ICLR 2026 论文列表

\begin{abstract} Developing generalist large language models (LLMs) capable of complex biological reasoning is a central challenge in computational biology. While existing LLMs excel at predictive tasks like cell type annotation and logically-constrained problems, enabling open-ended and mechanistic reasoning remains a challenge. A promising direction is Reinforcement Learning from Verifiable Rewards (RLVR), which has been shown to significantly enhance complex reasoning in general domains like mathematics and code synthesis. However, its application in biology is hindered, as most biological outcomes are non-verifiable. For example, verifying a generated gene sequence is usually infeasible. In this paper, we introduce CellDuality, a self-supervised framework that enables LLM agents for robust reasoning in single-cell biology. Our framework is built on the principle of complementary task duality, a self-verification process that leverages a bidirectional reasoning loop. First, the model performs a forward reasoning task by predicting a biological outcome (e.g., a cell's response to a drug). Then, in a complementary inverse task, it must reason backward from its own prediction to reconstruct the initial conditions (e.g., the original drug perturbation). The fidelity of this reconstruction serves as an intrinsic reward signal, creating a feedback loop that enforces logical and biological consistency. We use these intrinsic rewards to align the base LLM via reinforcement learning, without requiring ground-truth verification labels. We demonstrate that CellDuality achieves state-of-the-art performance and provides coherent biological explanations across a diverse suite of single-cell reasoning tasks. Critically, on the challenging out-of-distribution perturbation prediction benchmark, our self-supervised approach significantly outperforms the standard fine-tuning baseline and narrows the performance gap to a supervised RLVR baseline. Our work showcases a new path toward scalable training of biological foundation models.

AI models have made significant strides in recent years in their ability to describe and answer questions about real-world images. They have also made progress in the ability to converse with users in real-time using audio input. This raises the question: have we reached the point where AI models, connected to a camera and microphone, can converse with users in real-time about scenes and events that are unfolding live in front of the camera? This has been a long-standing goal in AI and is a prerequisite for real-world AI assistants and humanoid robots to interact with humans in everyday situations. In this work, we introduce a new dataset and benchmark, the Interactive Video Dataset (IVD), which allows us to assess the extent to which existing models can support these abilities, and to what degree these capabilities can be instilled through fine-tuning. The dataset is based on a simple question-answering setup, where users ask questions that the system has to answer, in real-time, based on the camera and audio input. We show that existing models fall far behind human performance on this task, and we identify the main sources for the performance gap. However, we also show that for many of the required perceptual skills, fine-tuning on this form of data can significantly reduce this gap.

Vision-language models have achieved remarkable success in multi-modal representation learning from large-scale pairs of visual scenes and linguistic descriptions. However, they still struggle to simultaneously express two distinct types of semantic structures: the hierarchy within a concept family (e.g., *dog* $\preceq$ *mammal* $\preceq$ *animal*) and the compositionality across different concept families (e.g., "a dog in a car" $\preceq$ *dog*, *car*). Recent works have addressed this challenge by employing hyperbolic space, which efficiently captures tree-like hierarchy, yet its suitability for representing compositionality remains unclear. To resolve this dilemma, we propose *PHyCLIP*, which employs an $\ell_1$-*P*roduct metric on a Cartesian product of *Hy*perbolic factors. With our design, intra-family hierarchies emerge within individual hyperbolic factors, and cross-family composition is captured by the $\ell_1$-product metric, analogous to a Boolean algebra. Experiments on zero-shot classification, retrieval, hierarchical classification, and compositional understanding tasks demonstrate that PHyCLIP outperforms existing single-space approaches and offers more interpretable structures in the embedding space.

This work identifies anisotropy in the singular value spectra of parameters, activations, and gradients as the fundamental barrier to low-bit training of large language models (LLMs). These spectra are dominated by a small fraction of large singular values, inducing wide numerical ranges that cause quantization bias and severe spectral distortion, ultimately degrading training performance. This work presents \emph{Metis}, a spectral-domain quantization framework that partitions anisotropic spectra into narrower sub-distributions for independent quantization, thereby reducing errors and preserving spectral structure. To minimize overhead, Metis leverages two key properties of the dominant spectral subspace: preservation via sparsely random sampling and preservation via random projection, reducing decomposition cost to a negligible level. On LLaMA-3 8B trained with 100B tokens, Metis enables robust W4A4G4 training with FP4 quantization of weights, activations, and gradients, yielding only a 0.4\% training loss gap and a 0.1\% degradation in downstream accuracy relative to BF16. Beyond matching BF16 fidelity, Metis also surpasses our implementation of Nvidia’s recently announced (yet to be publicly released) FP4 recipe, consistently achieving lower loss and higher downstream accuracy while incurring significantly lower computational overhead. The code implementation for Metis is available at: \url{https://anonymous.4open.science/r/Metis-quantization-644B}.

