ICLR 2026 论文列表

While foundation models (FMs), such as diffusion models and large vision-language models (LVLMs), have been widely applied in educational contexts, their ability to generate pedagogically effective visual explanations remains limited. Most existing approaches focus primarily on textual reasoning, overlooking the critical role of structured and interpretable visualizations in supporting conceptual understanding. To better assess the visual reasoning capabilities of FMs in educational settings, we introduce EduVisBench, a multi-domain, multi-level benchmark. EduVisBench features diverse STEM problem sets requiring visually grounded solutions, along with a fine-grained evaluation rubric informed by pedagogical theory. Our empirical analysis reveals that existing models frequently struggle with the inherent challenge of decomposing complex reasoning and translating it into visual representations aligned with human cognitive processes. To address these limitations, we propose EduVisAgent, a multi-agent collaborative framework that coordinates specialized agents for instructional planning, reasoning decomposition, metacognitive prompting, and visualization design. Experimental results show that EduVisAgent substantially outperforms all baselines, achieving a 40.2% improvement and delivering more educationally aligned visualizations.

Large language model (LLM) alignment is typically achieved through learning from human preference comparisons, making the quality of preference data critical to its success. Existing studies often pre-process raw training datasets to identify valuable preference pairs using external reward models or off-the-shelf LLMs, achieving improved overall performance but rarely examining whether individual, selected data point is genuinely beneficial. We assess data quality through individual influence on validation data using our newly proposed truncated influence function (TIF), which mitigates the over-scoring present in traditional measures and reveals that preference data quality is inherently a property of the model. In other words, a data pair that benefits one model may harm another. This leaves the need to improve the preference data selection approaches to be adapting to specific models. To this end, we introduce a set of candidate scoring functions (SFs) that are computationally simpler than TIF and positively correlated with it. These functions are also model dependent and can serve as potential indicators of individual data quality for preference data selection. Furthermore, we observe that these SFs inherently exhibit errors when compared to TIF. To this end, we combine them to offset their diverse error sources, resulting in a simple yet effective data selection rule that enables the models to achieve a more precise selection of valuable preference data. We conduct experiments across diverse alignment benchmarks and various LLM families, with results demonstrating that better alignment performance can be achieved using less data, showing the generality and robustness of our findings and new methods.

Modern video-text retrieval (VTR) models excel on in-distribution benchmarks but are highly vulnerable to real-world *query shifts*, where the distribution of query data deviates from the training domain, leading to a sharp performance drop. Existing image-focused robustness solutions are inadequate to handle this vulnerability in video, as they fail to address the complex spatio-temporal dynamics inherent in these shifts. To systematically evaluate this vulnerability, we first introduce a comprehensive benchmark featuring 12 distinct types of video perturbations across five severity degrees. Analysis on this benchmark reveals that query shifts amplify the *hubness phenomenon*, where a few gallery items become dominant "hubs" that attract a disproportionate number of queries. To mitigate this, we then propose HAT-VTR (Hubness Alleviation for Test-time Video-Text Retrieval), as our baseline test-time adaptation framework designed to directly counteract hubness in VTR. It leverages two key components: a *Hubness Suppression Memory* to refine similarity scores, and *multi-granular losses* to enforce temporal feature consistency. Extensive experiments demonstrate that HAT-VTR substantially improves robustness, consistently outperforming prior methods across diverse query shift scenarios, and enhancing model reliability for real-world applications.

Practical guidance on training Large Language Models (LLMs) to leverage Code Interpreter across diverse tasks remains lacking. We present R1-Code-Interpreter, an extension of a text-only LLM trained via multi-turn supervised fine-tuning (SFT) and reinforcement learning (RL) to autonomously generate multiple code queries during step-by-step reasoning. Unlike prior RL + tool-use efforts focused on narrow domains such as math or retrieval, we curate 144 diverse reasoning and planning tasks and show that training a general-purpose Code Interpreter across them presents significant challenges due to task heterogeneity and scarcity of effective samples. To address this, we introduce a multi-stage curriculum learning approach that partitions training samples by measured improvement potential. The RL training prioritizes samples with higher potential and gradually shifts to lower-potential ones, increasing the average RL gains from merely +3.4\% to +9.3\% across Qwen-2.5 models (3/7/14B). Our final model, R1-CI-14B, improves average accuracy on the 37 test tasks from 44.1\% to 72.4\%, outperforming text-only GPT-4o (58.6\%) and GPT-4o with Code Interpreter (70.9\%). Notably, R1-CI-14B also exhibits emergent self-checking behavior through code generation.

