ICML 2026 论文列表

Incorporating code into training corpora has become a widely acknowledged practice in the development of modern foundation language models (LMs). Compared with a general Internet corpus, code offers high-quality, well-structured signals that substantially augment the coding proficiency of models. Beyond programming skills, prior research has suggested that code data may also contribute to non-coding capabilities. Nevertheless, through a series of rigorous controlled experiments, we demonstrate that the influence of code on other domains, particularly reasoning, remains limited. Our principal findings are as follows: (1) Code corpus yields substantial gains in programming-related abilities but only competed with knowledge-intensive tasks. (2) We identify a core subset that functions as cognitive scaffolding for mathematical reasoning, especially for complex problem-solving scenarios. (3) Formal reasoning provides more pronounced improvements in challenging mathematical reasoning tasks, while natural language–based reasoning proves more effective for simpler reasoning problems. Finally, by probing the internal mechanisms of LMs, we reveal how training data modulates routing patterns, thereby shaping emergent model behavior. As a central driver of model capability, our findings disentangle domain-specific data into finer-grained, cross-domain ability dimensions and underscore promising directions for future data optimization.

It is widely recognized that, after generative pre-training, Transformer FeedForward layers implicitly function as semantic memory, encoding linguistic and factual knowledge, while the contexts in key–value (KV) cache contain raw events, serving as the source of models' episodic memory. In this work, we show that a same group of Transformer FeedForward-layer parameters can both be semantic and episodic memory, which is retrievable without explicitly attending to the related KV cache. To realize this idea, we introduce Hypermem, a hypernetwork that recurrently maps contexts into targeted updates of FeedForward parameters. We post-train the hypernetwork using continuation and random-access associative memory objectives, eliminating the need for test-time training. Extensive experiments demonstrate that our approach outperforms related methods, including MemoryLLM and generative adapter, on memory retrieval, long-context question answering, and personalization benchmarks, establishing a new state of the art for hypernetwork-based memory mechanisms. Our results suggest that directly bridging data and parameters provides a viable direction for exploring next-generation foundation models with more flexible and persistent memory capabilities.

Although Multimodal Large Language Models (MLLMs) demonstrate strong omni-modal perception, their ability to forecast future events from audio-visual cues remains largely unexplored, as existing benchmarks focus mainly on retrospective understanding. To bridge this gap, we introduce FutureOmni, the first benchmark designed to evaluate omni-modal future forecasting from audio-visual environments. The evaluated models are required to perform cross-modal causal and temporal reasoning, as well as effectively leverage internal knowledge to predict future events. FutureOmni is constructed via a scalable LLM-assisted, human-in-the-loop pipeline and contains 919 videos and 1,034 multiple-choice QA pairs across 8 primary domains. Evaluations on 13 omni-modal and 7 video-only models show that current systems struggle with audio-visual future prediction, particularly in speech-heavy scenarios, with the best accuracy of 64.8% achieved by Gemini 3 Flash. To mitigate this limitation, we curate a 7K-sample instruction-tuning dataset and propose an Omni-Modal Future Forecasting (OFF) training strategy. Evaluations on FutureOmni along with standard audio-visual and video-only benchmarks show that OFF improves future forecasting performance and generalization. Code and data are available at https://github.com/OpenMOSS/FutureOmni.

Large language models (LLMs) increasingly perform multi-step reasoning, where intermediate claims form implicit directed acyclic graphs whose node correctness is structurally conditioned on their ancestors. This makes factuality uncertainty structural, rather than a trivial accumulation of node-wise errors, and necessitates inference-time uncertainty quantification over the reasoning structure. While conformal prediction (CP) offers flexible user-specified factuality control, existing work remains post-hoc and cannot intervene during generation. To fill the gap between CP’s flexibility and its post-hoc limitation, we propose an *Inference-Time Conformal Reasoning (ITCR)* framework that integrates CP directly into reasoning graph generation. ITCR learns a structure-level factuality uncertainty function that aggregates claim-level factuality signals over reasoning graphs without complex modeling assumptions. We then design the non-conformity score based on graph-level factuality uncertainty and calibrate the conformal threshold to decide when to stop generation. We theoretically show such generation is nested, yielding valid coverage guarantees for factuality control. Experiments over multiple datasets and coverage objectives demonstrate empirically valid coverage. In downstream reasoning tasks, inference-time calibrated graphs yield more accurate generation than post-hoc pruned graphs.