Structured images (e.g., charts and geometric diagrams) remain challenging for multimodal large language models (MLLMs), as perceptual slips can cascade into erroneous conclusions. Intermediate visual cues can steer reasoning; however, existing cue-based methods are constrained with low-fidelity image processing and linear, rigid reasoning patterns, limiting their effectiveness on complex structured-image tasks. In this paper, we propose PixelCraft, a novel multi-agent system for high-fidelity image processing and flexible visual reasoning on structured images. The system comprises a dispatcher, a planner, a reasoner, critics, and a set of visual tool agents. To achieve high-fidelity processing, we construct a high-quality corpus and fine-tune an MLLM into a grounding model, whose pixel-level localizations are integrated with traditional computer vision (CV) algorithms in tool agents. Building on this foundation, PixelCraft facilitates flexible visual reasoning through a dynamic three-stage workflow of tool selection, agent discussion, and self-criticism. Moreover, unlike prior linear reasoning patterns that simply append historical images, PixelCraft maintains an image memory to allow the planner to adaptively revisit earlier visual steps, explore alternative reasoning branches, and dynamically adjust the reasoning trajectory during discussion. Extensive experiments on challenging chart and geometry benchmarks demonstrate that PixelCraft significantly improves visual reasoning performance for advanced MLLMs, setting a new standard for structured image reasoning.

Recent progress in large language models (LLMs) has outpaced the development of effective evaluation methods. Evaluating LLMs with static, task-specific benchmarks is increasingly fragile due to contamination and saturation, and it fails to capture interactive reasoning. We introduce Teach2Eval, which reframes evaluation as teaching: a candidate model guides weaker students, and the students’ gains constitute the score. This interaction yields robustness to contamination and exposes orthogonal abilities with fine-grained metrics across Application, Judgment, Guidance, and Reflection. The framework scales automatically by exploiting natural error distributions from weak students, requiring neither bespoke rubrics nor human graders. Across 30 LLMs and 60 datasets, Teach2Eval achieves Spearman above 0.95 with human-preference leaderboards (e.g., Chatbot Arena/LiveBench), surpassing direct baselines, while offering actionable training signals (capability hierarchies, early overfitting) at low cost.

The Transformer model architecture has become one of the most widely used in deep learning and the attention mechanism is at its core. The standard attention formulation uses a softmax operation applied to a scaled dot product between query and key vectors. We explore the role played by norms of the queries and keys, which can cause training instabilities when they arbitrarily increase. We demonstrate how this can happen even in simple Transformer models, in the presence of easy-to-learn spurious patterns in the data. We propose a new attention formulation, QUEry-modulated Spherical aTtention (QUEST), that constrains the keys to a hyperspherical latent space, while still allowing individual tokens to flexibly control the sharpness of the attention distribution. QUEST can be easily used as a drop-in replacement for standard attention. We focus on vision applications while also exploring other domains to highlight the method's generality. We show that (1) QUEST trains without instabilities and (2) produces models with improved performance (3) that are robust to data corruptions and adversarial attacks.

Multimodal Large Language Models (MLLMs) combine the linguistic strengths of LLMs with the ability to process multimodal data, enabling them to address a broader range of tasks. This progression highlights a shift from language-only reasoning to integrated vision–language reasoning in children's development. Inspired by the Wechsler Intelligence Scales, we introduce KidGym, a comprehensive 2D grid-based benchmark for assessing five essential capabilities of MLLMs: execution, perception reasoning, learning, memory, and planning. The benchmark comprises 12 unique tasks, each targeting at least one core capability, specifically designed to gauge MLLMs' adaptability and developmental potential, mirroring the stages of children's cognitive growth. Additionally, our tasks encompass diverse scenarios and objects with randomly generated layouts, ensuring more accurate and robust evaluation of MLLM capabilities. KidGym is designed to be fully user-customizable and extensible, allowing researchers to create new evaluation scenarios and adjust difficulty levels to accommodate the rapidly growing MLLM community. Through evaluation of state-of-the-art MLLMs using KidGym, we identified significant insights into model capabilities and revealed important limitations of current status. We release our benchmark at: https://kidgym.github.io/KidGym-Website/.

