ICLR 2026 Papers

While Vision–language models (VLMs) have demonstrated remarkable performance across multi-modal tasks, their choice of vision encoders presents a fundamental weakness: their low-level features lack the robust structural and spatial information essential for document understanding and web agents. To bridge this gap, we introduce DAVE, a vision encoder purpose-built for VLMs and tailored for these tasks. Our training pipeline is designed to leverage abundant unlabeled data to bypass the need for costly large-scale annotations for document and web images. We begin with a self-supervised pretraining stage on unlabeled images, followed by a supervised autoregressive pretraining stage, where the model learns tasks like parsing and localization from limited, high-quality data. Within the supervised stage, we adopt two strategies to improve our encoder’s alignment with both general visual knowledge and diverse document and web agentic tasks: (i) We introduce a novel model-merging scheme, combining encoders trained with different text decoders to ensure broad compatibility with different web agentic architectures. (ii) We use ensemble training to fuse features from pretrained generalist encoders (e.g., SigLIP2) with our own document and web-specific representations. Extensive experiments on classic document tasks, VQAs, web localization, and agent-based benchmarks validate the effectiveness of our approach, establishing DAVE as a strong vision encoder for document and web applications.

AI agents powered by large language models (LLMs) are being deployed at scale, yet we lack a systematic understanding of how the choice of backbone LLM affects agent security. The non-deterministic sequential nature of AI agents complicates security modeling, while the integration of traditional software with AI components entangles novel LLM vulnerabilities with conventional security risks. Existing frameworks only partially address these challenges as they either capture specific vulnerabilities only or require modeling of complete agents. To address these limitations, we introduce threat snapshots: a framework that isolates specific states in an agent's execution flow where LLM vulnerabilities manifest, enabling the systematic identification and categorization of security risks that propagate from the LLM to the agent level. We apply this framework to construct the $b^3$ benchmark, a security benchmark based on 194,331 unique crowdsourced adversarial attacks. We then evaluate 34 popular LLMs with it, revealing, among other insights, that enhanced reasoning capabilities improve security, while model size does not correlate with security. We release our benchmark, dataset, and evaluation code to facilitate widespread adoption by LLM providers and practitioners, offering guidance for agent developers and incentivizing model developers to prioritize backbone security improvements.

Ensembling Large Language Models (LLMs) has gained attention as a promising approach to surpass the performance of individual models by leveraging their complementary strengths. In particular, aggregating models’ next-token probability distributions to select the next token has been shown to be effective in various tasks. However, while successful for short-form answers, its application to long-form generation remains underexplored. In this paper, we show that using existing ensemble methods in long-form generation requires a careful choice of ensembling positions, since the standard practice of ensembling at every token often degrades performance. We identify two key factors for determining the ensembling positions: tokenization mismatch across models and consensus in their next-token probability distributions. Based on this, we propose $\textbf{SAFE}$, ($\textbf{S}$table $\textbf{A}$nd $\textbf{F}$ast LLM $\textbf{E}$nsembling), a framework that selectively ensembles by jointly considering these factors. To further improve stability, we apply a probability sharpening strategy when the ensemble distribution becomes overly smooth, enabling the selection of more confident tokens during ensembling. Our experiments on diverse benchmarks, including MATH500 and BBH, demonstrate that SAFE outperforms existing methods in both accuracy and efficiency, with gains achieved even when ensembling fewer than 1\% of tokens.

