sequence context. We introduce DEL, a plug-and-play method that adaptively selects the exit layer and speculation length during inference. DEL dynamically tracks the token acceptance rate if the tokens are drafted at each layer of an LLM and uses that knowledge to heuristically select the optimal exit layer and speculation length. Our experiments across a broad range of models and downstream tasks show that DEL achieves overall speedups of 2.16timessim2.50times over vanilla auto-regressive decoding and improves upon the state-of-the-art SD methods by up to 0.27times.

Test-time scaling (TTS), which involves dynamic allocation of compute during inference, offers a promising way to improve reasoning in large language models. While existing TTS methods work well, they often rely on long decoding paths or require a large number of samples to be generated, increasing the token usage and inference latency. We observe the surprising fact that for reasoning tasks, shorter traces are much more likely to be correct than longer ones. Motivated by this, we introduce First Finish Search (FFS), a training-free parallel decoding strategy that launches n independent samples and returns as soon as any one completes. We evaluate FFS alongside simple decoding, beam search, majority voting, and budget forcing on four reasoning models (DeepSeek-R1, R1-Distill-Qwen-32B, QwQ-32B and Phi-4-Reasoning-Plus) and across four datasets (AIME24, AIME25-I, AIME25-II and GPQA Diamond). With DeepSeek-R1, FFS achieves 82.23% accuracy on the AIME datasets, a 15% improvement over DeepSeek-R1's standalone accuracy, nearly matching OpenAI's o4-mini performance. Our theoretical analysis explains why stopping at the shortest trace is likely to yield a correct answer and identifies the conditions under which early stopping may be suboptimal. The elegance and simplicity of FFS demonstrate that straightforward TTS strategies can perform remarkably well, revealing the untapped potential of simple approaches at inference time.

Large Language Models (LLMs) are increasingly relied upon for solving complex reasoning tasks in domains such as mathematics, logic, and multi-step question answering. A growing line of work seeks to improve reasoning quality by scaling inference time compute particularly through Process Reward Models (PRMs), used to reward the reasoning at intermediate steps. While effective, these methods introduce substantial computational overhead, especially when generating large numbers of solutions in parallel. In this paper, we investigate whether PRMs can be used mid-generation to provide early signals that enable the rejection of suboptimal candidates before full generation of step is complete. We introduce the hypothesis that PRMs are also Partial Reward Models, meaning that the scores they assign to partially completed reasoning step are predictive of final output quality. This allows for principled early rejection based on intermediate token-level signals. We support this hypothesis both theoretically, by proving that the risk of discarding optimal beams decreases exponentially with generation length and empirically, by demonstrating a strong correlation between partial and final rewards across multiple reward models. On math reasoning benchmarks, our method achieves up to 1.4times-9times reduction in inference FLOPs without degrading final performance. These results suggest that early rejection is a powerful mechanism for improving the compute-efficiency of reasoning in LLMs.

Large pretrained models, coupled with fine-tuning, are slowly becoming established as the dominant architecture in machine learning. Even though these models offer impressive performance, their practical application is often limited by the prohibitive amount of resources required for every inference. Early-exiting dynamic neural networks (EDNN) circumvent this issue by allowing a model to make some of its predictions from intermediate layers (i.e., early-exit). Training an EDNN architecture is challenging as it consists of two intertwined components: the gating mechanism (GM) that controls early-exiting decisions and the intermediate inference modules (IMs) that perform inference from intermediate representations. As a result, most existing approaches rely on thresholding confidence metrics for the gating mechanism and strive to improve the underlying backbone network and the inference modules. Although successful, this approach has two fundamental shortcomings: 1) the GMs and the IMs are decoupled during training, leading to a train-test mismatch; and 2) the thresholding gating mechanism introduces a positive bias into the predictive probabilities, making it difficult to readily extract uncertainty information. We propose a novel architecture that connects these two modules. This leads to significant performance improvements on classification datasets and enables better uncertainty characterization capabilities.

Critic-free methods like GRPO reduce memory demands by estimating advantages from multiple rollouts but tend to converge slowly, as critical learning signals are diluted by an abundance of uninformative samples and tokens. To tackle this challenge, we propose the Dynamic Dual-Level Down-Sampling (D^3S) framework that prioritizes the most informative samples and tokens across groups to improve the efficient of policy optimization. D^3S operates along two levels: (1) the sample-level, which selects a subset of rollouts to maximize advantage variance (Var(A)). We theoretically proven that this selection is positively correlated with the upper bound of the policy gradient norms, yielding higher policy gradients. (2) the token-level, which prioritizes tokens with a high product of advantage magnitude and policy entropy (|A_{i,t}|times H_{i,t}), focusing updates on tokens where the policy is both uncertain and impactful. Moreover, to prevent overfitting to high-signal data, D^3S employs a dynamic down-sampling schedule inspired by curriculum learning. This schedule starts with aggressive down-sampling to accelerate early learning and gradually relaxes to promote robust generalization. Extensive experiments on Qwen2.5 and Llama3.1 demonstrate that integrating D^3S into advanced RL algorithms achieves state-of-the-art performance and generalization while requiring fewer samples and tokens across diverse reasoning benchmarks. Our code is added in the supplementary materials and will be made publicly available.

Inference accounts for the majority of latency and energy consumption in large language model (LLM) deployments, often exceeding 90% of total cost. While training-time efficiency has seen extensive progress, runtime optimization remains a key bottleneck, particularly under autoregressive decoding. Existing approaches -- such as pruning, quantization, early exits, and speculative decoding -- often require retraining, architectural changes, or disrupt decoding compatibility. We introduce QuickSilver, a modular, token-level framework that enables semantic adaptivity at inference time without altering model weights or structure. QuickSilver integrates four synergistic mechanisms: (i) Dynamic Token Halting, which halts computation for tokens with converged representations; (ii) KV Cache Skipping, which selectively suppresses memory writes to reduce attention overhead; and (iii) Contextual Token Fusion, which collapses redundant tokens into shared paths to shrink sequence length. Unlike speculative decoding or MoE routing, QuickSilver operates entirely on frozen, dense models and requires no auxiliary networks. Applied to GPT-2 and Llama-2 across WikiText-103 and C4, QuickSilver achieves up to 39.6% FLOP reduction with negligible perplexity degradation (<=0.2).

Large Language Models (LLMs) and Vision-Language Models (VLMs) have demonstrated remarkable capabilities. However, their deployment is hindered by significant computational costs. Existing structured pruning methods, while hardware-efficient, often suffer from significant accuracy degradation. In this paper, we argue that this failure stems from a stage-agnostic pruning approach that overlooks the asymmetric roles between the prefill and decode stages. By introducing a virtual gate mechanism, our importance analysis reveals that deep layers are critical for next-token prediction (decode) but largely redundant for context encoding (prefill). Leveraging this insight, we propose Prefill-Only Pruning (POP), a stage-aware inference strategy that safely omits deep layers during the computationally intensive prefill stage while retaining the full model for the sensitive decode stage. To enable the transition between stages, we introduce independent Key-Value (KV) projections to maintain cache integrity, and a boundary handling strategy to ensure the accuracy of the first generated token. Extensive experiments on Llama-3.1, Qwen3-VL, and Gemma-3 across diverse modalities demonstrate that POP achieves up to 1.37times speedup in prefill latency with minimal performance loss, effectively overcoming the accuracy-efficiency trade-off limitations of existing structured pruning methods.

In infilling tasks, sub-tokens, representing instances where a complete token is segmented into two parts, often emerge at the boundaries of prefixes, middles, and suffixes. Traditional methods focused on training models at the token level, leading to sub-optimal performance in character-level infilling tasks during the inference stage. Alternately, some approaches considered character-level infilling, but they relied on predicting sub-tokens in inference, yet this strategy diminished ability in character-level infilling tasks due to the large perplexity of the model on sub-tokens. In this paper, we introduce FIM-SE, which stands for Fill-In-the-Middle with both Starting and Ending character constraints. The proposed method addresses character-level infilling tasks by utilizing a line-level format to avoid predicting any sub-token in inference. In addition, we incorporate two special tokens to signify the rest of the incomplete lines, thereby enhancing generation guidance. Extensive experiments demonstrate that our proposed approach surpasses previous methods, offering a significant advantage. Code is available at https://github.com/SenseLLM/FIM-SE.

Token dropping is a recently-proposed strategy to speed up the pretraining of masked language models, such as BERT, by skipping the computation of a subset of the input tokens at several middle layers. It can effectively reduce the training time without degrading much performance on downstream tasks. However, we empirically find that token dropping is prone to a semantic loss problem and falls short in handling semantic-intense tasks. Motivated by this, we propose a simple yet effective semantic-consistent learning method (ScTD) to improve the token dropping. ScTD aims to encourage the model to learn how to preserve the semantic information in the representation space. Extensive experiments on 12 tasks show that, with the help of our ScTD, token dropping can achieve consistent and significant performance gains across all task types and model sizes. More encouragingly, ScTD saves up to 57% of pretraining time and brings up to +1.56% average improvement over the vanilla token dropping.

Diffusion Large Language Models (dLLMs) offer a compelling paradigm for natural language generation, leveraging parallel decoding and bidirectional attention to achieve superior global coherence compared to autoregressive models. While recent works have accelerated inference via KV cache reuse or heuristic decoding, they overlook the intrinsic inefficiencies within the block-wise diffusion process. Specifically, they suffer from spatial redundancy by modeling informative-sparse suffix regions uniformly and temporal inefficiency by applying fixed denoising schedules across all the decoding process. To address this, we propose Streaming-dLLM, a training-free framework that streamlines inference across both spatial and temporal dimensions. Spatially, we introduce attenuation guided suffix modeling to approximate the full context by pruning redundant mask tokens. Temporally, we employ a dynamic confidence aware strategy with an early exit mechanism, allowing the model to skip unnecessary iterations for converged tokens. Extensive experiments show that Streaming-dLLM achieves up to 68.2X speedup while maintaining generation quality, highlighting its effectiveness in diffusion decoding. The code is available at https://github.com/xiaoshideta/Streaming-dLLM.