Simultaneous speech translation requires translating source speech into a target language in real-time while handling non-monotonic word dependencies. Traditional approaches rely on supervised training with word-level aligned data, which is difficult to collect at scale and thus depends on synthetic alignments using language-specific heuristics that are suboptimal. We propose Hibiki-Zero, which eliminates the need for word-level alignments entirely. This fundamentally simplifies the training pipeline and enables seamless scaling to diverse languages with varying grammatical structures, removing the bottleneck of designing language-specific alignment heuristics. We first train on sentence-level aligned data to learn speech translation at high latency, then apply a novel reinforcement learning strategy using GRPO to optimize latency while preserving translation quality. Hibiki-Zero achieves state-of-the-art performance in translation accuracy, latency, voice transfer, and naturalness across four X-to-English tasks. Moreover, we demonstrate that our model can be adapted to support a new input language with less than 1000h of speech. We provide [examples](https://huggingface.co/spaces/kyutai/hibiki-zero-samples), [model weights](https://huggingface.co/kyutai/hibiki-zero-3b-pytorch-bf16), [inference code](https://github.com/kyutai-labs/hibiki-zero) and we release a [benchmark](https://huggingface.co/datasets/kyutai/Audio-NTREX-4L) containing 45h of multilingual data for speech translation evaluation.

Training large language models via self-play often suffers from a persistent iteration-collapse, where performance initially improves but subsequently regresses as training iterations increase. We analyze this phenomenon as arising from cross-iteration degeneration, where the task-generation distribution becomes increasingly confined to a narrow subset of familiar (seen) problems, weakening the effective learning signal and destabilizing training. To address this issue, we propose a plug-in approach that augments existing self-play pipelines with a one-class novelty reward. A Seen Detector trained on a historical buffer of previously used training problems identifies in-support instances and discourages redundant generation by the questioner, thereby steering exploration toward under-explored yet learnable regions. Experimental results show that the proposed method mitigates iteration-collapse during iterative training and yields consistent improvements.

Large Language Models (LLMs) excel at code generation, yet ensuring the functional correctness of their outputs remains a persistent challenge. Recent studies have applied Test-Driven Development (TDD) to refine code, leveraging execution feedback to guide the model toward correct solutions. However, such feedback is often noisy and uninformative, stemming from the scarcity of high-quality test cases and the abundance of noisy, auto-generated ones. In this work, we shift the focus from test-case generation to feedback quality. We introduce the Property-Generated Solver (PGS), a novel feedback-centric framework designed to generate highly effective feedback via two principles: it must provide semantic guidance beyond simple I/O mismatches through property validation, and be structurally minimal, to reduce cognitive load and isolate root causes. PGS operates by checking high-level program properties (e.g., a sorting function must produce a non-decreasing sequence) then providing the simplest failing counterexample to the LLM. This property-driven, minimal feedback steers LLMs toward correct and generalizable solutions. Across diverse benchmarks, PGS demonstrates superior performance, achieving a bug fix rate 1.4x-1.6x higher than the strongest debugging-based approaches and establishing a new state-of-the-art in automated code refinement.