Reinforcement learning (RL) has become central to training large language models (LLMs), yet the field lacks predictive scaling methodologies comparable to those established for pre-training. Despite rapidly rising compute budgets, there is no principled understanding of how to evaluate algorithmic improvements for scaling RL compute. We present the first large-scale systematic study, amounting to more than 400,000 GPU-hours, that defines a principled framework for analyzing and predicting RL scaling in LLMs. We fit sigmoidal compute-performance curves for RL training and ablate a wide range of common design choices to analyze their effects on asymptotic performance and compute efficiency. We observe: (1) Not all recipes yield similar asymptotic performance, Details such as loss aggregation, normalization, curriculum, and off-policy algorithm primarily modulate compute efficiency without materially shifting the asymptote, and (3) Stable, scalable recipes follow predictable scaling trajectories, enabling extrapolation from smaller-scale runs. Combining these insights, we propose a _best-practice_ recipe, ScaleRL, and demonstrate its effectiveness by successfully scaling and predicting validation performance on a single RL run scaled up to 100,000 GPU-hours. Our work provides both a _scientific framework_ for analyzing scaling in RL and a practical recipe that brings RL training closer to the predictability long achieved in pre-training.

Unified multimodal Large Language Models (LLMs) that can both understand and generate visual content hold immense potential. However, existing open-source models often suffer from a performance trade-off between these capabilities. We present Manzano, a simple and scalable unified framework that substantially reduces this tension by coupling a hybrid image tokenizer with a well-curated training recipe. A single shared vision encoder feeds two lightweight adapters that produce continuous embeddings for image-to-text understanding and discrete tokens for text-to-image generation within a common semantic space. A unified autoregressive LLM predicts high-level semantics in the form of text and image tokens, with an auxiliary diffusion decoder subsequently translating the image tokens into pixels. The architecture, together with a unified training recipe over understanding and generation data, enables scalable joint learning of both capabilities. Manzano achieves state-of-the-art results among unified models, and is competitive with specialist models, particularly on text-rich evaluation. Our studies show minimal task conflicts and consistent gains from scaling model size, validating our design choice of a hybrid tokenizer.

Recent findings reveal that token embeddings of Large Language Models (LLMs) exhibit strong hyperbolicity. This insight motivates leveraging LLMs for scientific taxonomy tasks, where maintaining and expanding hierarchical knowledge structures is critical. Although potential, generally-trained LLMs face challenges in directly handling domain-specific taxonomies, including computational cost and hallucination. Meanwhile, Small Language Models (SLMs) provide a more economical alternative if empowered with proper knowledge transfer. In this work, we introduce SS-Mono (Structure-Semantic Monotonization), a novel pipeline that combines local taxonomy augmentation from LLMs, self-supervised fine-tuning of SLMs with geometric constraints, and LLM calibration. Our approach enables efficient and accurate taxonomy expansion across root, leaf, and intermediate nodes. Extensive experiments on both leaf and non-leaf expansion benchmarks demonstrate that a fine-tuned SLM (e.g., DistilBERT-base-110M) consistently outperforms frozen LLMs (e.g., GPT-4o, Gemma-2-9B) and domain-specific baselines. These findings highlight the promise of lightweight yet effective models for structured knowledge enrichment in scientific domains.

Large vision–language models (VLMs) deliver state-of-the-art results on a wide range of multimodal tasks, yet they remain prone to visual hallucinations, producing content that is not grounded in the input image. Despite progress with visual supervision, reinforcement learning, and post-hoc attention reshaping, the representational origins of hallucinations remain unclear. Our study reveals that successful grounding emerges when adjacent visual tokens exhibit coherent alignment, while hallucinations arise when key vectors scatter isotropically, weakening cross-modal attention and blurring object boundaries. Building on this insight, we propose Depth and Spatial aware Cache Refinement (DSCR), a lightweight and training-free method that augments the Transformer's key-value (KV) cache with depth cues and 2D spatial proximity. DSCR clusters vectors within objects and separates those across surfaces, guiding attention toward relevant regions without any fine-tuning. Comprehensive evaluations show that DSCR consistently reduces hallucinations, delivering up to 23\% accuracy gains across MME, POPE, RePOPE, CHAIR, and a new depth-sensitive benchmark. Our findings highlight KV-coherence as a core factor behind hallucinations and demonstrate a practical, model-agnostic solution for enhancing VLM reliability.