Fine-tuning on open-source Large Language Models (LLMs) with proprietary data is now a standard practice for downstream developers to obtain task-specific models. Surprisingly, we reveal a new and concerning risk along with the practice: the provider of the open-source LLMs can later extract the private downstream fine-tuning data through simple backdoor training, only requiring black-box access to the fine-tuned downstream model. Our comprehensive experiments, across 4 popularly used open-source models with 3B to 32B parameters and 2 downstream datasets, suggest that the extraction performance can be strikingly high: in practical settings, as much as 76.3\% downstream fine-tuning data (queries) out of a total 5,000 samples can be perfectly extracted, and the success rate can increase to 94.9\% in more ideal settings. We further investigate several defense strategies, but none achieve satisfactory effectiveness in mitigating the risk. Overall, we highlight the emergency of this newly identified data breaching risk in fine-tuning, and we hope more follow-up research can push the progress of addressing this concerning risk.

Recent advances in learning rate~(LR) scheduling have demonstrated the effectiveness of decay-free approaches that eliminate the traditional decay phase while maintaining competitive performance. Model merging techniques have emerged as particularly promising solutions in this domain. We present Warmup-Stable and Merge (WSM), a general framework that establishes a formal connection between learning rate decay and model merging. WSM provides a unified theoretical foundation for emulating various decay strategies—including cosine decay, linear decay and inverse square root decay—as principled model averaging schemes, while remaining fully compatible with diverse optimization methods. Through extensive experiments, we identify merge duration—the training window for checkpoint aggregation—as the most critical factor influencing model performance, surpassing the importance of both checkpoint interval and merge quantity. Our framework consistently outperforms the widely-adopted Warmup-Stable-Decay (WSD) approach across multiple benchmarks, achieving significant improvements of +3.5\% on MATH, +2.9\% on HumanEval, and +5.5\% on MMLU-Pro. The performance advantages extend to supervised fine-tuning scenarios, highlighting WSM's potential for long-term model refinement.

Large Vision-Language Models (VLMs) have achieved remarkable performance across a wide range of tasks. However, their deployment in safety-critical domains poses significant challenges. Existing safety fine-tuning methods, which focus on textual or multimodal content, fall short in addressing challenging cases or disrupt the balance between helpfulness and harmlessness. Our evaluation highlights a safety reasoning gap: these methods lack safety visual reasoning ability, leading to such bottlenecks. To address this limitation and enhance both visual perception and reasoning in safety-critical contexts, we propose a novel dataset that integrates multi-image inputs with safety Chain-of-Thought (CoT) labels as fine-grained reasoning logic to improve model performance. Specifically, we introduce the Multi-Image Safety (MIS) dataset, an instruction-following dataset tailored for multi-image safety scenarios, consisting of training and test splits. Our experiments demonstrate that fine-tuning InternVL2.5-8B with MIS significantly outperforms both powerful open-source models and API-based models in challenging multi-image tasks requiring safety-related visual reasoning. This approach not only delivers exceptional safety performance but also preserves general capabilities without any trade-offs. Specifically, fine-tuning with MIS increases average accuracy by 0.83% across five general benchmarks and reduces the Attack Success Rate (ASR) on multiple safety benchmarks by a large margin.