Harmful fine-tuning can invalidate safety alignment of large language models, exposing significant safety risks. In this paper, we utilize the attention sink mechanism to mitigate harmful fine-tuning. Specifically, we first measure a statistic named sink divergence for each attention head and observe that different attention heads exhibit two different signs of sink divergence. To understand its safety implications, we conduct experiments and find that the number of attention heads of positive sink divergence increases along with the increase of the model's harmfulness when undergoing harmful fine-tuning. Based on this finding, we propose a separable sink divergence hypothesis -- attention heads associating with learning harmful patterns during fine-tuning are separable by their sign of sink divergence. Based on the hypothesis, we propose a fine-tuning-stage defense, dubbed Surgery. Surgery utilizes a regularizer for sink divergence suppression, which steers attention heads toward the negative sink divergence group, thereby reducing the model’s tendency to learn and amplify harmful patterns. Extensive experiments demonstrate that Surgery improves defense performance by 5.90\%, 11.25\%, and 9.55\% on the BeaverTails, HarmBench, and SorryBench benchmarks, respectively. Source code is available on https://github.com/Lslland/Surgery.

Reliable reward models (RMs) are critical for ensuring the safe alignment of large language models (LLMs). However, current RM evaluation methods focus solely on preference perception accuracies in given specific scenarios, obscuring the critical vulnerabilities of RMs in real-world scenarios. We identify the true challenge lies in assessing a novel dimension: Suitability, defined as conditional reliability under specific real-world perturbations. To this end, we introduce Reward Auditor, a hypothesis-testing framework specifically designed for RM suitability inference. Rather than answering “How accurate is the RM's preference perception for given samples?”, it employs scientific auditing to answer: “Can we infer RMs exhibit systematic vulnerabilities in specific real-world scenarios?". Under real-world perturbed scenarios, Reward Auditor quantifies statistical significance and effect size by auditing distribution degradation of RM preference perception confidence. This enables inference of both the certainty and severity of RM vulnerabilities across diverse real-world scenarios, thereby laying a solid foundation for building next-generation LLM alignment systems that are verifiably safe, more robust, and trustworthy.

Federated parameter-efficient fine-tuning (PEFT) enables customizing large language models on private data, yet it is vulnerable to backdoor poisoning—especially when privacy constraints prevent inspection of per-client real-valued updates. We exploit the intuition that poisoning leaves a similar backdoor imprint in which adapter coordinates become salient, so overlap in salient-index supports remains informative even without values. We introduce IndexGuard, an unsupervised index-only vetting primitive in which clients send only Top-$K$ salient update indices and the server operates on the induced overlap geometry, clustering clients and filtering cohesion-outlier groups before aggregation. We analyze support stability under bounded rescaling and separability under shared-trigger poisoning under non-IID drift. Across attacks, backbones, and PEFT variants, IndexGuard provides end-to-end mitigation, preserving clean accuracy while achieving performance comparable to centralized methods.

While prompt engineering is instrumental in maximizing the capabilities of Large Language Models (LLMs) during inference, the role of prompts during training remains critically underexplored. Prevailing fine-tuning paradigms typically treat training prompts as mere surface forms, assuming that semantically equivalent instructions yield identical learning outcomes. However, we reveal that this equivalence is deceptive: while paraphrased prompts often lead to comparable in-task performance, they induce drastically different cross-task impacts regarding catastrophic forgetting and generalization. Crucially, these impacts are positively correlated across tasks, indicating the existence of superior prompts that consistently yield better performance. Furthermore, we discover that these superior prompts can be robustly identified by task loss prior to learning. Leveraging these insights, we introduce State-Adaptive Prompt Optimization (SAPO), a lightweight yet effective training strategy that shifts task formulation from a static input to a dynamic, state-adaptive variable. Comprehensive experiments on diverse benchmarks confirm its effectiveness, which significantly mitigates forgetting while improving generalization, achieving substantial performance gains over state-of-the-art methods. These results provide insights into how training prompts shape learning dynamics and offer a practical recipe for robust fine-tuning. Our code is available at https://github.com/Eric8932/SAPO.