Inferring ground truth from noisy, crowdsourced labels is a fundamental challenge in machine learning. For decades, the dominant paradigm has relied on dataset-specific parameter estimation, a non-scalable method that fails to transfer knowledge. Recent efforts toward universal aggregation models do not account for the structural and behavioral complexities of human-annotated crowdsourcing, resulting in poor real-world performance. To address this gap, we introduce CrowdFM, a foundation model for crowdsourced label aggregation. At its core, CrowdFM is a bipartite graph neural network that is pre-trained on a vast, domain-randomized synthetic dataset. By leveraging a size-invariant initialization and attention-based message passing, it learns universal principles of collective intelligence and generalizes to new, unseen datasets. Extensive experiments on 22 real-world benchmarks show that our single, fixed model consistently matches or surpasses bespoke, per-dataset methods in both accuracy and efficiency. Furthermore, the representations learned by CrowdFM readily support diverse downstream applications, such as worker assessment and task assignment. Code and pre-trained models will be made publicly available upon acceptance.

Large Vision-Language Models (LVLMs), such as GPT-4o and LLaVA, have recently witnessed remarkable advancements and are increasingly being deployed in real-world applications. However, inheriting the sensitivity of visual neural networks, LVLMs remain vulnerable to adversarial attacks, which can result in erroneous or malicious outputs. While existing efforts utilize adversarial fine-tuning to enhance robustness, they often suffer from significant performance degradation on clean inputs. In this paper, we propose AdPO, a novel adversarial defense strategy for LVLMs based on preference optimization. For the first time, we reframe adversarial training as a preference optimization problem, aiming to enhance the model’s preference for generating normal outputs on clean inputs while rejecting the potential misleading outputs for adversarial examples. Notably, AdPO achieves this by solely modifying the image encoder, e.g., CLIP ViT, resulting in superior clean and adversarial performance in a variety of downstream tasks. Due to the computational cost of training large language models, we show that training on smaller LVLMs and transferring to larger ones achieves state-of-the-art performance with efficiency comparable to previous methods. Our comprehensive experiments confirm the effectiveness of the proposed AdPO which highlights the potential of preference-based learning in adversarially robust multimodal systems.

Writing competitive programming problems is exacting. Authors must: set constraints, input distributions, and edge cases that rule out shortcuts; target specific algorithms (e.g., max-flow, dynamic programming, data structures); and calibrate complexity beyond the reach of most competitors. We argue that this makes for an ideal test of general large language model capabilities and study whether they can do this reliably. We introduce AutoCode, which uses multiple rounds of validation to yield competition-grade problem statements and test cases. On held-out problems, AutoCode test suites approach 99% consistency with official judgments, a significant improvement over current state-of-the-art methods like HardTests, which achieve less than 81%. Furthermore, starting with a random seed problem, AutoCode can create novel variants with reference and brute-force solutions. By cross-verifying these generated solutions against test cases, we can further filter out malformed problems. Our system ensures high correctness, as verified by human experts. AutoCode successfully produces novel problems judged by Grandmaster-level (top 0.3%) competitive programmers to be of contest quality.

Modality alignment is critical for vision-language models (VLMs) to effectively integrate information across modalities. However, existing methods extract hierarchical features from text while representing each image with a single feature, leading to asymmetric and suboptimal alignment. To address this, we propose Alignment across Trees, a method that constructs and aligns tree-like hierarchical features for both image and text modalities. Specifically, we introduce a semantic-aware visual feature extraction framework that applies a cross-attention mechanism to visual class tokens from intermediate Transformer layers, guided by textual cues to extract visual features with coarse-to-fine semantics. We then embed the feature trees of the two modalities into hyperbolic manifolds with distinct curvatures to effectively model their hierarchical structures. To align across the heterogeneous hyperbolic manifolds with different curvatures, we formulate a KL distance measure between distributions on heterogeneous manifolds, and learn an intermediate manifold for manifold alignment by minimizing the distance. We prove the existence and uniqueness of the optimal intermediate manifold. Experiments on taxonomic open-set classification tasks across multiple image datasets demonstrate that our method consistently outperforms strong baselines under few-shot and cross-domain settings.

