ICLR 2026 论文列表

Humans organize knowledge into compact categories that balance compression with semantic meaning preservation. Large Language Models (LLMs) demonstrate striking linguistic abilities, yet whether they achieve this same balance remains unclear. We apply the Information Bottleneck principle to quantitatively compare how LLMs and humans navigate this compression-meaning trade-off. Analyzing embeddings from 40+ LLMs against classic human categorization benchmarks, we uncover three key findings. First, LLMs broadly align with human categories but miss fine-grained semantic distinctions crucial for human understanding. Second, LLMs demonstrate aggressive statistical compression, achieving ``optimal'' information-theoretic efficiency, while humans prioritize contextual richness and adaptive flexibility. Third, encoder models surprisingly outperform decoder models in human alignment, suggesting that generation and understanding rely on distinct mechanisms in current architectures. In addition, training dynamics analysis reveals that conceptual structure develops in distinct phases: rapid initial formation followed by architectural reorganization, with semantic processing migrating from deeper to mid-network layers as models discover more efficient encoding. These divergent strategies, where LLMs optimize for compression and humans for adaptive utility, reveal fundamental differences between artificial and biological intelligence, guiding development toward more human-aligned AI.

Distilling robust reasoning capabilities from large language models (LLMs) into smaller, computationally efficient student models remains an unresolved challenge. Despite recent advances, distilled models frequently suffer from superficial pattern memorization and subpar generalization. To overcome these limitations, we introduce a novel distillation framework that moves beyond simple mimicry to instill a deeper conceptual understanding. Our framework features two key innovations. \underline{\textit{First}}, to address pattern memorization, Explanatory Inversion (EI) generates targeted ``explanatory probes'' that compel the student to articulate the underlying logic behind an answer, rather than just memorizing it. \underline{\textit{Second}}, to improve generalization, Explanatory GRPO (\texttt{EXGRPO}) uses a reinforcement learning algorithm with a novel Dialogue Structure Utility Bonus, which explicitly rewards the student for maintaining a coherent reasoning process across these probes. Extensive evaluations on 12 datasets demonstrate significant improvements. Using Gemma-7b as the student model, our method yields an average \textbf{20.39\%} increase over zero-shot performance and a \textbf{6.02\%} improvement over the state-of-the-art distillation baselines. Moreover, models distilled with our method show remarkable training efficiency (e.g., surpassing vanilla fine-tuning with \textbf{10-25\%} training data) and strong generalization to out-of-distribution tasks.

Vision-language models for medical image segmentation often produce masks that conflict with the accompanying text, especially under multi-site/multi-lesion descriptions. We trace this failure to two factors: (i) highly templated and repetitive clinical language causes one-to-one hard contrastive learning to yield numerous false negatives, weakening cross-modal alignment; and (ii) predominantly vision-driven, one-way cross-attention lacks a language-dominant, spatially aware pathway, hindering effective injection of textual semantics into the spatial visual domain. To this end, we propose Consistency-enhanced Two-stage Segmentation (C2Seg). In the pretraining stage, Cluster-aware Contrastive Learning uses a frozen strong baseline to construct an intra-batch text similarity matrix as soft labels, thereby alleviating false negative conflicts and producing more discriminative visual representations. In the fusion stage, we introduce a Bidirectional Complementary Attention Module, where each modality dominates attention along its own path, fostering deep interaction and structural consistency between visual and textual representations. In order to enhance the expressive power of multimodal features, we further adopt KAN-based Attention Gating. Without updating the language encoder, our approach significantly improves text-mask consistency and segmentation accuracy on four public medical imaging datasets. Code is provided in the supplementary material.

Pretraining large language models (LLMs) on high-quality, structured data such as mathematics and code substantially enhances reasoning capabilities. However, existing math-focused datasets built from Common Crawl suffer from degraded quality due to brittle extraction heuristics, lossy HTML-to-text conversion, and the failure to reliably preserve mathematical structure. In this work, we intro- duce Nemotron-CC-Math, a large-scale, high-quality mathematical corpus constructed from Common Crawl using a novel, domain-agnostic pipeline specifically designed for robust scientific text extraction. Unlike previous efforts, our pipeline recovers math across various formats (e.g., MathJax, KaTeX, MathML) by leveraging layout-aware rendering with lynx and a targeted LLM-based cleaning stage. This approach preserves the structural integrity of equations and code blocks while removing boilerplate, standardizing notation into L A T EX representation, and correcting inconsistencies. We collected a large, high-quality math corpus, namely Nemotron-CC-Math-3+(133B tokens) and Nemotron-CC-Math-4+ (52B tokens). Notably, Nemotron-CC-Math-4+ not only surpasses all prior open math datasets-including Mega-Math, FineMath, and OpenWebMath-but also contains 5.5× more tokens than FineMath-4+, which was previously the highest-quality math pretraining dataset. When used to pretrain a Nemotron-T 8B model, our corpus yields +4.8 to +12.6. gains on MATH and +4.6 to +14.3 gains on MBPP+ over strong baselines, while also improving general-domain performance on MMLU and MMLU-Stem. We present the first pipeline to reliably extract scientific content—including math—from noisy web-scale data, yielding measurable gains in math, code, and general reasoning, and setting a new state of the art among open math pretraining corpora. To support open-source efforts, we release our code1 and datasets 2 .