Large Language Models (LLMs) typically generate outputs token by token using a fixed compute budget, leading to inefficient resource utilization. To address this shortcoming, recent advancements in mixture of expert (MoE) models, speculative decoding, and early exit strategies leverage the insight that computational demands can vary significantly based on the complexity and nature of the input. However, identifying optimal routing patterns for dynamic execution remains an open challenge, limiting the full potential of these adaptive methods. To address this need, we study adaptive computation in LLMs more systematically. We propose a novel framework that integrates smaller auxiliary modules within each Feed-Forward Network layer of the LLM. This design enables dynamic routing of tokens based on task complexity: tokens can be processed by either the small or big modules at each layer, or even bypass certain layers entirely. This allows us to introduce a novel notion of a token's difficulty, defined by its potential to benefit from additional computational resources. Importantly, by employing oracles to identify optimal patterns of adaptive computations, we gain valuable insights into the internal workings of LLMs and the routing processes in a simplified heterogeneous MoE setup. We show that trained routers operate differently from oracles and often yield suboptimal solutions. Notably, activating a large module in just one layer outperforms models that use large modules across all layers, underscoring the gap between practical implementations of routing in MoE models and theoretical optima for adaptive computation.

Modern reasoning models continue generating long after the answer is already determined. Across five model configurations, two families, and three benchmarks, we find that 52--88\% of chain-of-thought tokens are produced after the answer is recoverable from a partial prefix. This post-commitment generation reveals a structural phenomenon: the detection--extraction gap. Free continuations from early prefixes recover the correct answer even at 10\% of the trace, while forced extraction fails on 42\% of these cases. The answer is recoverable from the model state, yet prompt-conditioned decoding fails to extract it. We formalize this mismatch via a total-variation bound between free and forced continuation distributions, yielding quantitative estimates of suffix-induced shift. Exploiting this asymmetry, we propose Black-box Adaptive Early Exit (), which uses free continuations for both detection and extraction, truncating 70--78\% of serial generation while improving accuracy by 1--5\,pp across all models. For thinking-mode models, early exit prevents post-commitment overwriting, yielding gains of up to 5.8\,pp; a cost-optimized variant achieves 68--73\% reduction at a median of 9 API calls. Code is available at https://github.com/EdWangLoDaSc/know2say.

Due to the large number of parameters, the inference phase of Large Language Models (LLMs) is resource-intensive. Unlike traditional model compression, which needs retraining, recent dynamic computation methods show that not all components are required for inference, enabling a training-free pipeline. In this paper, we focus on the dynamic depth of LLM generation. A token-position aware layer skipping framework is proposed to save 1.5x times operations efficiently while maintaining performance. We first observed that tokens predicted later have lower perplexity and thus require less computation. Then, we propose a training-free algorithm called Position-Aware Depth Decay Decoding (D^3), which leverages a power-law decay function, leftlfloor L times (alpha^i) rightrfloor, to determine the number of layers to retain when generating token T_i. Remarkably, without any retraining, the D^3 achieves success across a wide range of generation tasks for the first time. Experiments on large language models (\ie the Llama) with 7 sim 70 billion parameters show that D^3 can achieve an average 1.5x speedup compared with the full-inference pipeline while maintaining comparable performance with nearly no performance drop (<1%) on the GSM8K and BBH benchmarks.

We investigate whether the success of a zero-shot Chain-of-Thought (CoT) process can be predicted before completion. We discover that a probing classifier, based on LLM representations, performs well even before a single token is generated, suggesting that crucial information about the reasoning process is already present in the initial steps representations. In contrast, a strong BERT-based baseline, which relies solely on the generated tokens, performs worse, likely because it depends on shallow linguistic cues rather than deeper reasoning dynamics. Surprisingly, using later reasoning steps does not always improve classification. When additional context is unhelpful, earlier representations resemble later ones more, suggesting LLMs encode key information early. This implies reasoning can often stop early without loss. To test this, we conduct early stopping experiments, showing that truncating CoT reasoning still improves performance over not using CoT at all, though a gap remains compared to full reasoning. However, approaches like supervised learning or reinforcement learning designed to shorten CoT chains could leverage our classifier's guidance to identify when early stopping is effective. Our findings provide insights that may support such methods, helping to optimize CoT's efficiency while preserving its benefits.

Tokenization is associated with many poorly understood shortcomings in language models (LMs), yet remains an important component for long sequence scaling purposes. This work studies how tokenization impacts model performance by analyzing and comparing the stochastic behavior of tokenized models with their byte-level, or token-free, counterparts. We discover that, even when the two models are statistically equivalent, their predictive distributions over the next byte can be substantially different, a phenomenon we term as "tokenization bias''. To fully characterize this phenomenon, we introduce the Byte-Token Representation Lemma, a framework that establishes a mapping between the learned token distribution and its equivalent byte-level distribution. From this result, we develop a next-byte sampling algorithm that eliminates tokenization bias without requiring further training or optimization. In other words, this enables zero-shot conversion of tokenized LMs into statistically equivalent token-free ones. We demonstrate its broad applicability with two use cases: fill-in-the-middle (FIM) tasks and model ensembles. In FIM tasks where input prompts may terminate mid-token, leading to out-of-distribution tokenization, our method mitigates performance degradation and achieves an approximately 18% improvement in FIM coding benchmarks, consistently outperforming the standard token healing fix. For model ensembles where each model employs a distinct vocabulary, our approach enables seamless integration, resulting in improved performance (up to 3.7%) over individual models across various standard baselines in reasoning, knowledge, and coding.

The prefill stage in long-context LLM inference remains a computational bottleneck. Recent token-ranking heuristics accelerate inference by selectively processing a subset of semantically relevant tokens. However, existing methods suffer from unstable token importance estimation, often varying between layers. Evaluating token-ranking quality independently from heuristic-specific architectures is challenging. To address this, we introduce an Answer-Informed Oracle, which defines ground-truth token importance by measuring attention from generated answers back to the prompt. This oracle reveals that existing heuristics exhibit high variance across layers: rankings can degrade sharply at specific layers, a failure mode invisible to end-to-end benchmarks. The diagnosis suggests a simple fix: aggregate scores across layers rather than relying on any single one. We implement this as Cross-Layer Attention Aggregation (CLAA), which closes the gap to the oracle upper bound and reduces Time-to-First-Token (TTFT) by up to 39\% compared to the Full KV Cache baseline.

Reinforcement Learning (RL) for Large Language Models (LLMs) often suffers from training collapse in long-horizon tasks due to exploding gradient variance. To mitigate this, a baseline is commonly introduced for advantage computation; however, traditional value models remain difficult to optimize, and standard group-based baselines overlook sequence heterogeneity. Although classic optimal baseline theory can achieve global variance reduction, it neglects token heterogeneity and requires prohibitive gradient-based computation. In this work, we derive the Optimal Token Baseline (OTB) from first principles, proving that gradient updates should be weighted inversely to their cumulative gradient norm. To ensure efficiency, we propose the Logit-Gradient Proxy that approximates the gradient norm using only forward-pass probabilities. Our method achieves training stability and matches the performance of large group sizes (N=32) with only N=4, reducing token consumption by over 65% across single-turn and tool-integrated reasoning tasks.

The deployment of Large Language Models (LLMs) faces a critical bottleneck when handling lengthy inputs: the prohibitive memory footprint of the Key Value (KV) cache. To address this bottleneck, the token pruning paradigm leverages attention sparsity to selectively retain a small, critical subset of tokens. However, existing approaches fall short, with static methods risking irreversible information loss and dynamic strategies employing heuristics that insufficiently capture the query-dependent nature of token importance. We propose FASA, a novel framework that achieves query-aware token eviction by dynamically predicting token importance. FASA stems from a novel insight into RoPE: the discovery of functional sparsity at the frequency-chunk (FC) level. Our key finding is that a small, identifiable subset of "dominant" FCs consistently exhibits high contextual agreement with the full attention head. This provides a robust and computationally free proxy for identifying salient tokens. %making them a powerful and efficient proxy for token importance. Building on this insight, FASA first identifies a critical set of tokens using dominant FCs, and then performs focused attention computation solely on this pruned subset. % Since accessing only a small fraction of the KV cache, FASA drastically lowers memory bandwidth requirements and computational cost. Across a spectrum of long-context tasks, from sequence modeling to complex CoT reasoning, FASA consistently outperforms all token-eviction baselines and achieves near-oracle accuracy, demonstrating remarkable robustness even under constraint budgets. Notably, on LongBench-V1, FASA reaches nearly 100\% of full-KV performance when only keeping 256 tokens, and achieves 2.56times speedup using just 18.9\% of the cache on AIME24.

Diffusion large language models (dLLMs) offer a promising alternative to autoregressive models, but their practical utility is severely hampered by slow, iterative sampling. We present SchED, a training-free, model-agnostic early-exit algorithm that aggregates full-span logit margins and halts decoding once a smooth, progress-dependent confidence threshold is met. We evaluated SchED on two dLLM families (Dream and LLaDA), in base and instruction-tuned variants across ten benchmarks spanning downstream tasks including multiple-choice question answering (MCQ), math, long-form QA/summarization, and translation. SchED delivers large, stable accelerations: on instruction-tuned models, it achieves 3.8-4.0times speedups while retaining 99.8-100% of the baseline score on average. On base models, SchED yields consistent speedup gains with 99.1-100% performance retention, with up to 2.34times under more aggressive settings. Using a conservative speed metric that heavily penalizes quality loss (QPS, γ{=}4), we show that SchED is robust and clearly outperforms prior confidence-based early-exit methods, which break down on long-form generation. An entropy analysis of the model's token predictions reveals that instruction tuning speeds up the decay of predictive entropy. By turning genuine confidence stabilization into computational savings, SchED makes dLLM decoding substantially more efficient.

Early stopping monitors global validation loss and halts all parameter updates simultaneously, which is computationally costly for large transformers due to the extended time required for validation inference. We propose GradES, a novel gradient-based early stopping approach that operates within transformer components (attention projections and Feed-Forward layer matrices). We found that different components converge at varying rates during fine-tuning. GradES tracks the magnitude of gradients in backpropagation for these matrices during training. When a projection matrix's gradients fall below a convergence threshold tau, we exclude that projection matrix from further updates individually, eliminating costly validation passes while allowing slow converging matrices to continue learning. By strategically freezing parameters when their gradients converge, GradES speeds up training time by 1.57--7.22times while simultaneously enhancing generalization through early prevention of overfitting, resulting in 1.2% higher average accuracy.

