ICLR 2026 论文列表

Reinforcement learning based post-training has recently emerged as a powerful paradigm for enhancing the alignment and reasoning capabilities of multimodal large language models (MLLMs). While *vision-centric* post-training is crucial for enhancing MLLMs’ intrinsic understanding of visual signals, current post-training paradigms are predominantly *text-centric*, where dense visual inputs are only leveraged to extract sparse cues for text-based reasoning. There exist a few approaches in this direction, however, they often still rely on text as an intermediate mediator or introduce additional visual generative designs. In this work, we introduce **Visual Jigsaw**, a generic *self-supervised* post-training framework designed to strengthen visual understanding in MLLMs. Visual Jigsaw is formulated as a general ordering task: visual inputs are partitioned, shuffled, and the model must reconstruct the visual information by producing the correct permutation in natural language. This naturally aligns with reinforcement learning from verifiable rewards (RLVR), requires no additional visual generative components, and derives its supervisory signal automatically without any annotations. We instantiate Visual Jigsaw across three visual modalities, including images, videos, and 3D data. Extensive experiments demonstrate substantial improvements in fine-grained perception, temporal reasoning, and 3D spatial understanding. Our findings highlight the potential of self-supervised vision-centric tasks in post-training MLLMs and aim to inspire further research on vision-centric pretext designs.

Referring Image Segmentation (RIS) aims to segment the object in an image uniquely referred to by a natural language expression. However, RIS training often contains hard-to-align and instance-specific visual signals; optimizing on such pixels injects misleading gradients and drives the model in the wrong direction. By explicitly estimating pixel-level vision–language alignment, the learner can suppress low-alignment regions, concentrate on reliable cues, and acquire more generalizable alignment features. In this paper, we propose Alignment-Aware Masked Learning (AML), a simple yet effective training strategy that quantifies region–referent alignment (PMME) and filters out unreliable pixels during optimization (AFM). Specifically, each sample first computes a similarity map between visual and textual features, and then masks out pixels falling below an adaptive similarity threshold, thereby excluding poorly aligned regions from the training process. AML does not require architectural changes and incurs no inference overhead, directing attention to the areas aligned with the textual description. Experiments on the RefCOCO (vanilla/+/g) datasets show that AML achieves state-of-the-art results across all 8 splits, and beyond improving RIS performance, AML also enhances the model’s robustness to diverse descriptions and scenarios.

In this paper, we propose a rotation-constrained compensation method to address the errors introduced by structured pruning of large language models (LLMs). LLMs are trained on massive datasets and accumulate rich semantic knowledge in their representation space. In contrast, pruning is typically carried out with only a small amount of calibration data, which makes output mismatches unavoidable. Although direct least-squares fitting can reduce such errors, it tends to overfit to the limited calibration set, destructively modifying pretrained weights. To overcome this difficulty, we update the pruned parameters under a rotation constraint. This constrained update preserves the geometry of output representations (i.e., norms and inner products) and simultaneously re-aligns the pruned subspace with the original outputs. Furthermore, in rotation-constrained compensation, removing components that strongly contribute to the principal directions of the output makes error recovery difficult. Since input dimensions with large variance strongly affect these principal directions, we design a variance-aware importance score that ensures such dimensions are preferentially kept in the pruned model. By combining this scoring rule with rotation-constrained updates, the proposed method effectively compensates errors while retaining the components likely to be more important in a geometry-preserving manner. In the experiments, we apply the proposed method to Llama-7B and Llama-2-13B, and evaluate it on WikiText-2 and multiple language understanding benchmarks. The results demonstrate consistently better perplexity and task accuracy compared with existing baselines. Codes are available at \url{https://anonymous.4open.science/r/anonymous-llm-pruning-D884/}.

Vision-language models trained on large-scale multimodal datasets show strong demographic biases, but the role of training data in producing these biases remains unclear. A major barrier has been the lack of demographic annotations in web-scale datasets such as LAION-400M. We address this gap by creating person-centric annotations for the full dataset, including over 276 million bounding boxes, perceived gender and race/ethnicity labels, and automatically generated captions. These annotations are produced through validated automatic labeling pipelines combining object detection, multimodal captioning, and finetuned classifiers. Using them, we uncover demographic imbalances and harmful associations, such as the disproportionate linking of men and individuals perceived as Black or Middle Eastern with crime-related and negative content. We also show that 60-70\% of gender bias in CLIP and Stable Diffusion can be linearly explained by direct co-occurrences in the data. Our resources establish the first large-scale empirical link between dataset composition and downstream model bias.