Parameter-efficient Mixture-of-Experts (MoE) architectures, such as LoRA-MoE, enable strong and generalizable fine-tuning. However, a critical problem arises when fine-tuning these architectures with advanced reinforcement learning algorithms such as Group Relative Policy Optimization (GRPO). Traditional supervised techniques are not naturally compatible with the GRPO objective, and naive combinations fail to effectively address routing collapse and the underutilization of MoE adapter parameters. To resolve this disconnect, we introduce Routing-Optimized Group Relative Policy Optimization (RO-GRPO), a mechanism-aware framework. It turns internal expert routing statistics collected during training into a direct reward signal, seamlessly integrating routing supervision into the reinforcement fine-tuning (RFT) process. This enables effective optimization of parameter utilization and improves performance on both unimodal and multimodal mathematical reasoning tasks, all without extra training stages. Our work provides the first demonstration that a scalar reward in GRPO can be engineered from a model's own internal mechanics to explicitly guide its optimization, extending alignment from mere behavior tuning to holistic mechanism alignment.

Long-term time series forecasting (LTSF) plays a crucial role in fields such as energy management, finance, and traffic prediction. Transformer-based models have adopted patch-based strategies to capture long-range dependencies, but accurately modeling shape similarities across patches and variables remains challenging due to scale differences. To address this, we introduce patch-mean decoupling (PMD), which separates the trend and residual shape information by subtracting the mean of each patch, preserving the original structure and ensuring that the attention mechanism captures true shape similarities. Futhermore, to more effectively model long-range dependencies and capture cross-variable relationships, we propose Trend Restoration Attention (TRA) and Proximal Variable Attention (PVA). The former module reintegrates the decoupled trend from PMD while calculating attention output. And the latter focuses cross-variable attention on the most relevant, recent time segments to avoid overfitting on outdated correlations. Combining these components, we propose PMDformer, a model designed to effectively capture shape similarity in long-term forecasting scenarios. Extensive experiments indicate that PMDformer outperforms existing state-of-the-art methods in stability and accuracy across multiple LTSF benchmarks.

Recent optimizers like Muon, Scion, and Gluon have pushed the frontier of large-scale deep learning by exploiting layer-wise linear minimization oracles (LMOs) over non-Euclidean norm balls, capturing neural network structure in ways traditional algorithms cannot. Yet, no principled distributed framework exists for these methods, and communication bottlenecks remain unaddressed. The very few distributed variants are heuristic, with no convergence guarantees in sight. We introduce EF21-Muon, the first communication-efficient, non-Euclidean LMO-based optimizer with rigorous convergence guarantees. EF21-Muon supports stochastic gradients, momentum, and bidirectional compression with error feedback–marking the first extension of error feedback beyond the Euclidean setting. It recovers Muon/Scion when compression is off and specific norms are chosen, providing the first efficient distributed implementation of this powerful family. Our theory covers non-Euclidean smooth and the more general (L0, L1)–smooth setting, matching best-known Euclidean rates and enabling faster convergence under suitable norm choices. We further extend the analysis to layer-wise (generalized) smoothness regimes, capturing the anisotropic structure of deep networks. Experiments on NanoGPT benchmarking EF21-Muon against uncompressed Muon/Scion/Gluon demonstrate up to 7× communication savings with no accuracy degradation.