The excessive use of visual tokens in existing Multimoal Large Language Models (MLLMs) often exhibits obvious redundancy and brings in prohibitively expensive computation. To gain insights into this problem, we first conduct extensive empirical studies on the attention behaviors of MLLMs, and summarize three main inference stages in MLLMs: (i) Early fusion between tokens is first accomplished quickly. (ii) Intra-modality modeling then comes to play. (iii) Multimodal reasoning} resumes and lasts until the end of inference. In particular, we reveal that visual tokens will stop contributing to reasoning when the text tokens receive enough image information, yielding obvious visual redundancy. Based on these generalized observations, we propose a simple yet effective method to improve the efficiency of MLLMs, termed dynamic visual-token exit (DyVTE). DyVTE uses lightweight hyper-networks to perceive the text token status and decide the removal of all visual tokens after a certain layer, thereby addressing the observed visual redundancy. To validate VTE, we apply it to a set of MLLMs, including LLaVA, VILA, Eagle and InternVL, and conduct extensive experiments on a bunch of benchmarks. The experiment results not only show the effectiveness of our VTE in improving MLLMs' efficiency, but also yield the general modeling patterns of MLLMs, well facilitating the in-depth understanding of MLLMs. Our code is anonymously released at https://github.com/DoubtedSteam/DyVTE.

Machine learning (ML) inference platforms are tasked with balancing two competing goals: ensuring high throughput given many requests, and delivering low-latency responses to support interactive applications. Unfortunately, existing platform knobs (e.g., batch sizes) fail to ease this fundamental tension, and instead only enable users to harshly trade off one property for the other. This paper explores an alternate strategy to taming throughput-latency tradeoffs by changing the granularity at which inference is performed. We present Apparate, a system that automatically applies and manages early exits (EEs) in ML models, whereby certain inputs can exit with results at intermediate layers. To cope with the time-varying overhead and accuracy challenges that EEs bring, Apparate repurposes exits to provide continual feedback that powers several novel runtime monitoring and adaptation strategies. Apparate lowers median response latencies by 40.5--91.5% and 10.0--24.2% for diverse CV and NLP classification workloads, and median time-per-token latencies by 22.6--77.9% for generative scenarios, without affecting throughputs or violating tight accuracy constraints.

We introduce GateSkip, a simple residual-stream gating mechanism that enables token-wise layer skipping in decoder-only LMs. Each Attention/MLP branch is equipped with a sigmoid-linear gate that condenses the branch's output before it re-enters the residual stream. During inference we rank tokens by the gate values and skip low-importance ones using a per-layer budget. While early-exit or router-based Mixture-of-Depths models are known to be unstable and need extensive retraining, our smooth, differentiable gates fine-tune stably on top of pretrained models. On long-form reasoning, we save up to 15\% compute while retaining over 90\% of baseline accuracy. On instruction-tuned models we see accuracy gains at full compute and match baseline quality near 50\% savings. The learned gates give insight into transformer information flow (e.g., BOS tokens act as anchors), and the method combines easily with quantization, pruning, and self-speculative decoding.

Large Language Models (LLMs) achieve impressive reasoning capabilities at the cost of substantial inference overhead, posing substantial deployment challenges. Although distilled Small Language Models (SLMs) significantly enhance efficiency, their performance suffers as they fail to follow LLMs' reasoning paths. Luckily, we reveal that only a small fraction of tokens genuinely diverge reasoning paths between LLMs and SLMs. Most generated tokens are either identical or exhibit neutral differences, such as minor variations in abbreviations or expressions. Leveraging this insight, we introduce **Roads to Rome (R2R)**, a neural token routing method that selectively utilizes LLMs only for these critical, path-divergent tokens, while leaving the majority of token generation to the SLM. We also develop an automatic data generation pipeline that identifies divergent tokens and generates token-level routing labels to train the lightweight router. We apply R2R to combine R1-1.5B and R1-32B models from the DeepSeek family, and evaluate on challenging math, coding, and QA benchmarks. With an average activated parameter size of 5.6B, R2R surpasses the average accuracy of R1-7B by 1.6x, outperforming even the R1-14B model. Compared to R1-32B, it delivers a 2.8x wall-clock speedup with comparable performance, advancing the Pareto frontier of test-time scaling efficiency. Our code is available at https://github.com/thu-nics/R2R.

Recent breakthroughs in large language models (LLMs) have led to notable successes in complex reasoning tasks, such as mathematical problem solving. A common strategy for improving performance is parallel thinking, in which multiple reasoning traces are generated and the final prediction is made using aggregation schemes like majority voting or best-of-N decoding. However, two key challenges persist. First, multi-sample decoding incurs substantial inference latency, especially for long-form outputs. Second, effective mechanisms for reliably assessing the correctness of individual reasoning traces are still limited. To address these challenges, we introduce One-Token Verification (OTV), a computational method that estimates reasoning correctness in a single forward pass during generation. OTV is activated by a learnable token and integrated into the LLM via low-rank adaptation to probe internal reasoning signals through the key-value cache, supporting token-level correctness estimation at any stage of generation without disrupting primary reasoning. Experiments on mathematical reasoning benchmarks demonstrate that OTV consistently surpasses existing verifiers. Additionally, OTV reduces token usage by up to 90% through correctness-guided early termination, prioritizing shorter, more reliable solutions.

The growing demand for large language models (LLMs) requires serving systems to handle many concurrent requests with diverse service level objectives (SLOs). This exacerbates head-of-line (HoL) blocking during the compute-intensive prefill phase, where long-running requests monopolize resources and delay higher-priority ones, leading to widespread time-to-first-token (TTFT) SLO violations. While chunked prefill enables interruptibility, it introduces an inherent trade-off between responsiveness and throughput: reducing chunk size improves response latency but degrades computational efficiency, whereas increasing chunk size maximizes throughput but exacerbates blocking. This necessitates an adaptive preemption mechanism. However, dynamically balancing execution granularity against scheduling overheads remains a key challenge. In this paper, we propose FlowPrefill, a TTFT-goodput-optimized serving system that resolves this conflict by decoupling preemption granularity from scheduling frequency. To achieve adaptive prefill scheduling, FlowPrefill introduces two key innovations: 1) Operator-Level Preemption, which leverages operator boundaries to enable fine-grained execution interruption without the efficiency loss associated with fixed small chunking; and 2) Event-Driven Scheduling, which triggers scheduling decisions only upon request arrival or completion events, thereby supporting efficient preemption responsiveness while minimizing control-plane overhead. Evaluation on real-world production traces shows that FlowPrefill improves maximum goodput by up to 5.6times compared to state-of-the-art systems while satisfying heterogeneous SLOs.

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.

Pre-tokenization, the initial step in many modern tokenization pipelines, segments text into smaller units called pretokens, typically splitting on whitespace and punctuation. While this process encourages having full, individual words as tokens, it introduces a fundamental limitation in most tokenization algorithms such as Byte Pair Encoding (BPE). Specifically, pre-tokenization causes the distribution of tokens in a corpus to heavily skew towards common, full-length words. This skewed distribution limits the benefits of expanding to larger vocabularies, since the additional tokens appear with progressively lower counts. To overcome this barrier, we propose BoundlessBPE, a modified BPE algorithm that relaxes the pretoken boundary constraint. Our approach selectively merges two complete pretokens into a larger unit we term a superword. Superwords are not necessarily semantically cohesive. For example, the pretokens " of" and " the" might be combined to form the superword " of the". This merging strategy results in a substantially more uniform distribution of tokens across a corpus than standard BPE, and compresses text more effectively, with an approximate 20% increase in bytes per token.

Real-time spoken dialogue systems face a fundamental tension between latency and response quality. End-to-end speech-to-speech (S2S) models respond immediately and naturally handle turn-taking, backchanneling, and interruption, but produce semantically weaker outputs. Cascaded pipelines (ASR -> LLM) deliver stronger responses at the cost of latency that grows with model size. We present RelayS2S, a hybrid architecture that runs two paths in parallel upon turn detection. The fast path -- a duplex S2S model -- speculatively drafts a short response prefix that is streamed immediately to TTS for low-latency audio onset, while continuing to monitor live audio events. The slow path -- a cascaded ASR -> LLM pipeline -- generates a higher-quality continuation conditioned on the committed prefix, producing a seamless utterance. A lightweight learned verifier gates the handoff, committing the prefix when appropriate or falling back gracefully to the slow path alone. Experiments show that RelayS2S achieves P90 onset latency comparable to the S2S model while retaining 99% cascaded response quality in average score, with benefits growing as the slow-path model scales. Because the prefix handoff requires no architectural modification to either component, RelayS2S serves as a lightweight, drop-in addition to existing cascaded pipelines. Our code and data are publicly available at: https://github.com/mailong25/relays2s

Diffusion Large Language Models (dLLMs) deliver strong long-context processing capability in a non-autoregressive decoding paradigm. However, the considerable computational cost of bidirectional full attention limits the inference efficiency. Although sparse attention is promising, existing methods remain ineffective. This stems from the need to estimate attention importance for tokens yet to be decoded, while the unmasked token positions are unknown during diffusion. In this paper, we present Focus-dLLM, a novel training-free attention sparsification framework tailored for accurate and efficient long-context dLLM inference. Based on the finding that token confidence strongly correlates across adjacent steps, we first design a past confidence-guided indicator to predict unmasked regions. Built upon this, we propose a sink-aware pruning strategy to accurately estimate and remove redundant attention computation, while preserving highly influential attention sinks. To further reduce overhead, this strategy reuses identified sink locations across layers, leveraging the observed cross-layer consistency. Experimental results show that our method offers more than 29times lossless speedup under 32K context length. The code is publicly available at: https://github.com/Longxmas/Focus-dLLM

Reinforcement learning improves LLM reasoning, yet sparse delayed reward over long sequences makes token-level credit assignment the key bottleneck. We study the verifiable-reward setting, where the final answer is checkable and multiple responses can be drawn per prompt. Reasoning tasks in math and medical QA align with this setup, where only a few decision tokens significantly impact the outcome. PPO offers token-level advantages with a learned value model, but it is complex to train both the actor and critic models simultaneously, and it is not easily generalizable, as the token-level values from the critic model can make training prone to overfitting. GRPO is critic-free and supports verifiable rewards, but spreads a single sequence-level return across tokens and ignores branching. We introduce Prefix-to-Tree (P2T), a simple procedure that converts a group of responses into a prefix tree and computes nonparametric prefix values \(V(s)\) by aggregating descendant outcomes. Built on P2T, we propose TEMPO (\textbf{Tree-Estimated Mean Prefix Value for Policy Optimization}), a critic-free algorithm that augments the group-relative outcome signal of GRPO with branch-gated temporal-difference corrections derived from the tree. At non-branch tokens, the temporal-difference (TD) term is zero, so TEMPO reduces to GRPO; at branching tokens, it supplies precise token-level credit without a learned value network or extra judges/teachers. On Qwen3-1.7B/4B, TEMPO outperforms PPO and GRPO on in-distribution (MATH, MedQA) and out-of-distribution (GSM-HARD, AMC23, MedMCQA, MMLU-Medical) benchmarks, and reaches higher validation accuracy with roughly the same wall-clock time.