Modern Large Language Models (LLMs) face fundamental challenges in processing long text sequences due to the quadratic complexity of attention mechanisms. Key-Value (KV) cache retention strategies mitigate this issue by selectively preserving salient KV pairs for autoregressive generation. However, existing methods fail to adequately and efficiently preserve the semantic integrity of the compressed representations. In this paper, we discover a prevalent phenomenon in LLM: within the key embedding space, while most tokens exhibit similarity with their contextual neighbors (we term position-determined tokens), a small subset of special tokens serving as semantic anchors consistently show local deviation property and form one or several clusters (we term semantic-anchored tokens). Motivated by this observation, we propose ProtoKV that separately processes these two token categories for KV cache compression. Within this framework, we first construct semantic prototypes based on the inherent characteristics of the two token categories, and subsequently form clusters of semantically similar tokens as basic compression units. This approach preserves semantic integrity with high computational efficiency. Experiments on LongBench demonstrate that ProtoKV achieves 2.11% higher accuracy than state-of-the-art methods under matched memory constraints.

Planning in open-world environments, where agents must act with partially observed states and incomplete knowledge, is a central challenge in embodied AI. Open-world planning involves not only sequencing actions but also determining what information the agent needs to sense to enable those actions. Existing approaches using Large Language Models (LLM) and Vision-Language Models (VLM) cannot reliably plan over long horizons and complex goals, where they often hallucinate and fail to reason causally over agent-environment interactions. Alternatively, classical PDDL planners offer correct and principled reasoning, but fail in open-world settings: they presuppose complete models and depend on exhaustive grounding over all objects, states, and actions; they cannot address misalignment between goal specifications (e.g., “heat the bread”) and action specifications (e.g., “toast the bread”); and they do not generalize across modalities (e.g., text, vision). To address these core challenges: (i) we extend symbolic PDDL into a flexible natural language representation that we term NL-PDDL, improving accessibility for non-expert users as well as generalization over modalities; (ii) we generalize regression-style planning to NL-PDDL with commonsense entailment reasoning to determine what needs to be observed for goal achievement in partially-observed environments with potential goal–action specification misalignment; and (iii) we leverage the lifted specification of NL-PDDL to facilitate open-world planning that avoids exhaustive grounding and yields a time and space complexity independent of the number of ground objects, states, and actions. Our experiments in three diverse domains — classical Blocksworld and the embodied ALFWorld environment with both textual and visual states — show that NL-PDDL substantially outperforms existing baselines, is more robust to longer horizons and more complex goals, and generalizes across modalities.

The evaluation of large language models faces significant challenges. Technical benchmarks often lack real-world relevance, while existing human preference evaluations suffer from unrepresentative sampling, superficial assessment depth, and single-metric reductionism. To address these issues, we introduce HUMAINE, a framework for multidimensional, demographically aware measurement of human-AI interaction. We collected multi-turn, naturalistic conversations from 23,404 participants that were stratified across 22 demographic groups, both in the US and UK, to evaluate 28 state-of-the-art models across five human-centric dimensions. We use a hierarchical Bayesian Bradley-Terry-Davidson (BTD) model, with post-stratification to census data, and our analysis reveals three key insights. $\textbf{(1)}$ We establish a clear performance hierarchy where $\texttt{google/gemini-2.5-pro}$ ranks first overall, with a 95.6\% posterior probability of being the top-ranked model. $\textbf{(2)}$ We uncover significant preference heterogeneity, with user age emerging as the primary demographic axis of disagreement; a model's perceived rank can shift substantially across age groups, exposing failures in generalisation that unrepresentative samples typically mask. $\textbf{(3)}$ We quantify the vast difference in discriminative power across evaluation dimensions, with ambiguous qualities like Trust, Ethics and Safety showing a 65\% tie rate, in stark contrast to the decisive 10\% tie rate for Overall Winner. Our work emphasises the need for a more multidimensional, demographically aware perspective in LLM evaluation. We release our complete dataset, interactive leaderboard, and open-source framework.

Compact pretrained bidirectional encoders remain the backbone of industrial NLP under tight compute and memory budgets. Their effectiveness stems from self-attention’s ability to deliver high-quality bidirectional contextualization with sequence-level parallelism, as popularized by BERT-style architectures. Recently, **Avey** was introduced as an autoregressive, attention-free alternative that naturally admits an encoder-only adaptation. In this paper, we reformulate Avey for the encoder-only paradigm and propose several innovations to its architecture, including decoupled static and dynamic parameterizations, stability-oriented normalization, and neural compression. Results show that this reformulated architecture compares favorably to four widely used Transformer-based encoders, consistently outperforming them on standard token-classification and information-retrieval benchmarks while scaling more efficiently to long contexts.

