ICLR 2026 论文列表

Winning competitive debates requires sophisticated reasoning and argument skills. There are unique challenges in the competitive debate: (1) The time constraints force debaters to make strategic choices about which points to pursue rather than covering all possible arguments; (2) The persuasiveness of the debate relies on the back-and-forth interaction between arguments, which a single final game status cannot evaluate. To address these challenges, we propose TreeDebater, a novel debate framework that excels in competitive debate. We introduce two tree structures: the Rehearsal Tree and Debate Flow Tree. The Rehearsal Tree anticipates the attack and defenses to evaluate the strength of the claim, while the Debate Flow Tree tracks the debate status to identify the active actions. TreeDebater allocates its time budget among candidate actions and uses the speech time controller and feedback from the simulated audience to revise its statement. The human evaluation on both the stage-level and the debate-level comparison shows that our TreeDebater outperforms the state-of-the-art multi-agent debate system, with a +15.6% improvement in stage-level persuasiveness with DeepSeek and +10% debate-level opinion shift win. Further investigation shows that TreeDebater shows better strategies in limiting time to important debate actions, aligning with the strategies of human debate experts.

Vision-Language Models (VLMs) have recently demonstrated remarkable capabilities in visual understanding and reasoning, but they also impose significant computational burdens due to long visual sequence inputs. Recent works address this issue by pruning unimportant visual tokens, achieving substantial computational reduction while maintaining model performance. The core of token pruning lies in determining token importance, with current approaches primarily relying on attention scores from vision encoders or Large Language Models (LLMs). In this paper, we analyze the effectiveness of attention mechanisms in both vision encoders and LLMs. We find that vision encoders suffer from attention sink, leading to poor focus on informative foreground regions, while in LLMs, although prior studies have identified attention bias toward token positions, text-to-vision attention demonstrates resistance to this bias and enables effective pruning guidance in middle layers. Based on these observations, we propose $\textbf{LearnPruner}$, a two-stage token pruning framework that first removes redundant vision tokens via a learnable pruning module after the vision encoder, then retains only task-relevant tokens in the LLM's middle layer. Experimental results show that our LearnPruner can preserve approximately 95\% of the original performance while using only 5.5\% of vision tokens, and achieve 3.2$\times$ inference acceleration, demonstrating a superior accuracy-efficiency trade-off.

Causal attention has become a foundational mechanism in autoregressive Vision-Language models (VLMs), unifying textual and visual inputs under a single generative framework. However, existing causal mask-based strategies are inherited from large language models (LLMs) where they are tailored for text-only decoding, and their adaptation to vision tokens is insufficiently addressed in the prefill stage. Strictly masking future positions for vision queries introduces overly rigid constraints, which hinder the model’s ability to leverage future context that often contains essential semantic cues for accurate inference. In this work, we empirically investigate how different causal masking strategies affect vision-language inference and then propose a family of future-aware attentions tailored for this setting. We first empirically analyze the effect of previewing future tokens for vision queries and demonstrate that rigid masking undermines the model’s capacity to capture useful contextual semantic representations. Based on these findings, we propose a lightweight attention family that aggregates future visual context into past representations via pooling, effectively preserving the autoregressive structure while enhancing cross-token dependencies. We evaluate a range of causal masks across diverse vision-language inference settings and show that selectively compressing future semantic context into past representations benefits the inference.

Understanding the negation in a sentence is an important part of compositional understanding and logic in natural language. Many practical AI applications, such as autonomous driving, include precise instruction with negations. For example, following instruction to an AI assistant ”locate a parking spot without a vehicle” requires the assistant to not confuse between presence and absence of vehicles. Al- though joint embedding-based Vision Language Models (VLMs) like CLIP have revolutionized multi-modal tasks, they struggle to interpret negation. To address this limitation, recently many works proposed to solve the problem through a data- centric approach by introducing additional datasets with hard-negative samples for both image and text data. Contrary to these approaches, we present a zero-shot approach to tackle the negation understanding problem. We probe the properties of CLIP text embeddings and show that they follow compositional arithmetic op- erations, which allow the addition or removal of semantic information directly in the embedding space. We then present a rule-based approach to extract negated text from given caption and then use it to explicitly remove corresponding se- mantic information from original embedding, improving negation understanding in VLMs. Our approach does not require expensive training process to induce negation understanding into the model, and achieves the state-of-the-art perfor- mance on popular benchmark for negation understanding. We improve baseline CLIP model performance on NegBench from 25.5% to 67.0% for MCQ and from 50.9% to 56.1% for retrieval tasks. Even NegCLIP model which is fine-tuned on negtion datasets, our approach boosts its MCQ accuracy from 54.03% to 66.22% and retrieval accuracy from 59.25% to 60.1% showing strong performance.