Recent work shows that, beyond discrete reasoning through explicit chain-of-thought steps, which are limited by the boundaries of natural languages, large language models (LLMs) can also reason continuously in latent space, allowing richer information per step and thereby improving token efficiency. Despite this promise, latent reasoning still faces two challenges, especially in training-free settings: 1) purely latent reasoning broadens the search distribution by maintaining multiple implicit paths, which diffuses probability mass, introduces noise, and impedes convergence to a single high-confidence solution, thereby hurting accuracy; and 2) overthinking persists even without explicit text, wasting tokens and degrading efficiency. To address these issues, we introduce SwiReasoning, a training-free framework for LLM reasoning which features two key innovations: 1) SwiReasoning dynamically switches between explicit and latent reasoning, guided by block-wise confidence estimated from entropy trends in next-token distributions, to balance exploration and exploitation and promote timely convergence. 2) By limiting the maximum number of thinking-block switches, SwiReasoning curbs overthinking and improves token efficiency across varying problem difficulties. On widely used mathematics and STEM benchmarks, SwiReasoning consistently improves average accuracy by 1.5%–2.8% across reasoning LLMs of different model families and scales. Furthermore, under constrained budgets, SwiReasoning improves average token efficiency by 56%-79%, with larger gains as budgets tighten. The code will be released publicly.

Recent methods have made notable progress in accelerating Large Vision-Language Models (LVLMs) by exploiting the inherent redundancy in visual inputs. Most existing approaches, however, focus narrowly on reducing image tokens before or within the Large Language Model (LLM) stage to lower computational cost. This overlooks other major bottlenecks, particularly the image encoder, which itself requires substantial computation. As a result, these methods fall short of achieving true end-to-end acceleration. Importantly, the image encoder is the primary contributor of input tokens to the LLM. Thus, reducing visual redundancy at the encoder stage not only speeds up the encoder itself but also significantly lightens the workload for the subsequent LLM. Motivated by this, we investigate how to jointly optimize the image encoder and the LLM along with other LVLM components for comprehensive acceleration. To mitigate the risk of performance degradation from token reduction, we propose a novel token merging strategy that recycles useful information from otherwise discarded tokens. Our approach, iLLaVA, delivers consistent improvements across both image and video understanding tasks, achieving up to a 2× throughput boost and a 4× reduction in prefilling time. Notably, iLLaVA enables a larger model (e.g., InternVL-2.5 26B) to surpass a smaller counterpart (e.g., InternVL-2.5 8B) in both accuracy and efficiency. Extensive comparisons with state-of-the-art token pruning and merging techniques demonstrate the clear superiority of our method. Finally, we provide detailed visualizations for the merging steps of iLLaVA , offering deeper insights into how different LVLM components contribute to efficient computation.

Recent studies have investigated attention dynamics in large vision language models (LVLMs), yet existing methods remain limited in reliably distinguishing hallucinated from correct outputs — primarily because they rely solely on forward-pass attention, ignoring gradient-based signals that reveal how token influence propagates through the model. To bridge this gap, we introduce \textbf{LVLMs-Saliency}, an \textit{gradient-aware diagnostic tool} that quantifies the grounding strength of each output token by fusing attention weights with their gradients. Through analysis, we identify a decisive pattern: \textit{Hallucinations occur when prior output tokens shows low saliency to the next token prediction}, indicating a failure of contextual memory. Building on this insight, we propose a dual-mechanism inference-time framework: (1) Saliency-Guided Rejection Sampling (SGRS), which dynamically filters candidate tokens during decoding by rejecting those with saliency below a context-adaptive threshold, thereby preventing coherence-breaking tokens from entering the sequence; and (2) Local Coherence Reinforcement (LocoRE), a lightweight plug-and-play module that strengthens attention from the current token to its most recent outputs, actively counteracting the “forgetting” behavior identified by LVLMs-Saliency. Experimental results demonstrate that our method significantly reduces hallucinations across multiple LVLMs, offering a robust and interpretable solution to improve model reliability.

Real-world visual signals are inherently variable in resolution, and it is natural to endow multimodal large language models (MLLMs) with such native-resolution perception capabilities. In principle, for general and straightforward multimodal understanding, low-resolution images are sufficient. While for images with nuanced details like documents and charts, it is crucial to preserve fine-grained details using high-resolution inputs, as naive resizing inevitably results in information loss. Recent advances employ sequence packing to process images of any resolution and aspect ratios. Despite these efforts, model performance degrades at both low and high resolutions, and high-resolution inputs incur substantial computational costs. We argue that the rigid use of a single patch size is the primary cause: when image resolution or information density varies, fixing patch size is intrinsically suboptimal. To address this issue, we introduce Adaptive Patching (AdaPatch), a simple yet effective strategy that adjusts patch size according to image resolution and information density and could be seamlessly plugged into pre-trained fixed-patch MLLMs without any training efforts. Extensive evaluations demonstrate consistent improvements in native resolution performance without additional training. Besides, we provide a training-based method to further adapt MLLMs with dynamic patch sizes and enhance the performance.