Generative planners based on flow matching (FM) produce high-quality paths in a single or a few ODE steps, but their sampling dynamics offer no formal safety guarantees and can yield incomplete paths near constraints. We present SafeFlowMatcher, a planning framework that couples FM with control barrier functions (CBFs) to achieve both real-time efficiency and certified safety. SafeFlowMatcher uses a two-phase prediction-correction (PC) integrator: (i) a prediction phase integrates the learned FM once (or a few steps) to obtain a candidate path without intervention; (ii) a correction phase refines this path with a vanishing time‑scaled vector field and a CBF-based quadratic program that minimally perturbs the vector field. We prove a barrier certificate for the resulting flow system, establishing forward invariance of a robust safe set and finite-time convergence to the safe set. In addition, by enforcing safety only on the executed path—rather than all intermediate latent paths—SafeFlowMatcher avoids distributional drift and mitigates local trap problems. Moreover, SafeFlowMatcher attains faster, smoother, and safer paths than diffusion- and FM-based baselines on maze navigation, locomotion, and robot manipulation tasks. Extensive ablations corroborate the contributions of the PC integrator and the barrier certificate.

Large language models (LLMs) exhibit strong abilities in multi-document reasoning, yet their evidence identification is highly sensitive to input order. We trace this limitation to attention mechanisms, where many heads overemphasize sequence boundaries and neglect central content. We systematically analyze attention distributions under document permutations and discover a small subset of heads that consistently prioritize task-relevant documents regardless of position. We formalize these as Expert Heads, identified via activation frequency and stability across permutations. Experiments on LLaMA, Mistral, and Qwen reveal architecture-specific patterns: mid-layer heads in LLaMA and Mistral dominate semantic integration, while deeper-layer heads in Qwen specialize in evidence selection. Moreover, Expert Heads exhibit concentrated focus during understanding and more distributed engagement during generation. Their activation strongly correlates with answer correctness, providing diagnostic signals for hallucination detection. Leveraging Expert Heads for document voting significantly improves retrieval and ranking on HotpotQA, 2WikiMultiHopQA, and MuSiQue, outperforming dense retrievers and LLM-based ranking with minimal overhead. Ablations confirm that even a small subset achieves robust gains. Our findings establish Expert Heads as a stable and interpretable mechanism for evidence integration, offering new directions for context pruning, hallucination mitigation, and head-guided training of LLMs

Large language models (LLMs) struggle with complex logical reasoning. Previous work has primarily explored single-agent methods, with their performance remains fundamentally limited by the capabilities of a single model. To our knowledge, this paper first introduce a multi-agent approach specifically to enhance the logical reasoning abilities of LLMs. Considering the respective strengths and weaknesses of symbolic and natural language reasoning, we propose a multi-agent framework where individual agents reason in both symbolic and natural languages and then engage in a debate. To ensure the accuracy of symbolic translation, we also leverage multiple agents to translate and debate in different symbolic languages. Due to the prohibitive communication and token costs of multi-turn interactions, we further propose an adaptive sparse communication strategy to ensure efficiency. Specifically, our method prunes unnecessary communication by assessing the agent confidence and information gains, allowing each agent to selectively maintain its memory with others' most valuable outputs to help generate answers. Extensive experiments demonstrate that not only our multi-agent debate framework outperforms previous methods in logical reasoning tasks, but also our sparse communication approach outperforms the fully-connected communication with 25% token costs reduced, improving both effectiveness and efficiency.

Comparing human and model performance offers a valuable perspective for understanding the strengths and limitations of embedding models, highlighting wherethey succeed and where they fail to capture meaning and nuance. However, such comparisons are rarely made, as human performance on embedding tasks is difficult to measure. To fill this gap, we introduce HUME: Human Evaluation Framework for Text Embeddings. While frameworks like MTEB provide broad model evaluation, they lack reliable estimates of human performance, limiting the interpretability of model scores. We measure human performance across 16 MTEB datasets spanning reranking, classification, clustering, and semantic textual similarity across linguistically diverse high- and low-resource languages. Humans achieve an average performance of 77.6% compared to 80.1% for the best embedding model, though with substantial variation: models reach high performance on some datasets while struggling on notably low-resource languages. Our human annotation also reveals multiple notable dataset issues. We also benchmark nine LLMs as annotators, finding they fall short of human performance (76.1% vs. 81.2%) despite offering scalability advantages. We provide human performance baselines, insights into task difficulty patterns, and an extensible evaluation framework that enables a more meaningful interpretation of the model and informs the development of both models and benchmarks.