Unified Foundation Models (UFMs), which support interleaved multimodal generation and understanding, have been proposed as a promising paradigm for reasoning about dynamic world states, yet it remains unclear whether the visual content they generate functions as grounded evidence for subsequent reasoning or merely as auxiliary output. Existing benchmarks largely evaluate generation and understanding as separate capabilities and do not test their functional dependence during reasoning. We introduce UFO, a benchmark designed to evaluate whether UFMs generate and use image and text cues as evidence for compositional multimodal reasoning. UFO spans three state-transition regimes, state determination, state reconstruction, and state augmentation, which correspond to progressively smaller transformations of the underlying world state. Our analysis reveals a significant modality gap, as models often achieve high prediction accuracy even when the generated visual cues exert limited influence on their decisions, indicating weakened evidential coupling and a reliance on textual shortcuts rather than robust cross-modal grounding.

Although reinforcement learning (RL) enhances the reasoning capabilities of large language models (LLMs), it is primarily learned from the model's self-generated distribution, limiting its ability to acquire reasoning skills beyond its initial knowledge. To overcome this, we propose a Difficulty-Aware Learning Strategy Allocation (DALSA) framework, which adaptively assigns appropriate learning strategies to samples based on their difficulty signals. DALSA is built on the key insight that samples beyond models' knowledge scope are better addressed through supervised fine-tuning (SFT), while those within the boundary but insufficiently mastered benefit more from RL, and well-learned samples are discarded to avoid redundant updates. To realize this principle, we extract a series of difficulty-aware training characteristics and employ a learnable strategy allocator to dynamically determine the optimal learning strategy for each sample based on its training dynamics. The allocator and the LLM are alternately optimized, enabling adaptive strategy allocation. Furthermore, two regularization techniques, anti-curriculum weighting and adversarial label smoothing, are integrated to alleviate the inherent limitations of RL and SFT, backed by comprehensive theoretical analyses. Extensive experiments on ten LLMs ranging from 1.5B to 70B across various tasks indicate that DALSA consistently outperforms baselines under both full and parameter-efficient fine-tuning settings.

Quantization-aware training (QAT) is an effective method to drastically reduce the memory footprint of LLMs while keeping performance degradation at an acceptable level. However, the optimal choice of quantization format and bit-width presents a challenge in practice. The full design space of quantization is not fully explored in the context of QAT, and the precise trade-off between quantization and downstream performance is poorly understood, as comparisons often rely solely on perplexity-based evaluations. In this work, we address these shortcomings with an empirical study of QAT in the low-bit regime. We show that k-means based weight quantization outperforms integer formats and can be implemented efficiently on standard hardware. Furthermore, we find that, under a fixed inference memory budget, the best performance on generative downstream tasks is achieved with $1$-bit quantized weights.

Tool-augmented reasoning has emerged as a promising direction for enhancing the reasoning capabilities of multimodal large language models (MLLMs). However, existing studies mainly focus on enabling models to perform tool invocation, while neglecting the necessity of invoking tools. We argue that tool usage is not always beneficial, as redundant or inappropriate invocations largely increase reasoning overhead and even mislead model predictions. To address this issue, we introduce AutoTool, a model that adaptively decides whether to invoke tools according to the characteristics of each query. Within a reinforcement learning framework, we design an explicit dual-mode reasoning strategy with mode-specific reward functions to guide the model toward producing accurate responses. Moreover, to prevent premature bias toward a single reasoning mode, AutoTool jointly explores and balances tool-assisted and text-centric reasoning throughout training, and promotes free exploration in later stages. Extensive experiments demonstrate that AutoTool exhibits outstanding performance and high efficiency, yielding a 21.8\% accuracy gain on V* benchmark compared to the base model, and a 44.9\% improvement in efficiency over existing tool-augmented methods on POPE benchmark. Code is available at https://github.com/MQinghe/AutoTool.

Structured pruning reduces LLM inference cost by removing low-importance architectural components. This can be viewed as learning a multiplicative gate for each component under an $\ell_0$ sparsity constraint. Due to the discreteness of the $\ell_0$ norm, prior work typically adopts stochastic hard-concrete relaxations to enable differentiable optimization; however, this stochasticity can introduce a train--test mismatch when sampled masks are discretized for deployment and restricting masks to a bounded, near-binary range. To address this, we propose Deterministic Differentiable Pruning (DDP), a mask-only optimization method that eliminates stochasticity by directly optimizing a deterministic soft surrogate of the discrete $\ell_0$ objective. Compared with prior approaches, DDP offers greater expressiveness, reduced train--test mismatch, and faster convergence. We apply our method to several dense and MoE models, including Qwen3-32B and Qwen3-30B-A3B, achieving a performance loss as small as 1\% on downstream tasks while outperforming previous methods at 20\% sparsity. We further demonstrate end-to-end inference speedups in realistic deployment settings with vLLM.