Despite recent advances, multimodal large language models continue to struggle with visual mathematical problem solving. Some recent works recognize that visual perception is a bottleneck in visual mathematical reasoning, but their solutions are limited to improving the extraction and interpretation of visual inputs. Notably, they all ignore the key issue of whether the extracted visual cues are faithfully integrated and properly utilized in subsequent reasoning. Motivated by this, we present CogFlow, a novel cognitive-inspired three-stage framework that incorporates a knowledge internalization stage, explicitly simulating the hierarchical flow of human reasoning: perception$\Rightarrow$internalization$\Rightarrow$reasoning. In line with this hierarchical flow, we holistically enhance all its stages. We devise synergistic visual rewards to boost perception capabilities in parametric and semantic spaces, jointly improving visual information extraction from symbols and diagrams. To guarantee faithful integration of extracted visual cues into subsequent reasoning, we introduce a visual-anchored reward model in the internalization stage, bridging perception and reasoning. Moreover, we design a visual-gated policy optimization algorithm to further enforce the reasoning is grounded with the visual knowledge, preventing models seeking shortcuts that appear coherent but are visually ungrounded reasoning chains. Moreover, we contribute a new dataset MathCog for model training, which contains samples with over 120K high-quality perception-reasoning aligned annotations. Comprehensive experiments and analysis on three commonly used visual mathematical reasoning benchmarks validate the superiority of the proposed CogFlow. Our data, code and models will be publicly available.

As hubs of human activity, urban surfaces consist of a wealth of semantic entities. Segmenting these various entities from satellite imagery is crucial for a range of downstream applications. Current advanced segmentation models can reliably segment entities defined by physical attributes (e.g., buildings, water bodies) but still struggle with socially defined categories (e.g., schools, parks). In this work, we achieve socio-semantic segmentation by vision-language model reasoning. To facilitate this, we introduce the Urban Socio-Semantic Segmentation dataset named \textbf{SocioSeg}, a new resource comprising satellite imagery, digital maps, and pixel-level labels of social semantic entities organized in a hierarchical structure. Additionally, we propose a novel vision-language reasoning framework called \textbf{SocioReasoner} that simulates the human process of identifying and annotating social semantic entities via cross-modal recognition and multi-stage reasoning. We employ reinforcement learning to optimize this non-differentiable process and elicit the reasoning capabilities of the vision-language model. Experiments demonstrate our approach's significant gains over state-of-the-art models and strong zero-shot generalization. The dataset and code will be made publicly available.

Large Language Models (LLMs) achieve strong performance on diverse tasks but often exhibit cognitive inertia, struggling to follow instructions that conflict with the standardized patterns learned during supervised fine-tuning (SFT). To evaluate this limitation, we propose Inverse IFEval, a benchmark that measures models’ Counter-intuitive Ability—their capacity to override training-induced biases and comply with adversarial instructions. Inverse IFEval introduces eight types of such challenges, including Question Correction, Intentional Textual Flaws, Code without Comments, and Counterfactual Answering. Using a human-in-the-loop pipeline, we construct a dataset of 1012 high-quality Chinese and English questions across 23 domains, evaluated under an optimized LLM-as-a-Judge framework. Experiments on existing leading LLMs demonstrate the necessity of our proposed Inverse IFEval benchmark. Our findings emphasize that future alignment efforts should not only pursue fluency and factual correctness but also account for adaptability under unconventional contexts. We hope that Inverse IFEval serves as both a diagnostic tool and a foundation for developing methods that mitigate cognitive inertia, reduce overfitting to narrow patterns, and ultimately enhance the instruction-following reliability of LLMs in diverse and unpredictable real-world scenarios.

