ICML 2026 论文列表

Large language models (LLMs) incur substantial inference latency due to autoregressive decoding, in which each token requires a full forward pass through all transformer layers. Early-exit methods that terminate computation at intermediate layers offer a promising remedy, yet existing approaches suffer from fundamental limitations. Confidence-based methods rely on evaluating the full LM head at every layer, introducing considerable overhead that can negate the expected speedup. Schedule-based methods avoid this cost through predetermined exit schedules, but their monotonically decreasing layer allocation collapses to shallow layers, thereby constraining the maximum generation length. Learned exit predictors further require costly task-specific training and are vulnerable to distribution shifts in unseen domains. We propose Window-Aware Vocabulary-Efficient Early-Exit (WAVE), a training-free framework that addresses these challenges through two key innovations. First, exit window scheduling identifies an optimal layer range for early-exit decisions via offline calibration, preventing premature convergence to shallow layers while substantially reducing the number of exit checks. Second, a proxy LM head constructs a lightweight vocabulary subset at the window’s starting layer, reducing per-layer exit overhead by 87\% relative to full LM head. WAVE requires no gradient-based training and enables immediate deployment with only a brief calibration phase. Experiments on Llama-2 7B demonstrate up to 1.4$\times$ average speedup while preserving output quality, with full compatibility with W4A16 quantization, establishing WAVE as a practical early-exit framework for accelerating LLMs inference without retraining.

Long-form TV dramas present a formidable challenge for comprehensive video understanding, where deciphering complex storyline often relies on **speaker recognition**, the task of accurately attributing each spoken utterance to its respective character. In this paper, we advance this field through two primary contributions. (1) We introduce **DramaSR-532K**, a large-scale benchmark comprising 532K annotated dialogue lines across more than 900 unique characters, necessitating the integration of auditory, linguistic, and visual cues for speaker recognition. (2) We propose **DramaSR-LRM**, a robust approach built upon a large reasoning model (LRM). DramaSR-LRM is designed to autonomously aggregate contextual evidence via multimodal tool-use, synthesizing diverse inputs to achieve high-fidelity attribution. Experimental results demonstrate that DramaSR-LRM significantly outperforms existing baselines, particularly on short utterances where acoustic biometrics are inherently unreliable. *All the data and code will be made publicly available.*

Grammar-constrained decoding enables large language models (LLMs) to reliably generate structured outputs such as JSON, SQL, and domain-specific programs. Existing systems often enforce constraints by executing byte-level parser logic inside the token-level decoding loop, introducing CPU-side control flow and CPU--GPU synchronization that become bottlenecks under continuous batching. We propose Gram2Token, a GPU-native framework that preprocesses deterministic byte-level grammar execution into token-level transitions before inference. Gram2Token aligns tokenizer byte sequences with grammar transitions through a trie and groups tokens with identical transition outcomes across preprocessed grammar states. These categories yield compact validity masks and transition tables, reducing run-time enforcement to category lookup, masking, and state update rather than parser-style byte traversal. Across four model families under schema-diverse continuous batching, Gram2Token achieves a geometric-mean throughput improvement of 1.38$\times$ over the strongest baseline, with a maximum speedup of 1.85$\times$, at the cost of additional preprocessing and time-to-first-token overhead. Break-even and grammar-complexity analyses show that this overhead is amortized by grammar reuse, longer outputs, and larger batches. These results show that token-level grammar preprocessing is an effective design point for high-throughput structured LLM serving. Code is available at https://github.com/Paradozile/Gram2Token.

Autoregressive Models (ARMs) have long dominated the landscape of Large Language Models. Recently, a new paradigm has emerged in the form of diffusion-based Large Language Models (dLLMs), which generate text by iteratively denoising masked segments. This approach has shown significant advantages and potential. However, dLLMs suffer from high inference latency. Traditional ARM acceleration techniques, such as Key-Value caching, are incompatible with dLLMs due to their bidirectional attention mechanism. To address this specific challenge, our work begins with a key observation that dLLM inference involves a static prompt and a partially dynamic response, where most tokens remain stable across adjacent denoising steps. Based on this, we propose dLLM-Cache, a training-free adaptive caching framework that combines long-interval prompt caching with partial response updates guided by feature similarity. This design enables efficient reuse of intermediate computations without compromising model performance. Extensive experiments on representative dLLMs, including LLaDA 8B and Dream 7B, show that dLLM-Cache achieves up to *9.1*$\times$ FLOPs reduction on LongBench-HotpotQA while maintaining competitive output quality. Notably, our method brings dLLM inference latency close to that of ARMs under many settings. *The code for this work is publicly available at: https://github.com/maomaocun/dLLM-cache*.