3D Gaussian Splatting has demonstrated remarkable potential in novel view synthesis. In contrast to small-scale scenes, large-scale scenes inevitably contain sparsely observed regions with excessively sparse initial points. In this case, supervising Gaussians initialized from low-frequency sparse points with high-frequency images often induces uncontrolled densification and redundant primitives, degrading both efficiency and quality. Intuitively, this issue can be mitigated with scheduling strategies, which can be categorized into two paradigms: modulating target signal frequency via densification and modulating sampling frequency via image resolution. However, previous scheduling strategies are primarily hardcoded, failing to perceive the convergence behavior of the scene frequency. To address this, we reframe scene reconstruction problem from the perspective of signal structure recovery, and propose SIG, a novel scheduler that Synchronizes Image supervision with Gaussian frequencies. Specifically, we derive the average sampling frequency and bandwidth of 3D representations, and then regulate the training image resolution and the Gaussian densification process based on scene frequency convergence. Furthermore, we introduce Sphere-Constrained Gaussians, which leverage the spatial prior of initialized point clouds to control Gaussian optimization. Our framework enables frequency-consistent, geometry-aware, and floater-free training, achieving state-of-the-art performance with a substantial margin in both efficiency and rendering quality in large-scale scenes.

Advancing machine intelligence requires developing the ability to perceive across multiple modalities, much as humans sense the world. We introduce OmniVinci, an initiative to build a strong, open-source, omni-modal LLM. We carefully study the design choices across model architecture and data curation. For model architecture, we present three key innovations: (i) OmniAlignNet for strengthening alignment between vision and audio embeddings in a shared omni-modal latent space; (ii) Temporal Embedding Grouping for capturing relative temporal alignment between vision and audio signals; and (iii) Constrained Rotary Time Embedding for encoding absolute temporal information in omni-modal embeddings. We introduce a curation and synthesis pipeline that generates 24M single-modal and omni-modal conversations. We find that modalities reinforce one another in both perception and reasoning. Our model, OmniVinci, improves over Qwen2.5-Omni with +19.05 on DailyOmni (cross-modal understanding), +1.7 on MMAR (audio), and +3.9 on Video-MME (vision), while using just 0.2T training tokens — a 6× reduction compared to Qwen2.5-Omni’s 1.2T. We finally demonstrate omni-modal advantages in downstream applications spanning robotics, medical AI, and smart factory.

Multimodal Large Language Models (MLLMs) encode images into visual tokens, aligning visual and textual signals within a shared latent space to facilitate cross-modal representation learning. The CLIP model is a widely adopted foundational vision language model whose vision encoder has played a critical role in the development of MLLMs such as LLaVA. However, the CLIP vision encoder suffers from notable limitations including being constrained to only handling fixed input resolutions and a failure to produce separated embeddings for dissimilar images. Replacing the vision encoder of an existing model typically incurs substantial computational costs because such a change often necessitates retraining the entire model pipeline. In this work, we identify two factors which underlie the limitations of the CLIP vision encoder: mesoscopic bias and interpolation bias. To address these issues, we propose QLIP, a drop-in replacement for CLIP that can be seamlessly integrated with existing MLLMs with only a few lines of code and can enhance both coarse-grained and fine-grained visual understanding, without re-training. QLIP is designed around an image quadtree which replaces the standard uniform grid patches with a novel content aware patchification. Our experimental results demonstrate that QLIP improves the general visual question answering accuracy of the LLaVA-1.5 model series across various model sizes—without requiring retraining or fine-tuning of the full MLLM. Notably, QLIP boosts detailed understanding performance on the challenging $V^*$ benchmark by up to 13.6%.