Large Vision Language Models (LVLMs) have recently emerged as powerful architectures capable of understanding and reasoning over both visual and textual information. These models typically rely on two key components: a Vision Transformer (ViT) and a Large Language Model (LLM). ViT encodes visual content into a sequence of image tokens and serves as the perceptual front-end -- the eyes of the model. In contrast, the LLM interprets these tokens to perform high-level reasoning, generates responses, and functions as the cognitive core -- the brain of the model. However, it remains unclear which visual tokens contribute most significantly to understanding and reasoning, and how effectively these signals are propagated from ViT to the LLM. While most existing works have focused on identifying attention sinks, low-semantic tokens receiving disproportionately high attention, within the LLM, we shift the focus to the vision encoder by identifying a class of high-norm visual tokens from ViT, referred to as ViT attention sinks -- a problem that has been rarely studied but is indeed very important for LVLMs. Our findings show that these ViT sinks encapsulate high-level semantic concepts from images, allowing the LLM to perform more effective understanding and reasoning. Despite their importance, these sink tokens are often overlooked in existing LVLM architectures. To explore their contribution, we present both qualitative and quantitative analyses of the information embedded in these sink tokens. We also propose both training-free and training-based approaches to better leverage how this information is interpreted by the LLM, and to what extent. By explicitly utilizing these tokens, we demonstrate substantial improvements across a range of LVLMs and visual reasoning tasks, including but not limited to mathematical problem solving, logical inference, and geometric understanding, highlighting the untapped potential of ViT attention sinks in enhancing visual reasoning.

Recent studies on 3D hand reconstruction have demonstrated the effectiveness of synthetic training data to improve estimation performance. However, most methods rely on game engines to synthesize hand images, which often lack diversity in textures and environments, and fail to include crucial components like arms or interacting objects. Generative models are promising alternatives to generate diverse hand images, but still suffer from misalignment issues. In this paper, we present SesaHand, which enhances controllable hand image generation from both semantic and structural alignment perspectives for 3D hand reconstruction. Specifically, for semantic alignment, we propose a pipeline with Chain-of-Thought inference to extract human behavior semantics from image captions generated by the Vision-Language Model. This semantics suppresses human-irrelevant environmental details and ensures sufficient human-centric contexts for hand image generation. For structural alignment, we introduce hierarchical structural fusion to integrate structural information with different granularity for feature refinement to better align the hand and the overall human body in generated images. We further propose a hand structure attention enhancement method to efficiently enhance the model's attention on hand regions. Experiments demonstrate that our method not only outperforms prior work in generation performance but also improves 3D hand reconstruction with the generated hand images.

As AI becomes more capable, we entrust it with more general and consequential tasks. The risks from failure grow more severe with increasing task scope. It is therefore important to understand the ways extremely capable AI models will fail: Will they fail by systematically pursuing goals we do not intend? Or will they fail by being a hot mess, and taking nonsensical actions that do not further any goal? We operationalize this question using a bias-variance decomposition of the errors made by AI models: An AI's *incoherence* on a task is measured over test-time randomness as the fraction of its error that stems from variance rather than bias in task outcome. Across all tasks and frontier models we measure, we find that the longer models spend reasoning and taking actions, *the more incoherent* they become. We observe that incoherence changes with model scale in a way that is task and experiment dependent. However, in several settings larger, more capable models are more incoherent than smaller models. Consequently, scale alone seems unlikely to eliminate incoherence. Instead, as more capable AIs pursue harder tasks, requiring more sequential action and thought, our results predict failures to be accompanied by more incoherent behavior. This suggests a future where AIs sometimes cause industrial accidents (due to unpredictable misbehavior), but are less likely to exhibit consistent pursuit of a misaligned goal. This increases the relative importance of alignment research targeting reward hacking or goal misspecification.

Spine disorders affect 619 million people globally and are a leading cause of disability, yet AI-assisted diagnosis remains limited by the lack of level-aware, multimodal datasets. Clinical decision-making for spine disorders requires sophisticated reasoning across X-ray, CT, and MRI at specific vertebral levels. However, progress has been constrained by the absence of traceable, clinically-grounded instruction data and standardized, spine-specific benchmarks. To address this, we introduce SpineMed, an ecosystem co-designed with practicing spine surgeons. It features SpineMed-450k, the first large-scale dataset explicitly designed for vertebral-level reasoning across imaging modalities with over 450,000 instruction instances, and SpineBench, a clinically-grounded evaluation framework. SpineMed-450k is curated from diverse sources, including textbooks, guidelines, open datasets, and $\sim$1,000 de-identified hospital cases, using a clinician-in-the-loop pipeline with a two-stage LLM generation method (draft and revision) to ensure high-quality, traceable data for question-answering, multi-turn consultations, and report generation. SpineBench evaluates models on clinically salient axes, including level identification, pathology assessment, and surgical planning. Our comprehensive evaluation of several recently advanced large vision-language models (LVLMs) on SpineBench reveals systematic weaknesses in fine-grained, level-specific reasoning. In contrast, our model fine-tuned on SpineMed-450k demonstrates consistent and significant improvements across all tasks. Clinician assessments confirm the diagnostic clarity and practical utility of our model's outputs.