Accelerating scientific discovery requires the identification of which experiments would yield the best outcomes before committing resources to costly physical validation. While existing benchmarks evaluate LLMs on scientific knowledge and reasoning, their ability to predict experimental outcomes---a task where AI could significantly exceed human capabilities---remains largely underexplored. We introduce SciPredict, a benchmark comprising 405 tasks derived from recent empirical studies in 33 specialized sub-fields of physics, biology, and chemistry. SciPredict addresses two critical questions: (a) *can LLMs predict the outcome of scientific experiments with sufficient accuracy?* and (b) *can such predictions be reliably used in the scientific research process?* Evaluations reveal fundamental limitations on both fronts. Model accuracies are 14-26\% and human expert performance is $\approx$20\%. Although some frontier models exceed human performance model accuracy is still far below what would enable reliable experimental guidance. Even within the limited performance, models fail to distinguish reliable predictions from unreliable ones, achieving only $\approx$20\% accuracy regardless of their confidence or whether they judge outcomes as predictable without physical experimentation. Human experts, in contrast, demonstrate strong calibration: their accuracy increases from $\approx$5\% to $\approx$80\% as they deem outcomes more predictable without conducting the experiment. SciPredict establishes a rigorous framework demonstrating that superhuman performance in experimental science requires not just better predictions, but better awareness of prediction reliability. For reproducibility all our data and code are provided at https://anonymous.4open.science/r/SciPredict-AI01.

Multimodal large language models (MLLMs) are trained on massive multimodal data, making data unlearning increasingly important as data owners may request the removal of specific content. In practice, these requests often arrive sequentially over time, giving rise to the challenging problem of *MLLM Lifelong Unlearning*. However, most existing benchmarks are limited in scale and scope, failing to capture the complexities of MLLM lifelong unlearning. To fill this gap, we introduce the MLUBench, a large-scale and comprehensive benchmark featuring 127 entities across 9 classes under lifelong unlearning requests. We perform extensive experiments using MLUBench and reveal that existing unlearning methods suffer from severe, cumulative degradation. More critically, we further identify the unique challenge of this problem: unlike in unimodal models, MLLM lifelong unlearning is constrained by the need to preserve multimodal alignment. Continually unlearning from one modality could degrade the entire model. To alleviate this challenge, we propose LUMoE, an effective and efficient method. Experiments demonstrate that LUMoE significantly mitigates the degradation problem faced by baselines. The source code and the MLUBench dataset are open-sourced in this [URL](https://github.com/lihe-maxsize/Lifelong_Unlearning_main).

Large language models are increasingly deployed in streaming scenarios, rendering conventional post-hoc safeguards ineffective as they fail to interdict unsafe content in real-time. While streaming safeguards based on token-level supervised training could address this, they necessitate expensive annotations and suffer from severe overfitting. In this work, we challenge the paradigm that streaming safety must rely on token-level supervised training. Instead, it is an inherent capability of well-trained post-hoc safeguards, as they already encode token-level risk signals in hidden representations. Hence, we introduce NExT-Guard, a framework that achieves streaming safeguards by monitoring interpretable latent features from Sparse Autoencoders (SAEs). It uses pretrained SAEs from publicly available base LLMs, enabling flexible, low-cost deployment without token-level supervision. Experimental results show that NExT-Guard outperforms both post-hoc and streaming safeguards based on supervised training, with superior robustness across models, SAE variants, and risk scenarios. These results make NExT-Guard a universal and scalable paradigm for real-time safety, accelerating the practical deployment of streaming safeguards.