Recent advances in multimodal time series learning underscore a paradigm shift from analytics centered on basic patterns toward advanced time series understanding and reasoning. However, existing multimodal time series datasets mostly remain at the level of surface alignment and question answering, without reaching the depth of genuine reasoning. The absence of well-defined tasks that genuinely require time series reasoning, along with the scarcity of high-quality data, has limited progress in building practical time series reasoning models (TSRMs). To this end, we introduce Time Series Reasoning Suite (TSR-Suite), which formalizes four atomic tasks that span three fundamental capabilities for reasoning with time series: (1) perception, acquired through scenario understanding and causality discovery; (2) extrapolation, realized via event-aware forecasting; and (3) decision-making, developed through deliberation over perception and extrapolation. TSR-Suite is the first comprehensive time series reasoning suite that supports not only thorough evaluation but also the data pipeline and training of TSRMs. It contains more than 23K samples, of which 2.3K are carefully curated through a human-guided hierarchical annotation process. Building on this foundation, we introduce TimeOmni-1, the first unified reasoning model designed to address diverse real-world problems demanding time series reasoning. The model is trained in multiple stages, integrating a mixture of task scenarios, novel reward functions, and tailored optimizations. Experiments show that TimeOmni-1 delivers strong out-of-distribution generalization across all tasks and achieves a high rate of valid responses. It significantly improves causality discovery accuracy (64.0% vs. 35.9% with GPT-4.1) and raises the valid response rate by over 6% compared to GPT-4.1 on the event-aware forecasting task.

The multi-commodity flow (MCF) problem is a fundamental topic in network flow and combinatorial optimization, with broad applications in transportation, communication, and logistics, etc. Nowadays, the rapid expansion of allocation systems has posed challenges for existing optimization engines in balancing optimality and tractability. In this paper, we present Pram, the first ML-based method that leverages the reasoning power of multimodal language models (MLMs) for addressing the trade-off dilemma—a great need of service providers. As part of our proposal, Pram (i) quickly computes high-quality allocations by dividing the original problem into local subproblems, which are then resolved by an MLM-powered "agent", and (ii) ensures global consistency by harmonizing these subproblems via a multi-agent reinforcement learning algorithm. Theoretically, we show that Pram, which learns to perform gradient descent in context, provably converges to the optimum within the family of MCF problems. Empirically, on real-world datasets and public topologies, Pram achieves performance comparable to, and in some cases even surpassing, linear programming solvers (very close to the optimal solution), and substantially lower runtimes (one to two orders of magnitude faster). Moreover, Pram exhibits strong robustness (<10\% performance degradation under failures or bursts), demonstrating MLM's generalization ability to unforeseen events. Pram is objective-agnostic and seamlessly integrates with mainstream allocation systems, providing a practical and scalable solution for future networks.

In the field of multimodal chain-of-thought (CoT) reasoning, existing approaches predominantly rely on reasoning on pure linguistic space, which inherently suffers from language bias and is largely confined to math or science domains. This narrow focus limits their ability to handle complex visual reasoning tasks that demand comprehensive understanding of image details. To address these limitations, this paper introduces VGR, a novel reasoning multimodal large language model (MLLM) that can replay the visual memory during thinking just like humans. Unlike traditional MLLMs, VGR first thinks the question and detects relevant regions that may help solve problems, then, the visual memory from the critical area is extracted to assist reasoning. To achieve this, we curate a large-scale SFT dataset called VGR-SFT that contains reasoning data with mixed vision grounding and language deduction. This teaches VGR to think and actively choose grounding areas for key information before answering, and we propose a dynamic visual memory replay stage to integrates the corresponding information into the reasoning process, enhancing multimodel comprehension. Experiments on the LLaVA-NeXT-7B baseline show that VGR achieves superior performance on multimodal benchmarks requiring comprehensive image detail understanding. Compared to the baseline, VGR uses only 30% of the image token count while delivering scores of +4.1 on MMStar, +7.1 on AI2D, and +12.9 improvement on ChartQA.