Text-to-video and image-to-video generation have made rapid progress in visual quality, but they remain limited in controlling the precise timing of motion. In contrast, audio provides temporal cues aligned with video motion, making it a promising condition for temporally controlled video generation. However, existing audio-to-video (A2V) models struggle with fine-grained synchronization due to indirect conditioning mechanisms or limited temporal modeling capacity. We present Syncphony, which generates 380×640 resolution, 24fps videos synchronized with diverse audio inputs. Our approach builds upon a pre-trained video backbone and incorporates two key components to improve synchronization: (1) Motion-aware Loss, which emphasizes learning at high-motion regions; (2) Audio Sync Guidance, which guides the full model using a visually aligned off-sync model without audio layers to better exploit audio cues at inference while maintaining visual quality. To evaluate synchronization, we propose CycleSync, a video-to-audio-based metric that measures the amount of motion cues in the generated video to reconstruct the original audio. Experiments on AVSync15 and The Greatest Hits datasets demonstrate that Syncphony outperforms existing methods in both synchronization accuracy and visual quality.

Fine-tuning vision-language models (VLMs) on robot teleoperation data to create vision-language-action (VLA) models is a promising paradigm for training generalist policies, but it suffers from a fundamental tradeoff: learning to produce actions often diminishes the VLM's foundational reasoning and multimodal understanding, hindering generalization to novel scenarios, instruction following, and semantic understanding. We argue that this catastrophic forgetting is due to a distribution mismatch between the VLM's internet-scale pretraining corpus and the robotics fine-tuning data. Inspired by this observation, we introduce VLM2VLA: a VLA training paradigm that first resolves this mismatch at the data level by representing low-level actions with natural language. This alignment makes it possible to train VLAs solely with Low-Rank Adaptation (LoRA), thereby minimally modifying the VLM backbone and averting catastrophic forgetting. As a result, the VLM can be fine-tuned on robot teleoperation data without fundamentally altering the underlying architecture and without expensive co-training on internet-scale VLM datasets. Through extensive Visual Question Answering (VQA) studies and over 800 real-world robotics experiments, we demonstrate that VLM2VLA preserves the VLM's core capabilities, enabling zero-shot generalization to novel tasks that require open-world semantic reasoning and multilingual instruction following.

Multimodal position encoding is essential for vision-language models, yet there has been little systematic investigation into multimodal position encoding. We conduct a comprehensive analysis of multimodal Rotary Positional Embedding (RoPE) by examining its two core components: position design and frequency allocation. Through extensive experiments, we identify three key guidelines: positional coherence, full frequency utilization, and preservation of textual priors—ensuring unambiguous layout, rich representation, and faithful transfer from the pre-trained LLM. Based on these insights, we propose Multi-Head RoPE (MHRoPE) and MRoPE-Interleave (MRoPE-I), two simple and plug-and-play variants that require no architectural changes. Our methods consistently outperform existing approaches across diverse benchmarks, with significant improvements in both general and fine-grained multimodal understanding.

Real-world settings where language models (LMs) are deployed---in domains spanning healthcare, finance, and other forms of knowledge work—require models to grapple with incomplete information and reason under uncertainty. Yet most LM evaluations focus on problems with well-defined answers and success criteria. This gap exists in part because natural problems involving uncertainty are difficult to construct: given that LMs have access to most of the same knowledge as humans, it is non-trivial to design questions for which LMs will struggle to produce correct answers, but which humans can answer reliably.As a result, LM performance on reasoning under uncertainty remains poorly characterized. To address this gap, we introduce \textsc{OpenEstimate}, an extensible, multi-domain benchmark for evaluating LMs on numerical estimation tasks that require models to synthesize significant amounts of background information and express predictions as probabilistic priors. We assess these priors for accuracy and calibration. Across six frontier models, we find that LM-elicited priors are worth the equivalent of about five samples from the underlying data distribution, and that posteriors computed using LM priors tend to be more accurate than those computed using a naive prior. At the same time, the relationship between model accuracy and confidence is weak across the board, indicating the value of developing new methods to improve calibration. The \textsc{OpenEstimate} benchmark thus offers a challenging evaluation for frontier LMs and a platform for developing models that are better at probabilistic estimation and reasoning under uncertainty.