With advances in decision-making and reasoning capabilities, multimodal agents show strong potential in computer application scenarios. Past evaluations have mainly assessed GUI interaction skills, while tool invocation abilities, such as those enabled by the Model Context Protocol (MCP), have been largely overlooked. Comparing agents with integrated tool invocation to those evaluated only on GUI interaction is inherently unfair. We present OSWorld-MCP, the first comprehensive and fair benchmark for assessing computer-use agents' tool invocation, GUI operation, and decision-making abilities in a real-world environment. We design a novel automated code-generation pipeline to create tools and combine them with a curated selection from existing tools. Rigorous manual validation yields 158 high-quality tools (covering 7 common applications), each verified for correct functionality, practical applicability, and versatility. Extensive evaluations of state-of-the-art multimodal agents on OSWorld-MCP show that MCP tools generally improve task success rates (e.g., from 8.3\% to 20.4\% for OpenAI o3 at 15 steps, from 40.1\% to 43.3\% for Claude 4 Sonnet at 50 steps), underscoring the importance of assessing tool invocation capabilities. However, even the strongest models have relatively low tool invocation rates, Only 36.3\%, indicating room for improvement and highlighting the benchmark's challenge. By explicitly measuring MCP tool usage skills, OSWorld-MCP deepens understanding of multimodal agents and sets a new standard for evaluating performance in complex, tool-assisted environments. We will release all code and data to the community.

The rapid development of visual generative models raises the need for more scalable and human-aligned evaluation methods. While the crowdsourced Arena platforms offer human preference assessments by collecting human votes, they are costly and time-consuming, inherently limiting their scalability. Leveraging vision-language model (VLMs) as substitutes for manual judgments presents a promising solution. However, the inherent hallucinations and biases of VLMs hinder alignment with human preferences, thus compromising evaluation reliability. Additionally, the static evaluation approach lead to low efficiency. In this paper, we propose K-Sort Eval, a reliable and efficient VLM-based evaluation framework that integrates posterior correction and dynamic matching. Specifically, we curate a high-quality dataset from thousands of human votes in K-Sort Arena, with each instance containing the outputs and rankings of K models. When evaluating a new model, it undergoes (K+1)-wise free-for-all comparisons with existing models, and the VLM provide the rankings. To enhance alignment and reliability, we propose a posterior correction method, which adaptively corrects the posterior probability in Bayesian updating based on the consistency between the VLM prediction and human supervision. Moreover, we propose a dynamic matching strategy, which balances uncertainty and diversity to maximize the expected benefit of each comparison, thus ensuring more efficient evaluation. Extensive experiments show that K-Sort Eval delivers evaluation results consistent with K-Sort Arena, typically requiring fewer than 90 model runs, demonstrating both its efficiency and reliability. The dataset and code are publicly available.

Reinforcement Learning from Verifiable Rewards (RLVR) is a promising method for enhancing the complex problem-solving abilities of large language models (LLMs). This is particularly evident in domains requiring long-horizon reasoning and precise execution, such as solving complex mathematical problems where solutions hinge on a fragile sequence of tool-based actions. However, current approaches are often crippled by two interconnected issues: the near-miss problem, where sparse rewards nullify the learning signal for almost-correct attempts, and the resulting exploration stagnation, which prevents the model from discovering better solutions. To address these challenges, we introduce HiPO (Hint-guided Policy Optimization), a novel RLVR framework that enables the agent to learn from its own rare successes. Our core insight is to capture an occasional successful trajectory within a training batch and repurpose its initial correct steps as an on-policy “hint”. This process transforms a single, stochastically-found success into a dense contrastive learning signal, effectively allowing the model to teach itself how to overcome the near-miss problem and break exploration stagnation. On a challenging suite of five mathematical reasoning benchmarks, HiPO improves the average avg@32 by +5.0 percentage points (pp) over the strong GRPO baseline. This improvement is driven by substantial absolute point gains on challenging datasets, including +10.3 pp on CMIMC 2025, +4.9 pp on BRUMO 2025, +4.6 pp on AIME 2024, and +3.1 pp on AIME 2025. Furthermore, HiPO demonstrates a new exploration paradigm, repurposing rare successes into reusable guidance to significantly accelerate skill acquisition for complex tasks, establishing a more efficient and scalable path for models to autonomously master intricate reasoning.

We introduce a sequence-conditioned critic for Soft Actor--Critic (SAC) that models trajectory context with a lightweight Transformer and trains on aggregated $N$-step targets. Unlike prior approaches that (i) score state--action pairs in isolation or (ii) rely on actor-side action chunking to handle long horizons, our method strengthens the critic itself by conditioning on short trajectory segments and integrating multi-step returns---without importance sampling (IS). The resulting sequence-aware value estimates capture temporal structure critical for extended-horizon and sparse-reward problems. On local-motion benchmarks, we further show that freezing critic parameters for several steps makes our update compatible with CrossQ's core idea, enabling stable training without a target network. Despite its simplicity---a 2-layer Transformer with 128--256 hidden units and a maximum update-to-data ratio (UTD) of $1$---the approach consistently outperforms standard SAC and strong off-policy baselines, with particularly large gains on long-trajectory control. These results highlight the value of sequence modeling and $N$-step bootstrapping on the critic side for long-horizon reinforcement learning.