As large language models (LLMs) are increasingly applied to scientific reasoning, the complexity of answer formats and the diversity of equivalent expressions make answer verification a critical yet challenging task. Existing verification studies in scientific domains suffer from two major limitations: (a) the absence of systematic evaluation standards and insufficient disciplinary coverage, which hinders their comprehensive assessment; and (b) heavy reliance on cumbersome rule design or prompt engineering, which reduces their effectiveness in complex reasoning scenarios or limits their cross-disciplinary generalization. To address these challenges, we propose solutions at both the data and model levels. On the data side, we construct **SCI-VerifyBench**, a cross-disciplinary benchmark covering mathematics, physics, biology, chemistry, and general scientific QA. The benchmark is built from real LLM responses and enhanced with domain-specific equivalence transformations that generate challenging and realistic data. Model-based and expert annotations ensure both quality and diversity, enabling rigorous evaluation of verification ability. On the model side, we emphasize the importance of reasoning for verification and introduce **SCI-Verifier**, a unified reasoning-augmented verifier for scientific domains. Through post-training, SCI-Verifier demonstrates strong logical reasoning and equivalence judgment capabilities while maintaining concise and stable outputs. Together, SCI-VerifyBench and SCI-Verifier provide a principled framework for scientific verification, offering both systematic evaluation and practical pathways to enhance the reliability and applicability of LLMs in scientific domains.

Autoregressive vision-language-action (VLA) models have recently demonstrated strong capabilities in robotic manipulation. However, their core process of action tokenization often involves a trade-off between reconstruction fidelity and inference efficiency. We introduce \textbf{FASTer}, a unified framework for efficient and generalizable robot learning that integrates a learnable tokenizer with an autoregressive policy built upon it. FASTerVQ encodes action chunks as single-channel images, capturing global spatio-temporal dependencies while maintaining a high compression ratio. FASTerVLA builds on this tokenizer with block-wise autoregressive decoding and a lightweight action expert, achieving both faster inference and higher task performance. Extensive experiments across simulated and real-world benchmarks show that FASTerVQ delivers superior reconstruction quality, high token utilization, and strong cross-task and cross-embodiment generalization, while FASTerVLA further improves overall capability, surpassing previous state-of-the-art VLA models in both inference speed and task performance.

Recent advances demonstrate that multimodal large language models (MLLMs) exhibit strong multimodal in-context learning (ICL) capabilities, enabling them to adapt to novel vision-language tasks from a few contextual examples. However, existing ICL approaches face challenges in reconciling generalization with robustness across diverse tasks and noisy contextual examples: manually selecting examples produces clean contexts but is labor-intensive and task-specific, while similarity-based retrieval improves scalability but could introduce irrelevant or structurally inconsistent samples that degrade ICL performance. To address these limitations, we propose ContextNav, the first agentic framework that integrates the scalability of automated retrieval with the quality and adaptiveness of human-like curation, enabling noise-robust and dynamically optimized contextualization for multimodal ICL. ContextNav unifies context management and noise-robust contextualization within a closed-loop workflow driven by graph-based tool orchestration. Specifically, it builds a resource-aware multimodal embedding pipeline, maintains a retrievable vector database, and applies agentic retrieval and structural alignment to construct noise-resilient contexts. An Operational Grammar Graph (OGG) further supports adaptive toolchain planning and optimization, enabling the agent to refine its strategies based on downstream feedback. Experimental results demonstrate that ContextNav achieves state-of-the-art performance across various datasets, underscoring the promise of agentic workflows for advancing scalable and robust contextualization in multimodal ICL. The code will be made publicly available on GitHub.