Diffusion-based large language models generate text by gradually filling in masked tokens, yet they remain slow because they usually decode only a few tokens per step. Parallel decoding, which unmasks multiple tokens simultaneously, promises acceleration but often degrades quality when too many tokens are predicted at once. We identify the root cause: when decoding is viewed as iterative token grouping, overly permissive grouping places interdependent tokens in the same step, violates the conditional independence assumption, and amplifies reliance on noisy context even when the top prediction is already correct. We introduce SPEED, a framework that enlarges safe parallel groups through complementary training and inference designs. At training time, a sharpened teacher distillation objective selectively aligns the student to teacher-correct positions using a temperature-scaled KL term together with a masked language modeling loss, producing a student that assigns more probability mass to correct token identities and elevates more positions above the decoding threshold. At inference time, Slow-Fast Decoding partitions tokens by sensitivity to revealed context using token-wise Jensen-Shannon Divergence computed with and without access to the preceding block, decoding low-sensitivity tokens jointly in parallel while deferring high-sensitivity tokens until sufficient context resolves them. Through extensive experiments, our framework attains up to 12.2$\times$ speedup on LLaDA-8B-Instruct and 6.7$\times$ on Dream-7B-Instruct with accuracy close to greedy decoding across standard reasoning and code benchmarks.

Large Language Models (LLMs) show strong reasoning abilities, yet their reliability is hindered by hallucinations, where fluent reasoning becomes factually or logically incorrect. Most existing uncertainty-based detectors rely on sequence-level averaging, which ignores the step-wise dynamics of reasoning and often misclassifies hard-but-correct or easy-but-wrong samples. We propose a dynamic perspective that models reasoning as a trajectory on a latent Evidence Manifold, where each step is supported by local evidence. Hallucinations are characterized as Evidence Drops, i.e., sudden declines in local evidence support that indicate topological deviations from this manifold. Based on this insight, we design a training-free and model-agnostic detector that identifies hallucinations via the worst-case Evidence Drop and enables step-level error localization. Experiments on GSM8K, MATH, and ProcessBench show consistent improvements over sequence-level uncertainty baselines in selective accuracy and risk–coverage trade-offs.

Reward models play a fundamental role in aligning large language models with human preferences. Existing methods predominantly follow two paradigms: scalar discriminative preference models, which are efficient but lack interpretability, and generative judging models, which offer richer reasoning at the cost of higher computational overhead. We observe that the log-probability margin between verdict tokens strongly correlates with prediction correctness, providing a reliable proxy for instance difficulty without additional inference cost. Building on this insight, we propose CAMEL, a confidence-gated reflection framework that performs a lightweight single-token preference decision first and selectively invokes reflection only for low-confidence instances. To induce effective self-correction, we train the model via reinforcement learning with counterfactual prefix augmentation, which exposes the model to diverse initial verdicts and encourages genuine revision. Empirically, CAMEL achieves state-of-the-art performance on three widely used reward-model benchmarks with 82.9% average accuracy, surpassing the best prior model by 3.2% and outperforming 70B-parameter models using only 14B parameters, while establishing a strictly better accuracy-efficiency Pareto frontier.