Tokenization is the first - and often underappreciated - layer of computation in language models. While Chain-of-Thought (CoT) prompting enables transformer models to approximate recurrent computation by externalizing intermediate steps, we show that the success of such reasoning is fundamentally bounded by the structure of tokenized inputs. This work presents a theoretical and empirical investigation into how tokenization schemes, particularly subword-based methods like byte-pair encoding (BPE), impede symbolic computation by merging or obscuring atomic reasoning units. We introduce the notion of Token Awareness to formalize how poor token granularity disrupts logical alignment and prevents models from generalizing symbolic procedures. Through systematic evaluation on arithmetic and symbolic tasks, we demonstrate that token structure dramatically affect reasoning performance, causing failure even with CoT, while atomically-aligned formats unlock strong generalization, allowing small models (e.g., GPT-4o-mini) to outperform larger systems (e.g., o1) in structured reasoning. Our findings reveal that symbolic reasoning ability in LLMs is not purely architectural, but deeply conditioned on token-level representations.

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.

Tokenization is an understudied and often neglected component of modern LLMs. Most published works use a single tokenizer for all experiments, often borrowed from another model, without performing ablations or analysis to optimize tokenization. Moreover, the tokenizer is generally kept unchanged when fine-tuning a base model. In this paper, we show that the size, pre-tokenization regular expression, and training data of a tokenizer can significantly impact the model's generation speed, effective context size, memory usage, and downstream performance. We train specialized Byte-Pair Encoding code tokenizers, and conduct extensive ablations on the impact of tokenizer design on the performance of LLMs for code generation tasks such as HumanEval and MBPP, and provide recommendations for tokenizer hyper-parameters selection and switching the tokenizer in a pre-trained LLM. We perform our experiments on models trained from scratch and from pre-trained models, verifying their applicability to a wide range of use-cases. We find that when fine-tuning on more than 50 billion tokens, we can specialize the tokenizer of a pre-trained LLM to obtain large gains in generation speed and effective context size.

Diffusion Language Models (DLMs) offer order-agnostic generation that can explore many possible decoding trajectories. However, current decoding methods commit to a single trajectory, limiting exploration in trajectory space. We introduce Order-Token Search to explore this space through jointly searching over generation order and token values. Its core is a likelihood estimator that scores denoising actions, enabling stable pruning and efficient exploration of diverse trajectories. Across mathematical reasoning and coding benchmarks, Order-Token Search consistently outperforms baselines on GSM8K, MATH500, Countdown, and HumanEval (3.1%, 3.8%, 7.9%, and 6.8% absolute over backbone), matching or surpassing diffu-GRPO post-trained d1-LLaDA. Our work establishes joint search as a key component for advancing decoding in DLMs.

The deployment of large language models (LLMs) in real-world applications is increasingly limited by their high inference cost. While recent advances in dynamic token-level computation allocation attempt to improve efficiency by selectively activating model components per token, existing methods rely on greedy routing--a myopic execute-or-skip mechanism that often leads to irreversible information loss and suboptimal token selection. This paper introduces informed routing, a new paradigm that proactively addresses these issues. The key insight is to assess not only a token's immediate importance but also its recoverability, i.e., how well its transformation can be approximated. To this end, we propose the Lightweight Feature Forecaster (LFF), a small predictive module that estimates a unit's output before routing decisions are made. This enables a flexible execute-or-approximate policy that preserves model fidelity while drastically reducing computation. Extensive experiments on both language modeling and reasoning tasks show that informed routing achieves state-of-the-art efficiency-performance trade-offs across multiple sparsity levels. Notably, even without final LoRA fine-tuning, our method matches or surpasses strong baselines that require full fine-tuning, all while reducing training time by over 50%. The code is available at: https://github.com/EIT-NLP/informed-routing

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

Agentic systems are evaluated on benchmarks where agents interact with environments to solve tasks. Most papers report a pass@1 score computed from a single run per task, assuming this gives a reliable performance estimate. We test this assumption by collecting 60,000 agentic trajectories on SWE-Bench-Verified, spanning three models and two scaffolds. We find substantial variance: single-run pass@1 estimates vary by 2.2 to 6.0 percentage points depending on which run is selected, with standard deviations exceeding 1.5 percentage points even at temperature 0. This variance has critical implications: reported improvements of 2--3 percentage points may reflect evaluation noise rather than genuine algorithmic progress. Through token-level analysis, we show that trajectories diverge early, often within the first few percent of tokens, and that these small differences cascade into different solution strategies. To enable reliable evaluation of agentic systems, we recommend three concrete practices: (1) estimate pass@1 from multiple independent runs per task, especially when measuring small improvements, (2) use statistical power analysis to determine the number of runs needed to detect expected effect sizes, and (3) consider metrics like pass@k (optimistic bound) and pass^k (pessimistic bound) with k>1 to better characterize the full performance envelope. While these practices increase evaluation cost, they are essential for distinguishing genuine scientific progress from statistical noise.

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/

Diffusion Large Language Models (dLLMs) offer fast, parallel token generation, but their standalone use is plagued by an inherent efficiency-quality tradeoff. We show that, if carefully applied, the attributes of dLLMs can actually be a strength for drafters in speculative decoding with autoregressive (AR) verifiers. Our core insight is that dLLM's speed from parallel decoding drastically lowers the risk of costly rejections, providing a practical mechanism to effectively realize the (elusive) lengthy drafts that lead to large speedups with speculative decoding. We present FailFast, a dLLM-based speculative decoding framework that realizes this approach by dynamically adapting its speculation length. It "fails fast" by spending minimal compute in hard-to-speculate regions to shrink speculation latency and "wins big" by aggressively extending draft lengths in easier regions to reduce verification latency (in many cases, speculating and accepting 70 tokens at a time!). Without any fine-tuning, FailFast delivers lossless acceleration of AR LLMs and achieves up to 4.9times speedup over vanilla decoding, 1.7times over the best naive dLLM drafter, and 1.4times over EAGLE-3 across diverse models and workloads. We open-source FailFast at https://github.com/ruipeterpan/failfast.

Practitioners have consistently observed three puzzling phenomena in transformer-based large language models (LLMs): attention sinks, value-state drains, and residual-state peaks, collectively referred to as extreme-token phenomena. These phenomena are characterized by certain so-called "sink tokens" receiving disproportionately high attention weights, exhibiting significantly smaller value states, and having much larger residual-state norms than those of other tokens. These extreme tokens give rise to various challenges in LLM inference, quantization, and interpretability. We elucidate the mechanisms behind extreme-token phenomena. First, we show that these phenomena arise in very simple architectures -- transformers with one to three layers -- trained on a toy model, the Bigram-Backcopy (BB) task. In this setting, we identify an active-dormant mechanism, where attention heads become sinks for specific input domains while remaining non-sinks for others. Our theoretical analysis of the training dynamics reveals that these phenomena are driven by a mutual reinforcement mechanism. Building on these insights, we propose strategies to mitigate extreme-token phenomena during pretraining, including replacing softmax with ReLU and Adam with SGD. Next, we extend our analysis to pretrained LLMs, including Llama and OLMo, showing that many attention heads exhibit a similar active-dormant mechanism as in the BB task, and that the mutual reinforcement mechanism also governs the emergence of extreme-token phenomena during LLM pretraining. Our results reveal that many of the static and dynamic properties of extreme-token phenomena predicted by the BB task align with observations in pretrained LLMs.

Carefully engineered system prompts play a critical role in guiding the behavior of LLM agents, but their considerable length introduces significant drawbacks, including increased inference latency, higher computational cost, and reduced effective context length. This raises the question of whether such lengthy prompts can be replaced by a drastically reduced number of tokens while preserving their behavioral effect on downstream tasks. To enable this, we propose a lightweight three-stage training framework that learns a single prompt-specific Behavior-Equivalent token ([BE]). The framework first trains [BE] to encode the natural-language content of the original system prompt via reconstruction, and then distills the prompt 's downstream behavior into this single token. Importantly, our method requires no access to model internals, no auxiliary compression models, and no labeled responses. Empirical evaluations on three datasets show that a single [BE] token achieves up to a 3000x reduction in prompt length, while retaining about 98% of the downstream performance of the original system prompts. This substantially reduces inference cost and leaves almost the entire context window available for user inputs.

While speculative decoding has recently appeared as a promising direction for accelerating the inference of large language models (LLMs), the speedup and scalability are strongly bounded by the token acceptance rate. Prevalent methods usually organize predicted tokens as independent chains or fixed token trees, which fails to generalize to diverse query distributions. In this paper, we propose DySpec, a faster speculative decoding algorithm with a novel dynamic token tree structure. We begin by bridging the draft distribution and acceptance rate from intuitive and empirical clues, and successfully show that the two variables are strongly correlated. Based on this, we employ a greedy strategy to dynamically expand the token tree at run time. Theoretically, we show that our method can achieve optimal results under mild assumptions. Empirically, DySpec yields a higher acceptance rate and speedup than fixed trees. DySpec can drastically improve the throughput and reduce the latency of token generation across various data distribution and model sizes, which significantly outperforms strong competitors, including Specinfer and Sequoia. Under low temperature setting, DySpec can improve the throughput up to 9.1times and reduce the latency up to 9.4times on Llama2-70B. Under high temperature setting, DySpec can also improve the throughput up to 6.21times, despite the increasing difficulty of speculating more than one token per step for draft model.