3D single object tracking (SOT) remains a highly challenging task due to the inherent crux in learning representations from point clouds to effectively capture both spatial shape features and temporal motion features. Most existing methods employ a category-specific optimization paradigm, training the tracking model individually for each object category to enhance tracking performance, albeit at the expense of generalizability across different categories. In this work, we propose a robust category-unified 3D SOT model, referred to as SpatioTemporal Point Relation Parsing model (*PointRePar*), which is capable of joint training across multiple categories while excelling in unified feature learning for both spatial shapes and temporal motions. Specifically, the proposed *PointRePar* captures and parses the latent point relations across both spatial and temporal domains to learn superior shape and motion characteristics for robust tracking. On the one hand, it models the multi-scale spatial point relations using a Mamba-based U-Net architecture with adaptive point-wise feature refinement. On the other hand, it captures both the point-level and box-level temporal relations to exploit the latent motion features. Extensive experiments across three benchmarks demonstrate that our *PointRePar* not only outperforms the existing category-unified 3D SOT methods significantly, but also compares favorably against the state-of-the-art category-specific methods. Codes will be released.

Large Audio Language Models (LALMs) integrate the audio modality directly into the model, rather than converting speech into text and inputting text to Large Language Models (LLMs). While jailbreak attacks on LLMs have been extensively studied, the security of LALMs with audio modalities remains largely unexplored. Currently, there is a lack of an adversarial audio dataset and a unified framework specifically designed to evaluate and compare attacks and LALMs. In this paper, we present JALMBench, a comprehensive benchmark to assess the safety of LALMs against jailbreak attacks. JALMBench includes a dataset containing 11,316 text samples and 245,355 audio samples (>1,000 hours). It supports 12 mainstream LALMs, 4 text-transferred and 4 audio-originated attack methods, and 5 defense methods. Using JALMBench, we provide an in-depth analysis of attack efficiency, topic sensitivity, voice diversity, and architecture. Additionally, we explore mitigation strategies for the attacks at both the prompt level and the response level. We find that LALM safety is strongly influenced by modality and architectural choices, demonstrating that text-based safety alignment can partially transfer to audio inputs and that interleaving audio-text strategy enables more robust cross-modal generalization of safety. Moreover, current general moderation for jailbreak only slightly improves security, and calls on the community to explore more defense methods for LALMs. Our work is the first to systematically uncover these design principles, providing a roadmap for building resilient multimodal language models.

Visual Place Recognition (VPR) requires robust retrieval of geotagged images despite large appearance, viewpoint, and environmental variation. Prior methods focus on descriptor fine-tuning or fixed sampling strategies yet neglect the dynamic interplay between spatial context and visual similarity during training. We present SAGE ($\underline{S}$patial-visual $\underline{A}$daptive $\underline{G}$raph $\underline{E}$xploration), a unified training pipeline that enhances granular spatial–visual discrimination by jointly improving local feature aggregation, organize samples during training, and hard sample mining. We introduce a lightweight Soft Probing module that learns residual weights from training data for patch descriptors before bilinear aggregation, boosting distinctive local cues. During training we reconstruct an online geo–visual graph that fuses geographic proximity and current visual similarity so that candidate neighborhoods reflect the evolving embedding landscape. To concentrate learning on the most informative place neighborhoods, we seed clusters from high-affinity anchors and iteratively expand them with a greedy weighted clique expansion sampler. Implemented with a frozen DINOv2 backbone and parameter-efficient fine-tuning, SAGE achieves SOTA across eight benchmarks. Notably, our method obtains 100\% Recall@10 on SPED only using 4096D global descriptors. The code and model are available at https://github.com/chenshunpeng/SAGE.

Mixture-of-Experts (MoE) architectures have demonstrated remarkable capabilities and scalability for large language models, but incur a prohibitive memory footprint due to duplicated expert parameters. Existing compression approaches, particularly those based on low-rank decomposition, typically operate at the granularity of individual experts. However, such per-expert methods overlook structural redundancies across experts, limiting their compression efficiency and effectiveness. In this work, we introduce TD-MoE (Tensor Decomposition for MoE Compression), a data-aware framework that jointly and holistically factorizes expert weights. Our contributions are threefold: (i) Cross-expert tensorization with joint 3D decomposition, which unifies all experts within a layer into a single tensor and captures shared structure beyond per-expert scope; (ii) A multi-linear whitening strategy, which decorrelates input and output features, yielding a more balanced and data-adaptive decomposition; (iii) A 3D rank allocation mechanism, which dynamically assigns 3D decomposition ranks across dimensions to best meet a target compression ratio while minimizing the reconstruction error. Extensive experiments on Qwen2-57B-A14 and Mixtral-8×7B across seven commonsense reasoning benchmarks demonstrate that TD-MoE achieves almost lossless performance under 20\% parameter reduction, and delivers more than 11\% and 14\% gains over state-of-the-art decomposition-based baselines at 40\% and 60\% compression. Further ablation studies validate the effectiveness of each component, highlighting the importance of joint factorization, whitening, and rank allocation. The code is available at \href{https://anonymous.4open.science/r/TD-MoE}{https://anonymous.4open.science/r/TD-MoE}.