Class Incremental Learning (CIL) enables models to adapt to evolving data distributions by learning new classes over time without revisiting previous data. While recent methods utilizing pre-trained models have shown promising results, they often assume access to fully labeled data for each incremental task, which is often impractical. In this paper, we instead tackle a more realistic scenario in which only unlabeled data and the class-name set are available for each new class. Although one could generate pseudo labels with a vision-language model and apply existing CIL methods, the inevitable noise in these pseudo labels tends to aggravate catastrophic forgetting. To overcome this challenge, we propose a method named N2L employing a regression objective with mean squared error loss, which can be solved in a recursive manner. To refine the pseudo labels, N2L applies feature dimensionality reduction to the extracted image features and iteratively updates the labels using a classifier trained on these reduced features. Furthermore, a bi-level weight adjustment strategy is proposed to downweight low-confidence pseudo labels via intra-class adjustment and compensate for pseudo-label class imbalance through inter-class adjustment. This incremental learning with adjustment can be solved recursively, yielding identical performance to joint training with unlabeled data and thereby mitigating forgetting. Our theoretical analysis supports the effectiveness of the pseudo label refinement process, and experiments on various datasets demonstrate that our proposed method outperforms SOTA methods. Code is provided in the appendix.

Omni-modal large language models (OLLMs) aim to unify audio, vision, and text understanding within a single framework. While existing benchmarks have advanced multimodal evaluation, it remains unclear whether OLLMs achieve modality-invariant reasoning or inherit modality-specific biases. We introduce \textbf{XModBench}, a large-scale tri-modal benchmark explicitly designed to measure cross-modal consistency. XModBench contains 60K multiple-choice questions across five task families and systematically covers all six cross-modality directions, enabling diagnosis of task competence, modality disparity, and directional imbalance. Experiments show that even the strongest model, Gemini 2.5 Pro, (i) struggles with spatial and temporal reasoning, achieving less than 60% accuracy, (ii) suffers from modality disparities, with performance dropping by over {20 points} on average when audio inputs replace text, and (iii) exhibits directional imbalance, with a {9-point gap} when using vision as context versus using text as context. The findings suggest that OLLMs fall short of modality-invariant reasoning, and XModBench provides a fundamental diagnostic tool for evaluating and improving their overall cross-modal competence.

Linear attention mechanisms have emerged as efficient alternatives to Softmax attention, exhibiting steady improvements in language modeling capabilities driven by increasingly sophisticated designs for decay matrices—though their structural complexity has typically been limited to the Diagonal-Plus-Rank-1 level. To further advance the understanding and capabilities of linear attention via more complex decay structures, this work makes two primary contributions: (1) We propose the HDLA linear attention mechanism, which utilizes efficient matrix decomposition to achieve a Diagonal-Plus-Rank-2 structure, thereby extending the decay matrix to a broader, more expressive, rank-enhanced and structured class. (2) We propose a more general chunk-wise parallel algorithm that accommodates both diagonal-plus-rank-$r_{ab}$ decay structure and key-value outer products of rank $r_{kv}$, thus providing a versatile foundation for future research. Comprehensive experiments demonstrate that, compared to linear attention baselines, HDLA sets new SOTA results on language modeling and retrieval tasks at 2.8B parameter scale, delivers at most 80\% and 58.2\% performance gains over baselines on retrieval-based MQAR and RULER tasks, and achieves an average score improvement of 4.39–7.66 on the synthetic MAD benchmark, respectively. Our proposed HDLA model, as well as the rank-generalized chunk-wise parallel algorithm, together provide a versatile algorithmic foundation and promising research prospects for the design of rank-enhanced, structured linear attention mechanisms.

We introduce Warping-Alone Field Transforms (WAFT), a simple and effective method for optical flow. WAFT is similar to RAFT but replaces cost volume with high-resolution warping, achieving better accuracy with lower memory cost. This design challenges the conventional wisdom that constructing cost volumes is nec- essary for strong performance. WAFT is a simple and flexible meta-architecture with minimal inductive biases and reliance on custom designs. Compared with existing methods, WAFT ranks 1st on Spring, Sintel, and KITTI benchmarks, achieves the best zero-shot generalization on KITTI, while being up to 4.1× faster than methods with similar performance. Code and model weights will be available upon acceptance.