Current language models (LMs) use a fixed, static subword tokenizer. This choice, often taken for granted, typically results in degraded efficiency and capabilities in languages other than English, and makes it challenging to apply LMs to new domains or languages. To address these issues, we propose retrofitting LMs with dynamic tokenization: a way to dynamically decide on token boundaries based on the input text. For encoder-style models, we introduce a subword-merging algorithm inspired by byte-pair encoding (BPE), but at a batch level. We merge frequent subword sequences in a batch, then apply a pretrained embedding-prediction hypernetwork to compute the token embeddings on-the-fly. When applied with word-level boundaries, this on average reduces token sequence lengths by >20% across 14 languages on XNLI with XLM-R while degrading its task performance by less than 2%. For decoder-style models, we apply dynamic tokenization in two ways: 1) for prefilling, maintaining performance of Mistral-7B almost completely with up to 40% sequence reduction - relative to the word-level; and 2) via an approximate nearest neighbor index, achieving fast generation with a one million token vocabulary, demonstrating scalability to even larger, dynamic vocabularies. Overall, our findings show that dynamic tokenization substantially improves inference speed and promotes fairness across languages, making a leap towards overcoming the limitations of static tokenization and enabling more equitable and adaptable LMs.

Language models (LMs) are bound to their tokenizer, which maps raw text to a sequence of vocabulary items (tokens). This restricts their flexibility: for example, LMs trained primarily on English may still perform well in other natural and programming languages, but have vastly decreased efficiency due to their English-centric tokenizer. To mitigate this, we should be able to swap the original LM tokenizer with an arbitrary one, on the fly, without degrading performance. Hence, in this work we define a new problem: Zero-Shot Tokenizer Transfer (ZeTT). The challenge at the core of ZeTT is finding embeddings for the tokens in the vocabulary of the new tokenizer. Since prior heuristics for initializing embeddings often perform at chance level in a ZeTT setting, we propose a new solution: we train a hypernetwork taking a tokenizer as input and predicting the corresponding embeddings. We empirically demonstrate that the hypernetwork generalizes to new tokenizers both with encoder (e.g., XLM-R) and decoder LLMs (e.g., Mistral-7B). Our method comes close to the original models' performance in cross-lingual and coding tasks while markedly reducing the length of the tokenized sequence. We also find that the remaining gap can be quickly closed by continued training on less than 1B tokens. Finally, we show that a ZeTT hypernetwork trained for a base (L)LM can also be applied to fine-tuned variants without extra training. Overall, our results make substantial strides toward detaching LMs from their tokenizer.

Diffusion Language Models (DLMs) have emerged as a compelling alternative to autoregressive approaches, enabling parallel text generation with competitive performance. Despite these advantages, there is a critical instability in DLMs: the moving sink phenomenon. Our analysis indicates that sink tokens exhibit low-norm representations in the Transformer's value space, and that the moving sink phenomenon serves as a protective mechanism in DLMs to prevent excessive information mixing. However, their unpredictable positions across diffusion steps undermine inference robustness. To resolve this, we propose a simple but effective extra sink token implemented via a modified attention mask. Specifically, we introduce a special token constrained to attend solely to itself, while remaining globally visible to all other tokens. Experimental results demonstrate that introducing a single extra token stabilizes attention sinks, substantially improving model performance. Crucially, further analysis confirms that the effectiveness of this token is independent of its position and characterized by negligible semantic content, validating its role as a robust and dedicated structural sink.

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.

Improving time-to-first-token (TTFT) is an essentially important objective in modern large language model (LLM) inference engines. Optimizing TTFT directly results in higher maximal QPS and meets the requirements of many critical applications. However, boosting TTFT is notoriously challenging since it is compute-bounded and the performance bottleneck shifts from the self-attention that many prior works focus on to the MLP part. In this work, we present SpecPrefill, a training free framework that accelerates the inference TTFT for both long and medium context queries based on the following insight: LLMs are generalized enough to preserve the quality given only a carefully chosen subset of prompt tokens. At its core, SpecPrefill leverages a lightweight model to speculate locally important tokens based on the context. These tokens, along with the necessary positional information, are then sent to the main model for processing. We evaluate SpecPrefill with a diverse set of tasks, followed by a comprehensive benchmarking of performance improvement both in a real end-to-end setting and ablation studies. SpecPrefill manages to serve Llama-3.1-405B-Instruct-FP8 with up to 7times maximal end-to-end QPS on real downstream tasks and 7.66times TTFT improvement.

Autoregressive Large Language Models (e.g., LLaMa, GPTs) are omnipresent achieving remarkable success in language understanding and generation. However, such impressive capability typically comes with a substantial model size, which presents significant challenges for autoregressive token-by-token generation. To mitigate computation overload incurred during generation, several early-exit and layer-dropping strategies have been proposed. Despite some promising success due to the redundancy across LLMs layers on metrics like Rough-L/BLUE, our careful knowledge-intensive evaluation unveils issues such as generation collapse, hallucination of wrong facts, and noticeable performance drop even at the trivial exit ratio of 10-15% of layers. We attribute these errors primarily to ineffective handling of the KV cache through state copying during early-exit. In this work, we observed the saturation of computationally expensive feed-forward blocks of LLM layers and proposed FFN-SkipLLM, which is a novel fine-grained skip strategy of autoregressive LLMs. More specifically, FFN-SkipLLM is an input-adaptive feed-forward skipping strategy that can skip 25-30% of FFN blocks of LLMs with marginal change in performance on knowledge-intensive generation tasks without any requirement to handle KV cache. Our extensive experiments and ablation across benchmarks like MT-Bench, Factoid-QA, and variable-length text summarization illustrate how our simple and ease-at-use method can facilitate faster autoregressive decoding.

One of the most striking findings in modern research on large language models (LLMs) is that scaling up compute during training leads to better results. However, less attention has been given to the benefits of scaling compute during inference. This survey focuses on these inference-time approaches. We explore three areas under a unified mathematical formalism: token-level generation algorithms, meta-generation algorithms, and efficient generation. Token-level generation algorithms, often called decoding algorithms, operate by sampling a single token at a time or constructing a token-level search space and then selecting an output. These methods typically assume access to a language model's logits, next-token distributions, or probability scores. Meta-generation algorithms work on partial or full sequences, incorporating domain knowledge, enabling backtracking, and integrating external information. Efficient generation methods aim to reduce token costs and improve the speed of generation. Our survey unifies perspectives from three research communities: traditional natural language processing, modern LLMs, and machine learning systems.

This paper localizes the policy routing mechanism in alignment-trained language models. An intermediate-layer attention gate reads detected content and triggers deeper amplifier heads that boost the signal toward refusal. In smaller models the gate and amplifier are single heads; at larger scale they become bands of heads across adjacent layers. The gate contributes under 1% of output DLA, but interchange testing (p<0.001) and knockout cascade confirm it is causally necessary. Interchange screening at n>=120 detects the same motif in twelve models from six labs (2B to 72B), though specific heads differ by lab. Per-head ablation weakens up to 58x at 72B and misses gates that interchange identifies; interchange is the only reliable audit at scale. Modulating the detection-layer signal continuously controls policy from hard refusal through evasion to factual answering. On safety prompts the same intervention turns refusal into harmful guidance, showing the safety-trained capability is gated by routing rather than removed. Thresholds vary by topic and by input language, and the circuit relocates across generations within a family while behavioral benchmarks register no change. Routing is early-commitment: the gate commits at its own layer before deeper layers finish processing the input. Under an in-context substitution cipher, gate interchange necessity collapses 70 to 99% across three models and the model switches to puzzle-solving. Injecting the plaintext gate activation into the cipher forward pass restores 48% of refusals in Phi-4-mini, localizing the bypass to the routing interface. A second method, cipher contrast analysis, uses plain/cipher DLA differences to map the full cipher-sensitive routing circuit in O(3n) forward passes. Any encoding that defeats detection-layer pattern matching bypasses the policy regardless of whether deeper layers reconstruct the content.

Transformer-based pre-trained language models (PLMs) mostly suffer from excessive overhead despite their advanced capacity. For resource-constrained devices, there is an urgent need for a spatially and temporally efficient model which retains the major capacity of PLMs. However, existing statically compressed models are unaware of the diverse complexities between input instances, potentially resulting in redundancy and inadequacy for simple and complex inputs. Also, miniature models with early exiting encounter challenges in the trade-off between making predictions and serving the deeper layers. Motivated by such considerations, we propose a collaborative optimization for PLMs that integrates static model compression and dynamic inference acceleration. Specifically, the PLM is slenderized in width while the depth remains intact, complementing layer-wise early exiting to speed up inference dynamically. To address the trade-off of early exiting, we propose a joint training approach that calibrates slenderization and preserves contributive structures to each exit instead of only the final layer. Experiments are conducted on GLUE benchmark and the results verify the Pareto optimality of our approach at high compression and acceleration rate with 1/8 parameters and 1/19 FLOPs of BERT.

Recently, the efficient deployment and acceleration of powerful vision transformers (ViTs) on resource-limited edge devices for providing multimedia services have become attractive tasks. Although early exiting is a feasible solution for accelerating inference, most works focus on convolutional neural networks (CNNs) and transformer models in natural language processing (NLP).Moreover, the direct application of early exiting methods to ViTs may result in substantial performance degradation. To tackle this challenge, we systematically investigate the efficacy of early exiting in ViTs and point out that the insufficient feature representations in shallow internal classifiers and the limited ability to capture target semantic information in deep internal classifiers restrict the performance of these methods. We then propose an early exiting framework for general ViTs termed LGViT, which incorporates heterogeneous exiting heads, namely, local perception head and global aggregation head, to achieve an efficiency-accuracy trade-off. In particular, we develop a novel two-stage training scheme, including end-to-end training and self-distillation with the backbone frozen to generate early exiting ViTs, which facilitates the fusion of global and local information extracted by the two types of heads. We conduct extensive experiments using three popular ViT backbones on three vision datasets. Results demonstrate that our LGViT can achieve competitive performance with approximately 1.8 times speed-up.

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.

Auto-regressive Large Language Models (LLMs) demonstrate remarkable performance across different domains such as vision and language processing. However, due to sequential processing through a stack of transformer layers, autoregressive decoding faces significant computation/latency challenges, particularly in resource-constrained environments like mobile and edge devices. Existing approaches in literature that aim to improve latency via skipping layers have two distinct flavors - 1) Early exit, and 2) Input-agnostic heuristics where tokens exit at pre-determined layers irrespective of input sequence. Both the above strategies have limitations - the former cannot be applied to handle KV Caching necessary for speed-ups in modern framework and the latter does not capture the variation in layer importance across tasks or more generally, across input sequences. To address both limitations, we propose FiRST, an algorithm that reduces inference latency by using layer-specific routers to select a subset of transformer layers adaptively for each input sequence - the prompt (during the prefill stage) decides which layers will be skipped during decoding. FiRST preserves compatibility with KV caching enabling faster inference while being quality-aware. FiRST is model-agnostic and can be easily enabled on any pre-trained LLM. Our approach reveals that input adaptivity is critical - indeed, different task-specific middle layers play a crucial role in evolving hidden representations depending on tasks. Extensive experiments show that FiRST significantly reduces latency while outperforming other layer selection strategies in quality metics. It retains competitive performance to base model (without layer skipping) and in some cases, even improves upon it. FiRST is thus a promising and efficient solution for LLM deployment in low-resource environments.