We present IncVGGT, a training-free incremental variant of VGGT that makes transformer-based 3D reconstruction feasible for long sequences in real-world applications. Vanilla VGGT relies on dense global attention, which causes memory to grow quadratically and requires excessive computation, making it impractical for long-sequence scenarios. Even evolved streaming variants, such as StreamVGGT, still suffer from rapidly growing cache and latency. IncVGGT addresses these challenges from two orthogonal directions: (1) register and fuse overlapping frames into composite views, reducing duplicate tokens, and (2) history-side pruning retains only the top-$k$ most relevant/maximum slots together with the most recent one, bounding cache growth. This incremental and memory-efficient design minimizes computation and memory occupation across arbitrarily long sequences. Compared to StreamVGGT, IncVGGT sustains arbitrarily long sequences with large efficiency gains (e.g., on 500-frame sequences, 58.5$\times$ fewer operators, 9$\times$ lower memory, 25.7$\times$ less energy, and 4.9$\times$ faster inference) while maintaining comparable accuracy. More importantly, unlike existing baselines that directly run out of memory beyond 300 (VGGT)–500 (StreamVGGT) frames, IncVGGT continues to operate smoothly even on 10k-frame inputs under an 80GB GPU, showing that our design truly scales to ultra-long sequences without hitting memory limits. These results highlight IncVGGT’s potential for deployment in resource-constrained edge devices for long-range 3D scenarios.

Speech-to-Speech (S2S) models have shown promising dialogue capabilities, but their ability to handle paralinguistic cues—such as emotion, tone, and speaker attributes—and to respond appropriately in both content and style remains underexplored. Progress is further hindered by the scarcity of high-quality and expressive demonstrations. To address this, we introduce a novel reinforcement learning (RL) framework for paralinguistic-aware S2S, ParaS2S, which evaluates and optimizes both content and speaking style directly at the waveform level. We first construct ParaS2SBench, a benchmark comprehensively evaluates S2S models’ output for content and style appropriateness from diverse and challenging input queries. It scores the fitness of input-output pairs and aligns well with human judgments, serving as an automatic judge for model outputs. With this scalable scoring feedback, we enable the model to explore and learn from diverse unlabeled speech via Group Relative Policy Optimization (GRPO). Experiments show that existing S2S models fail to respond appropriately to paralinguistic attributes, performing no better than pipeline-based baselines. Our RL-based strong baseline achieves a 11% relative improvement in response content and style’s appropriateness on ParaS2SBench over supervised fine-tuning (SFT), surpassing all prior models while requiring substantially fewer warm-up annotations than pure SFT.

Attention-based multiple instance learning (MIL) has emerged as a powerful framework for whole slide image (WSI) diagnosis, leveraging attention to aggregate instance-level features into bag-level predictions. Despite this success, we find that such methods exhibit a new failure mode: unstable attention dynamics. Across four representative attention-based MIL methods and two public WSI datasets, we observe that attention distributions oscillate across epochs rather than converging to a consistent pattern, degrading performance. This instability adds to two previously reported challenges: overfitting and over-concentrated attention distribution. To simultaneously overcome these three limitations, we introduce attention-stabilized multiple instance learning (ASMIL), a novel unified framework. ASMIL uses an anchor model to stabilize attention, replaces softmax with a normalized sigmoid function in the anchor to prevent over-concentration, and applies token random dropping to mitigate overfitting. Extensive experiments demonstrate that ASMIL achieves up to a 6.49% F1 score improvement over state-of-the-art methods. Moreover, integrating the anchor model and normalized sigmoid into existing attention-based MIL methods consistently boosts their performance, with F1 score gains up to 10.73%. All code and data are publicly available at https://anonymous.4open.science/r/ASMIL-5018/.