Recurrent neural networks (RNNs) with deep test-time memorization modules, such as Titans and TTT, represent a promising, linearly-scaling paradigm distinct from Transformers. While these expressive models do not yet match the peak performance of state-of-the-art Transformers, their potential has been largely untapped due to prohibitively slow training and low hardware utilization. Existing parallelization methods force a fundamental conflict governed by the chunksize hyperparameter: large chunks boost speed but degrade performance, necessitating a fixed, suboptimal compromise. To solve this challenge, we introduce TNT, a novel training paradigm that decouples training efficiency from inference performance through a two-stage process. Stage one is an efficiency-focused pre-training phase utilizing a hierarchical memory. A global module processes large, hardware-friendly chunks for long-range context, while multiple parallel local modules handle fine-grained details. Crucially, by periodically resetting local memory states, we break sequential dependencies to enable massive context parallelization. Stage two is a brief fine-tuning phase where only the local memory modules are adapted to a smaller, high-resolution chunksize, maximizing accuracy with minimal overhead. Evaluated on Titans and TTT models, TNT achieves a substantial acceleration in training speed—up to 17$\times$ faster than the most accurate baseline configuration—while simultaneously improving model accuracy. This improvement removes a critical scalability barrier, establishing a practical foundation for developing expressive RNNs and facilitating future work to close the performance gap with Transformers.

Masked diffusion models (MDMs) have achieved notable progress in modeling discrete data, while their potential in molecular generation remains underexplored. In this work, we explore their potential and introduce the surprising result that naively applying standards MDMs to molecules leads to severe performance degradation. We trace this critical issue to a *state-clashing*-where the forward diffusion trajectories of distinct molecules collapse into a common state, resulting in a mixture of reconstruction targets that cannot be learned with a typical reverse diffusion with unimodal predictions. To mitigate this, we propose **M**asked **E**lement-wise **L**earnable **D**iffusion (**MELD**) that orchestrates per-element corruption trajectories to avoid collisions between different molecular graphs. This is realized through a parameterized noise scheduling network that learns distinct corruption rates for individual graph elements, *i.e.*, atoms and bonds. Across extensive experiments, **MELD** is the first diffusion-based molecular generator to achieve 100% chemical validity in unconditional generation on QM9 and ZINC250K datasets, while markedly improving distributional and property alignment over standard MDMs.

Neural solvers have achieved impressive progress in addressing simple routing problems, particularly excelling in computational efficiency. However, their advantages under complex constraints remain nascent, for which current constraint-handling schemes via feasibility masking or implicit feasibility awareness can be inefficient or inapplicable for hard constraints. In this paper, we present Construct-and-Refine (CaR), the first general and efficient constraint-handling framework for neural routing solvers based on explicit learning-based feasibility refinement. Unlike prior construction-search hybrids that target reducing optimality gaps through heavy improvements yet still struggle with hard constraints, CaR achieves efficient constraint handling by designing a joint training framework that guides the construction module to generate diverse and high-quality solutions well-suited for a lightweight improvement process, e.g., 10 steps versus 5k steps in prior work. Moreover, CaR presents the first use of construction-improvement-shared representation, enabling potential knowledge sharing across paradigms by unifying the encoder, especially in more complex constrained scenarios. We evaluate CaR on typical hard routing constraints to showcase its broader applicability. Results demonstrate that CaR achieves superior feasibility, solution quality, and efficiency compared to both classical and neural state-of-the-art solvers. Our code, pre-trained models, and datasets are available at: https://github.com/jieyibi/CaR-constraint.

High-stakes, infrequent consumer decisions, such as housing selection, challenge conventional recommender systems due to sparse interaction signals, heterogeneous multi-criteria objectives, and high-dimensional feature spaces. This work presents an interactive preference elicitation framework that couples preferential Bayesian optimization (PBO) with two complementary components: (i) large language models (LLMs) that interpret natural language input to produce personalized probabilistic priors over feature utility weights to mitigate cold start, and (ii) an autoencoder (AE)-based latent representation that reduces effective dimensionality for sample-efficient exploration. The framework learns a latent utility function from user pairwise comparisons observed and integrated in real-time. We evaluate the developed method on rental real estate datasets from two major European cities. The results show that executing PBO in an AE latent space improves final pairwise ranking accuracy by 12%. For LLM-based preference prior generation, we find that direct, LLM-driven weight specification is outperformed by a static prior, while probabilistic weight priors that use LLMs only to rank feature importance achieve 25% better pairwise accuracy on average than a direct approach.