Reinforcement learning (RL) post-training has increasingly demonstrated strong ability to elicit reasoning behaviors in large language models (LLMs). For training efficiency, rollouts are typically generated in an off-policy manner using an older sampling policy and then used to update the current target policy. To correct the resulting discrepancy between the sampling and target policies, most existing RL objectives rely on a token-level importance sampling ratio, primarily due to its computational simplicity and numerical stability. However, we observe that token-level correction often leads to unstable training dynamics when the degree of off-policyness is large. In this paper, we revisit LLM policy optimization under off-policy conditions and show that the theoretically rigorous correction term is the prefix importance ratio, and that relaxing it to a token-level approximation can induce instability in RL post-training. To stabilize LLM optimization under large off-policy drift, we propose a simple yet effective objective, Minimum Prefix Ratio (MinPRO). MinPRO replaces the unstable cumulative prefix ratio with a non-cumulative surrogate based on the minimum token-level ratio observed in the preceding prefix. Extensive experiments on both dense and mixture-of-experts LLMs, across multiple mathematical reasoning benchmarks, demonstrate that MinPRO substantially improves training stability and peak performance in off-policy regimes.

The Mixture of Experts (MoE) is an effective architecture for scaling large language models by leveraging sparse expert activation, optimizing the trade-off between performance and efficiency. However, under expert parallelism, MoE suffers from inference inefficiencies due to imbalanced token-to-expert assignment, where some experts are overloaded while others remain underutilized. This imbalance leads to poor resource utilization and increased latency, as the most burdened expert dictates the overall delay, a phenomenon we define as the \textit{Straggler Effect}. To mitigate this, we propose Capacity-Aware Inference, including two key techniques: (1) \textit{Capacity-Aware Token Drop}, which discards overloaded tokens to regulate the maximum latency of MoE, and (2) \textit{Capacity-Aware Token Reroute}, which reallocates overflowed tokens to underutilized experts, balancing the token distribution. These techniques collectively optimize both high-load and low-load expert utilization, leading to a more efficient MoE inference pipeline. Extensive experiments demonstrate the effectiveness of our methods, showing significant improvements in inference efficiency, e.g., 0.2\% average performance increase and a 1.94times inference speedup on Mixtral-8times7B-Instruct.

Frontier LLMs are increasingly utilised across academia, society and industry. A commonly used unit for comparing models, their inputs and outputs, and estimating inference pricing is the token. In general, tokens are used as a stable currency, assumed to be broadly consistent across tokenizers and contexts, enabling direct comparisons. However, tokenization varies significantly across models and domains of text, making naive interpretation of token counts problematic. We quantify this variation by providing a comprehensive empirical analysis of tokenization, exploring the compression of sequences to tokens across different distributions of textual data. Our analysis challenges commonly held heuristics about token lengths, finding them to be overly simplistic. We hope the insights of our study add clarity and intuition toward tokenization in contemporary LLMs.

Large Language Models have revolutionized natural language processing, yet serving them efficiently in data centers remains challenging due to mixed workloads comprising latency-sensitive (LS) and best-effort (BE) jobs. Existing inference systems employ iteration-level first-come-first-served scheduling, causing head-of-line blocking when BE jobs delay LS jobs. We introduce QLLM, a novel inference system designed for Mixture of Experts (MoE) models, featuring a fine-grained, priority-aware preemptive scheduler. QLLM enables expert-level preemption, deferring BE job execution while minimizing LS time-to-first-token (TTFT). Our approach removes iteration-level scheduling constraints, enabling the scheduler to preempt jobs at any layer based on priority. Evaluations on an Nvidia A100 GPU show that QLLM significantly improves performance. It reduces LS TTFT by an average of 65.5times and meets the SLO at up to 7 requests/sec, whereas the baseline fails to do so under the tested workload. Additionally, it cuts LS turnaround time by up to 12.8times without impacting throughput. QLLM is modular, extensible, and seamlessly integrates with Hugging Face MoE models.

Inference with transformer-based language models begins with a prompt processing step. In this step, the model generates the first output token and stores the KV cache needed for future generation steps. This prompt processing step can be computationally expensive, taking 10s of seconds or more for billion-parameter models on edge devices when prompt lengths or batch sizes rise. This degrades user experience by introducing significant latency into the model's outputs. To reduce the time spent producing the first output (known as the ``time to first token'', or TTFT) of a pretrained model, we introduce a novel method called KV Prediction. In our method, a small auxiliary model is used to process the prompt and produce an approximation of the KV cache used by a base model. This approximated KV cache is then used with the base model for autoregressive generation without the need to query the auxiliary model again. We demonstrate that our method produces a pareto-optimal efficiency-accuracy trade-off when compared to baselines. On TriviaQA, we demonstrate relative accuracy improvements in the range of 15%-50% across a range of TTFT FLOPs budgets. We also demonstrate accuracy improvements of up to 30% on HumanEval python code completion at fixed TTFT FLOPs budgets. Additionally, we benchmark models on an Apple M2 Pro CPU and demonstrate that our improvement in FLOPs translates to a TTFT speedup on hardware. We release our code at https://github.com/apple/corenet/tree/main/projects/kv-prediction .

Large language models (LLMs) are typically developed through large-scale pre-training followed by task-specific fine-tuning. Recent advances highlight the importance of an intermediate mid-training stage, where models undergo multiple annealing-style phases that refine data quality, adapt optimization schedules, and extend context length. This stage mitigates diminishing returns from noisy tokens, stabilizes convergence, and expands model capability in late training. Its effectiveness can be explained through gradient noise scale, the information bottleneck, and curriculum learning, which together promote generalization and abstraction. Despite widespread use in state-of-the-art systems, there has been no prior survey of mid-training as a unified paradigm. We introduce the first taxonomy of LLM mid-training spanning data distribution, learning-rate scheduling, and long-context extension. We distill practical insights, compile evaluation benchmarks, and report gains to enable structured comparisons across models. We also identify open challenges and propose avenues for future research and practice.

Chain-of-thoughts (CoT) requires large language models (LLMs) to generate intermediate steps before reaching the final answer, and has been proven effective to help LLMs solve complex reasoning tasks. However, the inner mechanism of CoT still remains largely unclear. In this paper, we empirically study the role of CoT tokens in LLMs on two compositional tasks: multi-digit multiplication and dynamic programming. While CoT is essential for solving these problems, we find that preserving only tokens that store intermediate results would achieve comparable performance. Furthermore, we observe that storing intermediate results in an alternative latent form will not affect model performance. We also randomly intervene some values in CoT, and notice that subsequent CoT tokens and the final answer would change correspondingly. These findings suggest that CoT tokens may function like variables in computer programs but with potential drawbacks like unintended shortcuts and computational complexity limits between tokens. The code and data are available at https://github.com/solitaryzero/CoTs_are_Variables.

Large Language Models (LLMs) have exhibited remarkable performance on reasoning tasks. They utilize autoregressive token generation to construct reasoning trajectories, enabling the development of a coherent chain of thought. In this work, we explore the impact of individual tokens on the final outcomes of reasoning tasks. We identify the existence of ``critical tokens'' that lead to incorrect reasoning trajectories in LLMs. Specifically, we find that LLMs tend to produce positive outcomes when forced to decode other tokens instead of critical tokens. Motivated by this observation, we propose a novel approach - cDPO - designed to automatically recognize and conduct token-level rewards for the critical tokens during the alignment process. Specifically, we develop a contrastive estimation approach to automatically identify critical tokens. It is achieved by comparing the generation likelihood of positive and negative models. To achieve this, we separately fine-tune the positive and negative models on various reasoning trajectories, consequently, they are capable of identifying identify critical tokens within incorrect trajectories that contribute to erroneous outcomes. Moreover, to further align the model with the critical token information during the alignment process, we extend the conventional DPO algorithms to token-level DPO and utilize the differential likelihood from the aforementioned positive and negative model as important weight for token-level DPO learning.Experimental results on GSM8K and MATH500 benchmarks with two-widely used models Llama-3 (8B and 70B) and deepseek-math (7B) demonstrate the effectiveness of the propsoed approach cDPO.

Conditional computation is a popular strategy to make Transformers more efficient. Existing methods often target individual modules (e.g., mixture-of-experts layers) or skip layers independently of one another. However, interpretability research has demonstrated that the middle layers of Transformers exhibit greater redundancy, and that early layers aggregate information into token positions. Guided by these insights, we propose a novel architecture that dynamically skips a variable number of layers from the middle outward. In particular, a learned gating mechanism determines whether to bypass a symmetric span of central blocks based on the input, and a gated attention mechanism prevents subsequent tokens from attending to skipped token positions. Residual norms are controlled with a 'sandwich' or 'perilayernorm' scheme and gate sparsity with an adaptive regularization loss. We had aimed to reduce compute requirements for 'simpler' tokens and potentially foster an emergent multi-level representational hierarchy but, at the scales investigated, our approach does not achieve improvements in the trade-off between validation cross-entropy and estimated FLOPs compared to dense baselines with fewer layers. We release our code at https://github.com/tim-lawson/skip-middle.

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.

Modern Large Language Models achieve impressive reasoning capabilities with long Chain of Thoughts, but they incur substantial computational cost during inference, and this motivates techniques to improve the performance-cost ratio. Among these techniques, Speculative Decoding accelerates inference by employing a fast but inaccurate draft model to autoregressively propose tokens, which are then verified in parallel by a more capable target model. However, due to unnecessary rejections caused by token mismatches in semantically equivalent steps, traditional token-level Speculative Decoding struggles in reasoning tasks. Although recent works have shifted to step-level semantic verification, which improve efficiency by accepting or rejecting entire reasoning steps, existing step-level methods still regenerate many rejected steps with little improvement, wasting valuable target compute. To address this challenge, we propose Arbitrage, a novel step-level speculative generation framework that routes generation dynamically based on the relative advantage between draft and target models. Instead of applying a fixed acceptance threshold, Arbitrage uses a lightweight router trained to predict when the target model is likely to produce a meaningfully better step. This routing approximates an ideal Arbitrage Oracle that always chooses the higher-quality step, achieving near-optimal efficiency-accuracy trade-offs. Across multiple mathematical reasoning benchmarks, Arbitrage consistently surpasses prior step-level Speculative Decoding baselines, reducing inference latency by up to sim2times at matched accuracy.