Mid-training is increasingly used to improve the reasoning capabilities of large language models (LLMs), yet its design choices and interaction with evaluation and reinforcement learning (RL) remain poorly understood. Prior work often focuses on narrow domain gains, overlooking retention of general abilities, long-context performance, and RL compatibility. We present $\textbf{PRISM}$ (Demystifying Retention and Interaction in Mid-Training), a holistic empirical study that analyzes mid-training design choices, what to evaluate, and how domain mixtures and training stages interact across model families. Experiments on Granite-3.3 8B, LLaMA-3.1 8B, and Mistral-7B/24B base models show that a relatively small, high-quality mid-training phase of $\textbf{$\sim$27B}$ tokens acts as a critical stabilizing stage for reasoning. Across models, PRISM yields consistent gains of $\textbf{$\sim$6–10}$ points on coding benchmarks and $\textbf{$\sim$17–30}$ points on mathematical reasoning benchmarks while preserving general performance. RL applied on top of PRISM-mid-trained models produces stable, monotonic improvements, adding a further $\textbf{$\sim$3–8}$ points across coding and math tasks such as LiveCodeBench, Codeforces, AIME and MATH500, and $\textbf{$\sim$17–20}$ points on science (GPQA-Diamond), whereas RL applied directly to base models is substantially less effective. Our results demonstrate that retention-aware mid-training is a necessary intermediate step for reliable reasoning enhancement and RL scaling, and provide practical guidance for designing robust mid-training pipelines for modern LLMs.

Large language models (LLMs) are increasingly used as scalable judges in pairwise evaluation, but they remain prone to miscalibration and biases. We propose SCOPE (Selective Conformal Optimized Pairwise Evaluation), a framework that calibrates an acceptance threshold so that, under exchangeability, the error rate among non-abstained judgments is at most a user-specified level $\alpha$. To supply SCOPE with a bias-neutral uncertainty signal, we introduce Bidirectional Preference Entropy (BPE), which queries the judge under both response positions and converts the order-averaged preference probability into an entropy-based score. Across various pairwise judging benchmarks, BPE outperforms standard confidence proxies in calibration and discrimination, while SCOPE consistently satisfies the target risk bound (empirical FDR $\approx 0.097$--$0.099$ at $\alpha=0.10$) and retains substantial coverage. Compared to vanilla baselines, SCOPE accepts up to $2.4\times$ more judgments under the same risk constraint, demonstrating that BPE enables reliable and high-coverage LLM-based evaluation.

Large reasoning models (LRMs) often generate long, seemingly coherent reasoning traces yet still produce incorrect answers, making hallucination detection challenging. Although trajectories contain useful signals, directly using trace text or vanilla hidden states for detection is brittle: traces vary in form and detectors can overfit to superficial patterns rather than answer validity. We introduce Answer-agreement Representation Shaping (ARS), which learns detection-friendly trace-conditioned representations by explicitly encoding answer stability. ARS generates counterfactual answers through small latent interventions, specifically, perturbing the trace-boundary embedding, and labels each perturbation by whether the resulting answer agrees with the original. It then learns representations that bring answer-agreeing states together and separate answer-disagreeing ones, exposing latent instability indicative of hallucination risk. The shaped embeddings are plug-and-play with existing embedding-based detectors and require no human annotations during training. Experiments demonstrate that ARS consistently improves detection and achieves substantial gains over strong baselines.

Generative Graph Language Models (GLMs) must reconcile topology with causal language modeling. Linearization obscures multi-hop connectivity, while encoder-based methods bottleneck token-level reasoning during generation. Viewing context modeling as a form of message passing, we introduce **Weaver**, an encoder-free framework that extends the attention mechanism of decoder-only LLMs to enable graph reasoning. Weaver maps graph distances into rotary positional embeddings so that structurally connected nodes become proximate in attention space, propagating information over graph topology as if it were sequential context. To achieve this, we combine: 1) a masking mechanism for causal tokens with graph structures; 2) a unified geometric encoding that couples sequential position and graph distance in joint rotary embeddings (Graph-over-Tokens RoPE); and 3) a design principle to prioritize local information to resolve positional ambiguity under graph symmetries. On zero-shot benchmarks, Weaver achieves state-of-the-art performance among generative GLMs, with gains of up to 30\% over prior generative methods on heterophilic graphs, while matching specialized discriminative models on citation networks---all within a unified decoder-only framework.