We **color-coded** the added changes to the **paper** and **Appendix** for the comfort of our reviewers. Multimodal Language Models (MMLMs) typically undergo post-training alignment to prevent harmful content generation. However, these alignment stages focus primarily on the *assistant* role, leaving the *user* role unaligned, and sticking to a fixed input prompt structure of special tokens, making the model vulnerable when inputs deviate from these expectations. We introduce Role-Modality Attacks (RMA), a novel class of adversarial attacks that exploit role confusion between the *user* and *assistant* and alter the position of the *image* token to elicit harmful outputs. Unlike existing attacks that modify query content, RMAs manipulate the input structure without altering the query itself. We systematically evaluate these attacks across multiple Vision Language Models (VLMs) on eight distinct settings, showing that they can be composed to create stronger adversarial prompts, as also evidenced by their increased projection in the negative refusal direction in the residual stream, a property observed in prior successful attacks. Finally, for mitigation, we propose an adversarial training approach that makes the model robust against input prompt perturbations. By training the model on a range of harmful and benign prompts all perturbed with different RMA settings, the model loses its sensitivity to Role Confusion and Modality Manipulation attacks and is trained to only pay attention to the query content in the input prompt structure, effectively reducing Attack Success Rate (ASR) while preserving the model’s general utility.

Multimodal Large Language Models (MLLMs) have achieved notable success in enhancing translation performance by integrating multimodal information. However, existing research primarily focuses on image-guided methods, whose applicability is constrained by the scarcity of multilingual image-text pairs. The speech modality overcomes this limitation due to its natural alignment with text and the abundance of existing speech datasets, which enable scalable language coverage. In this paper, we propose a \textbf{Speech-guided Multimodal Machine Translation (SMMT)} framework that integrates speech and text as fused inputs into an MLLM to improve translation quality. To mitigate reliance on low-resource data, we introduce a \textbf{Self-Evolution Mechanism}. The core components of this framework include a text-to-speech model, responsible for generating synthetic speech, and an MLLM capable of classifying synthetic speech samples and iteratively optimizing itself using positive samples. Experimental results demonstrate that our framework surpasses all existing methods on the Multi30K multimodal machine translation benchmark, achieving new state-of-the-art results. Furthermore, on general machine translation datasets, particularly the FLORES-200, it achieves average state-of-the-art performance in 108 translation directions. Ablation studies on CoVoST-2 confirms that differences between synthetic and authentic speech have negligible impact on translation quality. We will open-source our model to support the wider community.

Despite recent progress in text-to-image (T2I) generation, existing models often struggle to faithfully capture user intentions from short and under-specified prompts. While prior work has attempted to enhance prompts using large language models (LLMs), these methods frequently generate stylistic or unrealistic content due to insufficient grounding in visual semantics and real-world composition. Inspired by recent advances in reasoning for language model, we propose RePrompt, a novel reprompting framework that introduces explicit reasoning into the prompt enhancement process via reinforcement learning. Instead of relying on handcrafted rules or stylistic rewrites, our method trains a language model to generate structured, self-reflective prompts by optimizing for image-level outcomes. The tailored reward models assesse the generated images in terms of human preference, semantic alignment, and visual composition, providing indirect supervision to refine prompt generation. Our approach enables end-to-end training without human-annotated data. Experiments on GenEval and T2I-Compbench show that RePrompt significantly boosts spatial layout fidelity and compositional generalization across diverse T2I backbones, establishing new state-of-the-art results. Code: https://github.com/microsoft/DKI_LLM/tree/main/RePrompt.

The recent advancement of Large Language Models (LLMs) has revolutionized the algorithm design patterns. A new paradigm, LLM-based Automated Algorithm Design (LLM-AAD), has emerged to generate code implementations for high-quality algorithms. Unlike the traditional expert-driven algorithm development, in the LLM-AAD paradigm, ideas behind the algorithm are often implicitly embedded within the generated code. Therefore, measuring similarity for algorithms may help identify whether a generated algorithm is innovative or merely a syntactic refinement of an existing code implementation. However, directly applying existing code similarity metrics to algorithms raises a critical limitation: they do not necessarily reflect the similarity between algorithms. To address this, we introduce a novel perspective that defines algorithm similarity through the lens of its problem-solving behavior. We represent the problem-solving trajectory of an algorithm as the sequence of intermediate solutions progressively generated by the algorithm. The behavioral similarity is calculated by the resemblance between two problem-solving trajectories. Our approach focuses on how an algorithm solves a problem, not just its code implementation or final output. We demonstrate the utility of our similarity measure in two use cases. (i) Improving LLM-AAD: Integrating our similarity measure into a search method demonstrates promising results across two AAD tasks, proving the effectiveness of maintaining behavioral diversity in the algorithm search. (ii) Algorithm analysis. Our similarity metric provides a new perspective for analyzing algorithms, revealing distinctions in their problem-solving behaviors.