Reasoning models excel by generating long chain-of-thoughts, but decoding the resulting thousands of tokens is slow. Token-level speculative decoding (SD) helps, but its benefit is capped, because the chance that an entire gamma-token guess is correct falls exponentially as gamma grows. This means allocating more compute for longer token drafts faces an algorithmic ceiling -- making the speedup modest and hardware-agnostic. We raise this ceiling with Lookahead Reasoning, which exploits a second, step-level layer of parallelism. Our key insight is that reasoning models generate step-by-step, and each step needs only to be semantically correct, not exact token matching. In Lookahead Reasoning, a lightweight draft model proposes several future steps; the target model expands each proposal in one batched pass, and a verifier keeps semantically correct steps while letting the target regenerate any that fail. Token-level SD still operates within each reasoning step, so the two layers of parallelism multiply. We show Lookahead Reasoning lifts the peak speedup of SD both theoretically and empirically. Across GSM8K, AIME, and other benchmarks, Lookahead Reasoning improves the speedup of SD from 1.4x to 2.1x while preserving answer quality, and its speedup scales better with additional GPU throughput. Our code is available at https://github.com/hao-ai-lab/LookaheadReasoning

Reinforcement Learning (RL) has significantly improved large language model reasoning, but existing RL fine-tuning methods rely heavily on heuristic techniques such as entropy regularization and reweighting to maintain stability. In practice, they often experience late-stage performance collapse, leading to degraded reasoning quality and unstable training. We derive that the magnitude of token-wise policy gradients in RL is negatively correlated with token probability and local policy entropy. Building on this result, we prove that training instability is driven by a tiny fraction of tokens, approximately 0.01\%, which we term spurious tokens. When such tokens appear in correct responses, they contribute little to the reasoning outcome but inherit the full sequence-level reward, leading to abnormally amplified gradient updates. Motivated by this observation, we propose Spurious-Token-Aware Policy Optimization (STAPO) for large-scale model refining, which selectively masks such updates and renormalizes the loss over valid tokens. Across six mathematical reasoning benchmarks using Qwen 1.7B, 8B, and 14B base models, STAPO consistently demonstrates superior entropy stability and achieves an average performance improvement of 7.13\% over GRPO, 20-Entropy and JustRL.

The autoregressive nature of large language models (LLMs) limits inference speed. Each forward pass generates only a single token and is often bottlenecked by memory bandwidth. Speculative decoding alleviates this issue using a draft-then-verify approach to accelerate token generation. However, the overhead introduced during the draft phase and the training cost of the draft model limit the efficiency and adaptability of speculative decoding. In this work, we introduce PARallel Draft (PARD), a novel speculative decoding method that enables low-cost adaptation of autoregressive draft models into parallel draft models. PARD enhances inference efficiency by predicting multiple future tokens in a single forward pass of the draft phase, and incorporates a conditional drop token method to accelerate training. Its target-independence property allows a single draft model to be applied to an entire family of different models, minimizing the adaptation cost. Our proposed conditional drop token method can improves draft model training efficiency by 3x. On our optimized inference framework, PARD accelerates LLaMA3.1-8B inference by 4.08x, achieving 311.5 tokens per second.

Token reduction is an effective way to accelerate long-video vision-language models (VLMs), but most existing methods are designed for dense Transformers and do not directly account for hybrid architectures that interleave attention with linear-time state-space blocks (e.g., Mamba). We study query-conditioned token reduction for hybrid video VLMs and analyze reduction behavior through two properties: layerwise sparsity (how many tokens capture query-relevant information) and importance stability (whether token-importance rankings persist across depth). Although token importance is sparse within each layer, the set of important tokens changes across layers, so aggressive early pruning is unreliable. Motivated by this, we propose a low-to-high progressive reduction schedule and a unified language-aware scoring mechanism for both attention and Mamba blocks (using an implicit-attention proxy for Mamba), enabling all-layer token reduction in hybrids. Under an aggressive compression setting (retaining 25% of visual tokens), our approach delivers substantial prefilling speedups (3.8--4.2x) with near-baseline accuracy at test time, and light finetuning under reduction further improves performance on long-context video benchmarks.

We present Universal Conditional Logic (UCL), a mathematical framework for prompt optimization that transforms prompt engineering from heuristic practice into systematic optimization. Through systematic evaluation (N=305, 11 models, 4 iterations), we demonstrate significant token reduction (29.8%, t(10)=6.36, p < 0.001, Cohen's d = 2.01) with corresponding cost savings. UCL's structural overhead function O_s(A) explains version-specific performance differences through the Over-Specification Paradox: beyond threshold S* = 0.509, additional specification degrades performance quadratically. Core mechanisms -- indicator functions (I_i in {0,1}), structural overhead (O_s = gamma * sum(ln C_k)), early binding -- are validated. Notably, optimal UCL configuration varies by model architecture -- certain models (e.g., Llama 4 Scout) require version-specific adaptations (V4.1). This work establishes UCL as a calibratable framework for efficient LLM interaction, with model-family-specific optimization as a key research direction.

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

With the success of language pretraining, it is highly desirable to develop more efficient architectures of good scalability that can exploit the abundant unlabeled data at a lower cost. To improve the efficiency, we examine the much-overlooked redundancy in maintaining a full-length token-level presentation, especially for tasks that only require a single-vector presentation of the sequence. With this intuition, we propose Funnel-Transformer which gradually compresses the sequence of hidden states to a shorter one and hence reduces the computation cost. More importantly, by re-investing the saved FLOPs from length reduction in constructing a deeper or wider model, we further improve the model capacity. In addition, to perform token-level predictions as required by common pretraining objectives, Funnel-Transformer is able to recover a deep representation for each token from the reduced hidden sequence via a decoder. Empirically, with comparable or fewer FLOPs, Funnel-Transformer outperforms the standard Transformer on a wide variety of sequence-level prediction tasks, including text classification, language understanding, and reading comprehension. The code and pretrained checkpoints are available at https://github.com/laiguokun/Funnel-Transformer.

Linearizing pretrained large language models (LLMs) primarily relies on intra-layer hybrid attention mechanisms to alleviate the quadratic complexity of standard softmax attention. Existing methods perform token routing based on sliding-window partitions, resulting in position-based selection and fails to capture token-specific global importance. Meanwhile, linear attention further suffers from distribution shift caused by learnable feature maps that distort pretrained feature magnitudes. Motivated by these limitations, we propose STILL, an intra-layer hybrid linearization framework for efficiently linearizing LLMs. STILL introduces a Self-Saliency Score with strong local-global consistency, enabling accurate token selection using sliding-window computation, and retains salient tokens for sparse softmax attention while summarizing the remaining context via linear attention. To preserve pretrained representations, we design a Norm-Preserved Feature Map (NP-Map) that decouples feature direction from magnitude and reinjects pretrained norms. We further adopt a unified training-inference architecture with chunk-wise parallelization and delayed selection to improve hardware efficiency. Experiments show that STILL matches or surpasses the original pretrained model on commonsense and general reasoning tasks, and achieves up to a 86.2% relative improvement over prior linearized attention methods on long-context benchmarks.

This work introduces EE-Tuning, a lightweight and economical solution to training/tuning early-exit large language models (LLMs). In contrast to the common approach of full-parameter pre-training, EE-Tuning augments any pre-trained (and possibly fine-tuned) standard LLM with additional early-exit layers that are tuned in a parameter-efficient manner, which requires significantly less computational resources and training data. Our implementation of EE-Tuning achieves outstanding training efficiency via extensive performance optimizations, as well as scalability due to its full compatibility with 3D parallelism. Results of systematic experiments validate the efficacy of EE-Tuning, confirming that effective early-exit LLM inference can be achieved with a limited training budget. In hope of making early-exit LLMs accessible to the community, we release the source code of our implementation of EE-Tuning at https://github.com/pan-x-c/EE-LLM.

Diffusion Large Language Models (dLLMs) enable breakthroughs in reasoning and parallel decoding but suffer from prohibitive quadratic computational complexity and memory overhead during inference. Current caching techniques accelerate decoding by storing full-layer states, yet impose substantial memory usage that limit long-context applications. Our analysis of attention patterns in dLLMs reveals persistent cross-layer sparsity, with pivotal tokens remaining salient across decoding steps and low-relevance tokens staying unimportant, motivating selective cache eviction. We propose Sparse-dLLM, the first training-free framework integrating dynamic cache eviction with sparse attention via delayed bidirectional sparse caching. By leveraging the stability of token saliency over steps, it retains critical tokens and dynamically evicts unimportant prefix/suffix entries using an attention-guided strategy. Extensive experiments on LLaDA and Dream series demonstrate Sparse-dLLM achieves up to 10times higher throughput than vanilla dLLMs, with comparable performance and similar peak memory costs, outperforming previous methods in efficiency and effectiveness.

Transformer-based models generally allocate the same amount of computation for each token in a given sequence. We develop a simple but effective "token dropping" method to accelerate the pretraining of transformer models, such as BERT, without degrading its performance on downstream tasks. In short, we drop unimportant tokens starting from an intermediate layer in the model to make the model focus on important tokens; the dropped tokens are later picked up by the last layer of the model so that the model still produces full-length sequences. We leverage the already built-in masked language modeling (MLM) loss to identify unimportant tokens with practically no computational overhead. In our experiments, this simple approach reduces the pretraining cost of BERT by 25% while achieving similar overall fine-tuning performance on standard downstream tasks.