Recent large language models (LLMs) have shown rapid progress on reading-based question answering (QA), where the evidence is explicitly provided or trivially retrievable. In contrast, real-world questions are often not paired with accurate evidence documents. The useful evidence resides in a massive collection of data lakes, necessitating searching as a prerequisite for answering. However, there is a lack of a comprehensive benchmark that requires searching and reasoning over a large collection of data lakes. To this end, we introduce LakeQA, a comprehensive benchmark for search-centric question answering over data lakes that jointly emphasizes \emph{searching} and \emph{reasoning} capabilities. LakeQA is built on a heterogeneous collection of ~9.5 TB text resources from Wikipedia and open-source government data, spanning structured and unstructured data. To ensure the quality of LakeQA's tasks, each sample is annotated by at least one Ph.D level expert. Each task requires long-horizon multi-hop reasoning with implicit intermediate steps: agents need to discover the correct document(s) and then compose evidence across sources to produce the answer. Intensive experiment results on seven frontier LLMs have demonstrated that LakeQA is challenging. For instance, GPT-5.2 only obtains an exact matching score of 14.73% on LakeQA. Overall LakeQA provides a realistic testbed for developing LLM agents that can both \emph{find} and \emph{analyze} data in modern data lakes.

A central challenge in large language model (LLM) editing is capability preservation: methods that successfully change targeted behavior can quietly game the editing proxy and corrupt general capabilities, producing degenerate behaviors reminiscent of proxy/reward hacking. We present CrispEdit, a scalable and principled second-order editing algorithm that treats capability preservation as an explicit constraint, unifying and generalizing several existing editing approaches. CrispEdit formulates editing as constrained optimization and enforces the constraint by projecting edit updates onto the low-curvature subspace of the capability-loss landscape. At the crux of CrispEdit is expressing capability constraint via Bregman divergence, whose quadratic form yields the Gauss–Newton Hessian exactly and even when the base model is not trained to convergence. We make this second-order procedure efficient at the LLM scale using Kronecker-factored approximate curvature (K-FAC) and a novel matrix-free projector that exploits Kronecker structure to avoid constructing massive projection matrices. Across standard model-editing benchmarks and safety unlearning tasks, CrispEdit achieves high edit success while keeping capability degradation below 1% on average across datasets, significantly improving over prior editors.

Post-training of reasoning LLMs is a holistic process that typically consists of an offline SFT stage followed by an online reinforcement learning (RL) stage. However, SFT is often optimized in isolation to maximize SFT performance alone. We show that, after identical RL training, models initialized from stronger SFT checkpoints can significantly underperform those initialized from weaker ones. We propose PEAR ($\textbf{P}$olicy $\textbf{E}$valuation–inspired $\textbf{A}$lgorithm for Offline Learning Loss $\textbf{R}$eweighting), an SFT-stage method that corrects this mismatch and better prepares the model for RL. PEAR uses importance sampling to reweight the SFT loss, with three variants operating at the token, block, and sequence levels. It can be used to augment standard SFT objectives and incurs little additional training overhead once probabilities for the offline data are collected. We conduct controlled experiments on verifiable reasoning games and mathematical reasoning tasks on Qwen2.5/3 and DeepSeek-distilled models. PEAR consistently improves post-RL performance over canonical SFT, with pass@8 gains up to a 14.6% on AIME-2025. Our results suggest that PEAR is an effective step toward more holistic LLM post-training by designing and evaluating SFT with downstream RL in mind rather than in isolation.

Large language model (LLM) agents are increasingly deployed in long-running settings where improving through experience at test time becomes important. A common approach is to update an explicit memory after each interaction to guide future decisions. However, most existing methods rely on hand-designed prompting rules, making it difficult to align memory updates with downstream objectives over multi-step horizons consistently. We propose MemoPilot, a plug-in memory copilot that explicitly trains the memory update process to improve a frozen LLM's performance across sequential interactions. We formulate memory updating as a multi-turn decision problem and optimize it end-to-end with multi-turn GRPO. Our training recipe introduces (i) a turn-wise reward signal and (ii) a context-independent, turn-level advantage estimation across rollouts, enabling finer-grained credit assignment and more stable training in multi-turn settings. We evaluate MemoPilot on two testbeds: multi-round Rock-Paper-Scissors (RPS) and Limit Texas Hold'em (LHE). Across both environments, MemoPilot substantially improves test-time learning of a frozen player over strong baselines, ranking first in Elo ratings on both games (1762 on LHE and 1590 on RPS) and outperforming all baseline memory methods and proprietary models, including Deepseek-V3.2. Our code is publicly available.