Large Language Models (LLMs) reproduce social biases, yet prevailing evaluations score models in isolation, obscuring how biases persist across families and releases. We introduce Bias Similarity Measurement (BSM), which treats fairness as a relational property between models, unifying scalar, distributional, behavioral, and representational signals into a single similarity space. Evaluating 30 LLMs on 1M+ prompts, we find that instruction tuning primarily enforces abstention rather than altering internal representations; small models gain little accuracy and can become less fair under forced choice; and open-weight models can match or exceed proprietary systems. Family signatures diverge: Gemma favors refusal, LLaMA 3.1 approaches neutrality with fewer refusals, and converges toward abstention-heavy behavior overall. Counterintuitively, Gemma 3 Instruct matches GPT-4–level fairness at far lower cost, whereas Gemini’s heavy abstention suppresses utility. Beyond these findings, BSM offers an auditing workflow for procurement, regression testing, and lineage screening, and extends naturally to code and multilingual settings. Our results reframe fairness not as isolated scores but as comparative bias similarity, enabling systematic auditing of LLM ecosystems. Code is available at https://anonymous.4open.science/r/bias_llm-0A8E.

The rapid advancement of Multimodal Large Language Models (MLLMs) has made aligning them with human preferences a critical challenge. Reward Models (RMs) are a core technology for achieving this goal, but a systematic guide for building state-of-the-art Multimodal Reward Models (MRMs) is currently lacking in both academia and industry. Through exhaustive experimental analysis, this paper aims to provide a clear “recipe” for constructing high-performance MRMs. We systematically investigate every crucial component in the MRM development pipeline, including reward modeling paradigms (e.g., Naive-RM, Critic-based RM, and Generative RM), reward head architecture, training strategies, data curation (covering over ten multimodal and text-only preference datasets), backbone model and model scale, and ensemble methods. Based on these experimental insights, we introduce BaseReward, a powerful and efficient baseline for multimodal reward modeling. BaseReward adopts a simple yet effective architecture, built upon a Qwen2.5-VL backbone, featuring an optimized two-layer reward head, and is trained on a carefully curated mixture of high-quality multimodal and text-only preference data. Our results show that BaseReward establishes a new state-of-the-art (SOTA) on major benchmarks such as MM-RLHF-Reward Bench, VL-Reward Bench, and Multimodal Reward Bench, outperforming previous open-source and proprietary models. Furthermore, to validate its practical utility beyond static benchmarks, we integrate BaseReward into a real-world reinforcement learning pipeline, successfully enhancing an MLLM’s performance across various perception, reasoning, and conversational tasks. This work not only delivers a top-tier MRM but, more importantly, provides the community with a clear, empirically backed guide for developing robust reward models for the next generation of MLLMs.

Recent work has demonstrated that controlled pretraining experiments are a powerful tool for understanding learning, reasoning, and memorization in large language models (LLMs). However, the computational cost of pretraining presents a significant constraint. To overcome this constraint, we propose to conduct multiple pretraining experiments simultaneously during a *single* training run. We demonstrate the feasibility of this approach by conducting ten experiments during the training of a 1.5B parameter model on 210B tokens. Although we only train a single model, we can replicate the results from multiple previous works on data contamination, poisoning, and memorization. We also conduct novel investigations into knowledge acquisition, mathematical reasoning, and watermarking. For example, we dynamically update the training data until the model acquires a particular piece of knowledge. Remarkably, the influence of the ten experiments on the model's training dynamics and overall performance is minimal. However, interactions between different experiments may act as a potential confounder in our approach. We propose to test for interactions with continual pretraining experiments, finding them to be negligible in our setup. Overall, our findings suggest that performing multiple pretraining experiments in a single training run can enable rigorous scientific experimentation with large models on a compute budget.

Descriptive Multimodal Emotion Recognition (DMER) has garnered increasing research attention. Unlike traditional discriminative paradigms that rely on predefined emotion taxonomies, DMER aims to describe human emotional state using free-form natural language, enabling finer-grained and more interpretable emotion representations. However, this free-form prediction paradigm introduces new challenges regarding its evaluation. Previous works depend on ground-truth descriptions, but emotions are inherently tied to diverse human behaviors, and generating a comprehensive and accurate description is inherently demanding. Other researchers reformulate this problem into a more tractable human preference learning task, but pairwise preference annotation involves substantial manual effort. This leads to a question: *can we leverage multimodal LLMs (MLLMs) to achieve more cost-efficient preference annotation?* To answer this, we propose **EmoPrefer**, a pioneering work exploring the potential of LLMs in decoding human emotion preferences. Specifically, we construct the first emotion preference dataset, **EmoPrefer-Data**, featuring high-quality preference annotations from experts. Additionally, we introduce **EmoPrefer-Bench**, which evaluates the performance of various MLLMs and prompting techniques in preference prediction, while also revealing new strategies to enhance their performance. To the best of our knowledge, this is the first work exploring the capabilities of LLMs in understanding human emotion preferences. Our work advances the field of DMER and lays the foundation for more intelligent human-computer interaction. Our data and code are released at https://github.com/zeroQiaoba/AffectGPT/tree/master/EmoPrefer.