Cultural awareness capabilities has emerged as a critical capability for Multimodal Large Language Models (MLLMs). However, current benchmarks lack progressed difficulty in their task design and are deficient in cross-lingual tasks. Moreover, current benchmarks often use real-world images. Each real-world image typically contains one culture, making these benchmarks relatively easy for MLLMs. Based on this, we propose C$^3$B ($\textbf{C}$omics $\textbf{C}$ross-$\textbf{C}$ultural $\textbf{B}$enchmark), a novel multicultural, multitask and multilingual cultural awareness capabilities benchmark. C$^3$B comprises over 2000 images and over 18000 QA pairs, constructed on three tasks with progressed difficulties, from basic visual recognition to higher-level cultural conflict understanding, and finally to cultural content generation. We conducted evaluations on 11 open-source MLLMs, revealing a significant performance gap between MLLMs and human performance. The gap demonstrates that C$^3$B poses substantial challenges for current MLLMs, encouraging future research to advance the cultural awareness capabilities of MLLMs.

Large Language Models (LLMs) are increasingly expected to be culturally customizable and personally aligned for natural language understanding (NLU). However, existing methods, from supervised fine-tuning (SFT) to personalized RLHF and prompting, either require costly large-scale annotations or remain constrained by pretraining distributions. Moreover, acquiring annotations that reflect subjective, diverse, and evolving user preferences is both expensive and labor-intensive. To address these limitations, we propose \textit{\textbf{H}eterogeneous-\textbf{C}onsistency \textbf{C}o-Alignment} (HCC) is a training-free annotation paradigm that leverages two heterogeneous models, which consists of an LLM, rich in knowledge yet often prone to overconfidence, is paired with a task-specialised lightweight model guided by a small user-preference set to verify and co-align misaligned outputs over unlabeled corpora. For verification, HCC introduces the reference-free \textit{\textbf{C}onsistent}-\textit{\textbf{A}nd}-\textit{\textbf{I}nconsistent} (\textbf{CAI}) Ratio, an uncertainty signal derived from inter-model agreements (consistent samples) and disagreements (inconsistent samples) to determine when refinement is needed. For co-alignment, HCC employs a non-parametric, embedding-based preference assignment scheme to recalibrate inconsistent samples according to user preferences. Across eight NLU datasets and both open- and closed-source LLMs, HCC consistently improves annotation quality and, in several tasks, even enables \textit{Llama-3-8B} to surpass \textit{GPT-3.5/4o} after co-alignment. Moreover, CAI correlates strongly with accuracy and reliably tracks pre-/post-alignment gains, offering a reference-free signal for scaling preference-aligned annotation.

The LLM unlearning aims to eliminate the influence of undesirable data without affecting causally unrelated information. This process typically involves using a **forget set** to remove target information, alongside a **retain set** to maintain non-target capabilities. While recent localization-based methods demonstrate promise in identifying important nodes (neurons) to be unlearned, they fail to disentangle nodes responsible for forgetting undesirable knowledge or retaining essential skills, often treating them as a single entangled group. As a result, these methods apply uniform interventions, risking catastrophic over-forgetting or incomplete erasure of the target knowledge. To address this, we turn to circuit discovery, a mechanistic interpretability technique, and propose the **C**onflict-guided **L**ocalization for LLM **U**nlearning fram**E**work (**CLUE**). This framework identifies the forget and retain circuit composed of important nodes, and then the circuits are transformed into conjunctive normal forms (CNF). The assignment of each node in the CNF satisfiability solution reveals whether it should be forgotten or retained. We then provide targeted fine-tuning strategies for different categories of nodes. Extensive experiments demonstrate that, compared to existing localization methods, CLUE achieves superior forget efficacy and retain utility through precise neural localization.