The core learning signal used in language model distillation is the standard Kullback-Leibler (KL) divergence between the student and teacher distributions. Traditional KL divergence tends to be dominated by the teacher’s highest-probability modes, thereby diminishing the influence of less-probable yet potentially informative components of the output distribution. We propose a new tail-aware divergence that decouples the contribution of the teacher model's top-$K$ predicted probabilities from that of lower-probability predictions, while maintaining the same computational profile as the KL Divergence. Our decoupled approach reduces the impact of teacher modes and, consequently, increases the contribution of the distribution's tail. Experimental results demonstrate that our modified distillation method yields competitive performance in both pre-training and supervised distillation for mathematical reasoning of decoder models across various datasets. Furthermore, the distillation process is efficient and can be performed on modest academic budgets for large datasets, drastically reducing the computational costs typically associated with large-scale distillation.

Training data plays a central role in large language models (LLMs) optimization, motivating extensive research on data scheduling strategies. Most existing approaches concentrate on adjusting the overall data distribution but neglect the underlying interactions between samples during training. However, we argue that such interactions cannot be overlooked, as real-world data samples frequently exhibit directional influences on each other, making the training order crucial. Intuitively, we can prioritize train-units with greater influence to improves learning efficiency. In this work, we propose $D^3$, a **D**ynamic, **D**irectional graph-constrained **D**ata scheduling framework. $D^3$ formulates the complex interactions among train-units as a dynamic influence graph, where edges represent loss-based dependencies. It then solves a constrained optimization problem over this graph to derive the training order, which ensures that the data sequence respects the evolving information flow throughout training. Our approach is theoretically motivated and yields consistent improvements over existing data scheduling methods across both pre-training and post-training phases. Furthermore, for scalability, $D^3$ also employs an efficient approximation algorithm that keeps the additional computational overhead within a manageable range. For future research, the code is available at https://github.com/xuyj233/D3.

Multimodal Large Language Models (MLLMs) demonstrate impressive cross-modal capabilities, yet their substantial size poses significant deployment challenges. Knowledge distillation (KD) is a promising solution for compressing these models, but existing methods primarily rely on static next-token alignment, neglecting the dynamic token interactions, which embed essential capabilities for multimodal understanding and generation. To this end, we introduce **Align-TI**, a novel KD framework designed from the perspective of **T**oken **I**nteractions. Our approach is motivated by the insight that MLLMs rely on two primary interactions: vision-instruction token interactions to extract relevant visual information, and intra-response token interactions for coherent generation. Accordingly, Align-TI introduces two components: IVA enables the student model to imitate the teacher's instruction-relevant visual information extract capability by aligning on salient visual regions. TPA captures the teacher's dynamic generative logic by aligning the sequential token-to-token transition probabilities. Extensive experiments demonstrate Align-TI's superiority. Notably, our approach achieves 2.6% relative improvement over Vanilla KD, and our distilled Align-TI-2B even outperforms LLaVA-1.5-7B (a much larger MLLM) by 7.0%, establishing a new state-of-the-art distillation framework for training parameter-efficient MLLMs.

Generating accurate and calibrated confidence estimates is critical for deploying LLMs in high-stakes or user-facing applications, and remains an open challenge. Prior research has often framed confidence as a problem of eliciting a model’s “self-knowledge”, i.e., the ability of an LLM to judge whether its own answers are correct; this approach implicitly assumes that there is some privileged information about the answer’s correctness that is accessible to the model itself. However, we find that whether trained or training-free, an LLM attempting to predict the correctness of its own outputs generally performs no better than an unrelated LLM attempting the same task. Moreover, we hypothesize that a key factor in predicting model correctness, i.e., building a “Correctness Model” (CM), is exposure to a target model’s historical predictions. We use multiple methods to inject this historical correctness information, including training an LLM to predict the confidences of many other LLMs, i.e., creating a Generalized Correctness Model (GCM). We use GCMs and CMs as a lens for studying the source of correctness prediction ability and its generalization, studying the importance of answer phrasing, world-knowledge, performance history, in-context examples, and post-hoc calibration for correctness prediction. We evaluate GCMs based on Qwen3-8B across 5 model families and the MMLU, TriviaQA, and Spider datasets, as well as on a downstream selective prediction task, finding that reliable LLM confidence estimation is a cross-model skill learned by encoding correctness history rather than a model-specific skill reliant on introspection. Code: https://github.com/The-Inscrutable-X/CalibratedModelAgnosticCorrectness.