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

Recent advancements in reinforcement learning from human feedback have shown that utilizing fine-grained token-level reward models can substantially enhance the performance of Proximal Policy Optimization (PPO) in aligning large language models. However, it is challenging to leverage such token-level reward as guidance for Direct Preference Optimization (DPO), since DPO is formulated as a sequence-level bandit problem. To address this challenge, this work decomposes the sequence-level PPO into a sequence of token-level proximal policy optimization problems and then frames the problem of token-level PPO with token-level reward guidance, from which closed-form optimal token-level policy and the corresponding token-level reward can be derived. Using the obtained reward and Bradley-Terry model, this work establishes a framework of computable loss functions with token-level reward guidance for DPO, and proposes a practical reward guidance based on the induced DPO reward. This formulation enables different tokens to exhibit varying degrees of deviation from reference policy based on their respective rewards. Experiment results demonstrate that our method achieves substantial performance improvements over DPO, with win rate gains of up to 7.5 points on MT-Bench, 6.2 points on AlpacaEval 2, and 4.3 points on Arena-Hard. Code is available at https://github.com/dvlab-research/TGDPO.

Diffusion Large Language Models (DLLMs) enable fully parallel token decoding but often remain impractical at inference time due to the many denoising iterations required to refine an information-free, fully masked initialization into coherent text. Most existing acceleration methods focus on traversing this generative trajectory more efficiently via improved solvers or sampling strategies. We advance a complementary perspective: shorten the trajectory itself by starting closer to the target distribution through context-aware initialization. We propose a training-free interface that injects prompt-conditioned priors from a lightweight auxiliary model into the diffusion initialization, and instantiate it with two mechanisms: discrete token injection and representation-level embedding interpolation. Because injected priors can be imperfect and unmask-only decoding can over-commit early, we also introduce a simple confidence-based remasking mechanism as a form of prior skepticism. Preliminary evidence on GSM8K suggests that context-aware initialization can substantially reduce denoising iterations (about 35\% fewer function evaluations in our setting), while also exposing a key open challenge: naive warm-starting can degrade final accuracy relative to strong diffusion baselines. We use these findings to motivate a research agenda around calibration, revision mechanisms, and representation alignment for reliable warm-started diffusion decoding.

Large Language Models (LLMs) are widely deployed in real-world applications, yet little is known about their training dynamics at the token level. Evaluation typically relies on aggregated training loss, measured at the batch level, which overlooks subtle per-token biases arising from (i) varying token-level dynamics and (ii) structural biases introduced by hyperparameters. While weight decay is commonly used to stabilize training, we reveal that it silently introduces performance biases detectable only at the token level. In fact, we empirically show across different dataset sizes, model architectures and sizes ranging from 270M to 3B parameters that as weight decay increases, low-frequency tokens are disproportionately depreciated. This is particularly concerning, as these neglected low-frequency tokens represent the vast majority of the token distribution in most languages, calling for novel regularization techniques that ensure fairness across all available tokens.

Large Language Models (LLMs) have revolutionized a wide range of domains such as natural language processing, computer vision, and multi-modal tasks due to their ability to comprehend context and perform logical reasoning. However, the computational and memory demands of LLMs, particularly during inference, pose significant challenges when scaling them to real-world, long-context, and real-time applications. Key-Value (KV) cache management has emerged as a critical optimization technique for accelerating LLM inference by reducing redundant computations and improving memory utilization. This survey provides a comprehensive overview of KV cache management strategies for LLM acceleration, categorizing them into token-level, model-level, and system-level optimizations. Token-level strategies include KV cache selection, budget allocation, merging, quantization, and low-rank decomposition, while model-level optimizations focus on architectural innovations and attention mechanisms to enhance KV reuse. System-level approaches address memory management, scheduling, and hardware-aware designs to improve efficiency across diverse computing environments. Additionally, the survey provides an overview of both text and multimodal datasets and benchmarks used to evaluate these strategies. By presenting detailed taxonomies and comparative analyses, this work aims to offer useful insights for researchers and practitioners to support the development of efficient and scalable KV cache management techniques, contributing to the practical deployment of LLMs in real-world applications. The curated paper list for KV cache management is in: https://github.com/TreeAI-Lab/Awesome-KV-Cache-Management{https://github.com/TreeAI-Lab/Awesome-KV-Cache-Management}.

Efficient Key-Value (KV) cache management is essential for processing long text sequences in large language models (LLMs), where memory constraints often limit performance. Conventional KV eviction strategies, such as top-k selection based on attention scores, depend on static heuristics that fail to capture the evolving implicit dependencies among tokens during inference. To overcome this, we propose GraphKV, a graph-based framework that redefines token selection for KV cache compression. In GraphKV, tokens are modeled as nodes with importance scores, and edges represent their similarity relationships. Through a decay-signal-propagation mechanism, token importance is dynamically updated by propagating information across the graph, enabling adaptive retention of the most contextually significant tokens. GraphKV can be seamlessly utilized in existing KV cache eviction methods such as SnapKV and PyramidKV in a plug-and-play manner. Codes will be released on Github.

Extensive LLM applications demand efficient structured generations, particularly for LR(1) grammars, to produce outputs in specified formats (e.g., JSON). Existing methods primarily parse LR(1) grammars into a pushdown automaton (PDA), leading to runtime execution overhead for context-dependent token processing, especially inefficient under large inference batches. To address these issues, we propose Pre^3 that exploits deterministic pushdown automata (DPDA) to optimize the constrained LLM decoding efficiency. First, by precomputing prefix-conditioned edges during the preprocessing, Pre^3 enables ahead-of-time edge analysis and thus makes parallel transition processing possible. Second, by leveraging the prefix-conditioned edges, Pre^3 introduces a novel approach that transforms LR(1) transition graphs into DPDA, eliminating the need for runtime path exploration and achieving edge transitions with minimal overhead. Pre^3 can be seamlessly integrated into standard LLM inference frameworks, reducing time per output token (TPOT) by up to 40% and increasing throughput by up to 36% in our experiments. Our code is available at https://github.com/ModelTC/lightllm.

Reinforcement learning (RL) has become a key driver of progress in large language models, but scaling RL to long chain-of-thought (CoT) trajectories is increasingly constrained by backpropagation over every generated token. Even with optimized rollout engines, full-token updates can consume a large fraction of total training cost, turning token length into a hidden tax on RL. We introduce Not All Tokens Are Needed (NAT), a unified framework that makes the token budget a first-class optimization primitive. NAT updates the policy using only a selected subset of generated tokens while preserving the learning signal of full-sequence RL. The core idea is an unbiased partial-token policy-gradient estimator via Horvitz-Thompson reweighting, which ensures statistically correct gradients despite subsampling. We instantiate NAT with two simple, plug-and-play token selection schemes: Uniform Random Sampling (URS) and Random Prefix Cutting (RPC), both of which reduce forward and backward compute and memory without modifying the reward computation or rollout pipeline. Across mathematical reasoning benchmarks, NAT matches full-token GRPO performance while using as few as 50% of tokens, providing an efficient and orthogonal pathway to scaling RL beyond the limits imposed by long trajectories. In our experiments, RPC saves 18% peak GPU memory and 29% forward and backward RL training time for Qwen3-8B.

Large Vision-Language Models (LVLMs) have advanced multimodal learning but face high computational costs due to the large number of visual tokens, motivating token pruning to improve inference efficiency. The key challenge lies in identifying which tokens are truly important. Most existing approaches rely on attention-based criteria to estimate token importance. However, they inherently suffer from certain limitations, such as positional bias. In this work, we explore a new perspective on token importance based on token transitions in LVLMs. We observe that the transition of token representations provides a meaningful signal of semantic information. Based on this insight, we propose TransPrune, a training-free and efficient token pruning method. Specifically, TransPrune progressively prunes tokens by assessing their importance through a combination of Token Transition Variation (TTV)-which measures changes in both the magnitude and direction of token representations-and Instruction-Guided Attention (IGA), which measures how strongly the instruction attends to image tokens via attention. Extensive experiments demonstrate that TransPrune achieves comparable multimodal performance to original LVLMs, such as LLaVA-v1.5 and LLaVA-Next, across eight benchmarks, while reducing inference TFLOPs by more than half. Moreover, TTV alone can serve as an effective criterion without relying on attention, achieving performance comparable to attention-based methods. The code will be made publicly available upon acceptance of the paper at https://github.com/liaolea/TransPrune.

Large Language Models (LLMs), despite their recent impressive accomplishments, are notably cost-prohibitive to deploy, particularly for applications involving long-content generation, such as dialogue systems and story writing. Often, a large amount of transient state information, referred to as the KV cache, is stored in GPU memory in addition to model parameters, scaling linearly with the sequence length and batch size. In this paper, we introduce a novel approach for implementing the KV cache which significantly reduces its memory footprint. Our approach is based on the noteworthy observation that a small portion of tokens contributes most of the value when computing attention scores. We call these tokens Heavy Hitters (H_2). Through a comprehensive investigation, we find that (i) the emergence of H_2 is natural and strongly correlates with the frequent co-occurrence of tokens in the text, and (ii) removing them results in significant performance degradation. Based on these insights, we propose Heavy Hitter Oracle (H_2O), a KV cache eviction policy that dynamically retains a balance of recent and H_2 tokens. We formulate the KV cache eviction as a dynamic submodular problem and prove (under mild assumptions) a theoretical guarantee for our novel eviction algorithm which could help guide future work. We validate the accuracy of our algorithm with OPT, LLaMA, and GPT-NeoX across a wide range of tasks. Our implementation of H_2O with 20% heavy hitters improves the throughput over three leading inference systems DeepSpeed Zero-Inference, Hugging Face Accelerate, and FlexGen by up to 29times, 29times, and 3times on OPT-6.7B and OPT-30B. With the same batch size, H2O can reduce the latency by up to 1.9times. The code is available at https://github.com/FMInference/H2O.

Tokenization is a fundamental step in natural language processing, breaking text into units that computational models can process. While learned subword tokenizers have become the de-facto standard, they present challenges such as large vocabularies, limited adaptability to new domains or languages, and sensitivity to spelling errors and variations. To overcome these limitations, we investigate a hierarchical architecture for autoregressive language modelling that combines character-level and word-level processing. It employs a lightweight character-level encoder to convert character sequences into word embeddings, which are then processed by a word-level backbone model and decoded back into characters via a compact character-level decoder. This method retains the sequence compression benefits of word-level tokenization without relying on a rigid, predefined vocabulary. We demonstrate, at scales up to 7 billion parameters, that hierarchical transformers match the downstream task performance of subword-tokenizer-based models while exhibiting significantly greater robustness to input perturbations. Additionally, during continued pretraining on an out-of-domain language, our model trains almost twice as fast, achieves superior performance on the target language, and retains more of its previously learned knowledge. Hierarchical transformers pave the way for NLP systems that are more robust, flexible, and generalizable across languages and domains.

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.