Does reinforcement learning (RL) teach large language models (LLMs) genuinely new skills, or does it merely activate existing ones? This question lies at the core of ongoing debates about the role of RL in LLM post-training. On one side, strong empirical results can be achieved with RL alone even without preceding supervised finetuning; on the other, critics argue that RL contributes little beyond reweighting existing reasoning strategies. This work provides concrete evidence that LLMs can acquire genuinely new skills during RL by composing existing ones, mirroring one of the central mechanisms by which humans acquire new cognitive skills \citep{Anderson1982Acquisition}. To mitigate data contamination and other confounding factors and to allow precise control over task complexity, we develop a synthetic framework for our investigation. Specifically, we define a skill as the ability to infer the output of a string transformation function $f(x)$ given $x$. Once an LLM has already learned $f$ and $g$ prior to RL, our experiments reveal that RL enables it to learn unseen compositions of them $h(x)=g(f(x))$. Further, this compositional ability generalizes to more difficult problems such as compositions of $>2$ functions unseen during training. Our experiments provide surprising evidence that this compositional ability, acquired on the source task, transfers to a different target task. This transfer occurs even though the model has never trained with RL on any compositional problems in the target task, as long as it has acquired the target task's atomic skills prior to RL on the source task. Our qualitative analysis shows that RL fundamentally changes the reasoning behaviors of the models. In contrast, neither of the findings is observed in next-token prediction training with the same data. Our systematic experiments provide fresh insights into the learning behaviors of widely-used post-training approaches for LLMs. They suggest the value of building base models with the necessary basic skills, followed by RL with appropriate incentivization to acquire more advanced skills that generalize better to complex and out-of-domain problems.

Increasing test-time computation is a straightforward approach to enhancing the quality of responses in Large Language Models (LLMs). While Best-of-N sampling and Self-Consistency with majority voting are simple and effective, they require a fixed number of sampling responses for each query, regardless of its complexity. This could result in wasted computation for simpler questions and insufficient exploration for more challenging ones. In this work, we argue that model confidence of responses can be used for improving the efficiency of test-time scaling. Unfortunately, LLMs are known to be overconfident and provide unreliable confidence estimation. To address this limitation, we introduce Self-Calibration by distilling Self-Consistency-derived confidence into the model itself. This enables reliable confidence estimation at test time with one forward pass. We then design Calibrated Test-Time Scaling (CaTS), adapting common repeated sampling methods, such as self-consistency and Best-of-N to handle queries of various difficulty. We also show that CaTS-SC is provably better than vanilla self-consistency. Experiments on three LLMs across nine datasets demonstrate the effectiveness of our approach. Specifically, applying confidence-based Early Stopping (CaTS-ES) to Best-of-N improves MathQA accuracy from 73.7 to 83.6 with a sample budget of 16 responses, demonstrating the effectiveness of the confidence-based sampling strategy at inference time.

Vision-language alignment in multi-modal large language models (MLLMs) relies on supervised fine-tuning (SFT) or reinforcement learning (RL). To align multi-modal large language models (MLLMs) in the post-training stage, supervised fine-tuning (SFT) is a stable choice but requires human annotations and lacks task generalizations, while Reinforcement Learning (RL) searches for better answers from reward signals but suffers from computational overhead and instability. To achieve balance among scalability, efficiency, and alignment generalizations, we propose MergeMix, a unified paradigm that bridges SFT and RL with an efficient Token Merge based Mixup augmentation. As for the Mixup policy, we generate contextual aligned mixed images with the corresponding labels according to the merged attention maps with cluster regions. Then, we enhance the preference-driven paradigm for MLLMs by building preference pairs with raw images and MergeMix-generated ones and optimizing the soft preference margin with the mixed SimPO loss. Extensive experiments demonstrate that MergeMix not only achieves dominant classification accuracy as an augmentation method but also improves generalization abilities and alignment of MLLMs, providing a new learning paradigm for preference alignment with training efficiency and stability.