Reasoning distillation, a cost-effective approach for enhancing student model performance, has attracted increasing attention. It typically leverages a large teacher model to generate reasoning paths, which are then used to fine-tune a student model so that it mimics the teacher's behavior in training contexts. However, previous approaches have lacked a detailed analysis of the origins of the distilled model's capabilities. It remains unclear whether the student can maintain consistent behaviors with the teacher in novel test-time contexts, or whether it regresses to its original output patterns, raising concerns about the generalization of distillation models. To analyse this question, we introduce a cross-model Reasoning Distillation Provenance Tracing framework. For each action (e.g., a sentence) produced by the distilled model, we obtain the predictive probabilities assigned by the teacher, the original student, and the distilled model under the same context. By comparing these probabilities, we classify each action into four categories: (i) teacher-originated actions, (ii) student-originated actions, (iii) pre-existing actions in both models not enhanced by distillation, and (iv) pre-existing actions boosted through distillation. By systematically disentangling the provenance of each action, we experimentally demonstrate that, in test-time contexts, the distilled model can indeed generate teacher-originated actions, which correlate with and plausibly explain observed performance on distilled model. Building on this analysis, we further propose a teacher-guided data selection method. Unlike prior approach that rely on heuristics (e.g., selecting data most aligned with the student's original distribution), our method directly compares teacher–student divergences on the training data, providing a principled selection criterion. We validate the effectiveness of our approach across multiple representative teacher models (Deepseek-R1-671B, QwQ-32B, GPT-OSS-120B) and diverse student models (Qwen2.5-7B-Instruct, Qwen4-4B-Base, Qwen3-8B-Base, Qwen3-4B-Instruct-2507). The results highlight the utility of our provenance-tracing framework and underscore its promise for reasoning distillation. We hope to share Reasoning Distillation Provenance Tracing, along with our insights into reasoning distillation, with the community.

Token-level attention tuning -- a class of training-free methods including Post-hoc Attention Steering (PASTA) and Attention Calibration (ACT) -- has emerged as a promising approach for improving frozen LLMs via interpretable interventions. However, these methods rely on auxiliary heuristics to identify important task-specific tokens, which can introduce bias and limit applicability when token importance is ambiguous or when optimized kernels make attention maps inaccessible. We propose a simpler alternative: intervening only on the initial token (e.g., <BOS> in LLaMA). We theoretically show that adding lightweight biases to this token’s attention logits systematically shifts and reshapes downstream attention patterns -- an effect amplified by its natural role as an attention sink. Empirically, we find that this tuning can improve LLM performance and better elicit pretrained knowledge, with stronger effects in early layers and distinct scaling preferences across attention heads. Building on these findings, we introduce ZeroTuning, a training-free method that improves LLM performance by applying head-specific attention adjustments to the initial token, requiring no parameter updates. We present two variants: a supervised mode that calibrates on validation examples, and an unsupervised mode that directly minimizes output entropy. ZeroTuning requires no KV-cache or decoding changes and is kernel-agnostic (works with SDPA and FlashAttention). It requires only four lines of modification to standard \texttt{LlamaAttention} code, achieves gains across 15 datasets, and outperforms prior, more complex methods. For example, on Llama-3.1-8B, it yields relative improvements of 19.9% on classification, 4.5% on question answering, and 2.1% on dialogue. ZeroTuning also works out of the box with quantized inference and maintains its improvements as context length increases. Our work provides a lightweight tool for inference-time improvement, advancing both optimization and interpretability. Our code and runnable demo are available at https://anonymous.4open.science/r/ZeroTuning.

We present MetaSpatial, the first reinforcement learning (RL) framework for enhancing 3D spatial reasoning in vision-language models (VLMs), enabling real-time 3D scene layout generation without post-processing. MetaSpatial addresses two key challenges: (i) the need for extensive post-processing, as existing VLMs lack inherent 3D spatial reasoning to generate realistic layouts; and (ii) the inefficiency of supervised fine-tuning (SFT) for layout generation due to scarcity of perfect annotations. Our core contribution is the 3D Spatial Policy Optimization (3D-SPO) algorithm, which incorporates physics-aware modulation into advantage estimates at the object level and trajectory-level reward from a training-only multi-turn refinement pipeline. This design enhances temporal credit assignment and encourages spatially consistent policy learning. Empirical evaluations across models of varying scales demonstrate that MetaSpatial improves spatial coherence, physical plausibility, and formatting stability, leading to more realistic and functionally coherent object placements applicable to metaverse environments.