LLM-based search agents are increasingly trained on entity-centric synthetic data to solve complex, knowledge-intensive tasks. However, prevailing training methods like Group Relative Policy Optimization (GRPO) discard this rich entity information, relying instead on sparse, outcome-based rewards. This critical limitation renders them unable to distinguish informative "near-miss" samples—those with substantially correct reasoning but a flawed final answer—from complete failures, thus discarding valuable learning signals. We address this by leveraging the very entities discarded during training. Our empirical analysis reveals a strong positive correlation between the number of ground-truth entities identified during an agent's reasoning process and final answer accuracy. Building on this insight, we introduce Entity-aware Group Relative Policy Optimization (E-GRPO), a novel framework that formulates a dense entity-aware reward function. E-GRPO assigns partial rewards to incorrect samples proportional to their entity match rate, enabling the model to effectively learn from these ''near-misses''. Experiments on diverse question-answering (QA) and deep research benchmarks show that E-GRPO consistently and significantly outperforms the GRPO baseline. Furthermore, our analysis reveals that E-GRPO not only achieves superior accuracy but also induces more efficient reasoning policies that require fewer tool calls, demonstrating a more effective and sample-efficient approach to aligning search agents.

Vision token pruning has proven to be an effective acceleration technique for the Efficient Vision Language Model (VLM). However, existing pruning methods demonstrate excellent performance preservation in visual question answering (VQA) and suffer substantial degradation on visual grounding (VG) tasks. Our analysis of the VLM’s processing pipeline reveals that strategies utilizing global semantic similarity and attention scores lose the global spatial reference frame, which is derived from the interactions of tokens' positional information. Motivated by these findings, we propose N\"uwa, a two-stage token pruning framework that enables efficient feature aggregation while maintaining spatial integrity. In the first stage, after the vision encoder, we apply three operations, namely separation, alignment, and aggregation, which are inspired by swarm intelligence algorithms to retain information-rich global spatial anchors. In the second stage, within the LLM, we perform text-guided pruning to retain task-relevant visual tokens. Extensive experiments demonstrate that N\"uwa achieves state-of-the-art performance on multiple VQA benchmarks (from 94\% to 95\%) and yields substantial improvements on visual grounding tasks (from 7\% to 47\%). Code is released.

Fine-Grained Visual Classification (FGVC) aims to distinguish visually similar subcategories within a broad category, and poses significant challenges due to subtle inter-class differences, large intra-class variations, and data scarcity. Existing methods often struggle to effectively capture both part-level detail and spatial relational features, particularly across rigid and non-rigid object categories. To address these issues, we propose Part-level Semantic-guided Contrastive Learning (PSCL), a novel framework that integrates three key components. (1) The Part Localization Module (PLM) leverages clearCLIP to enable text-controllable region selection, achieving decoupled and semantically guided spatial feature extraction. (2) The Multi-scale Multi-part Branch Progressive Reasoning (MMBPR) module captures discriminative features across multiple parts and scales, while reducing inter-branch redundancy. (3) The Visual-Language Contrastive Learning based on Multi-grained Text Features (VLCL-MG) module introduces intermediate-granularity category concepts to improve feature alignment and inter-class separability. Extensive experiments on five publicly available FGVC datasets demonstrate the superior performance and generalization ability of PSCL, validating the effectiveness of its modular design and the synergy between vision and language. Code is available at: https://anonymous.4open.science/r/PSCL-3E1F.

Building 4D video representations to model underlying spacetime constitutes a crucial step toward understanding dynamic scenes, yet there is no consensus on the paradigm: current approaches resort to additional estimators such as depth, flow, or tracking, or to heavy per-scene optimization, making them brittle and hard to generalize. In a video, its atomic unit, the pixel, follows a continuous 3D trajectory that unfolds over time, acting as the atomic primitive of dynamics. Recognizing this, we propose to represent any video as a Trajectory Field: a dense mapping that assigns each pixel in each frame to a parametric 3D trajectory. To this end, we introduce Trace Anything, a neural network that predicts the trajectory field in a feed-forward manner. Specifically, for each video frame, the model outputs a series of control point maps, defining parametric trajectories for each pixel. Together, our representation and model directly construct a 4D video representation in a single forward pass, without additional estimators or global alignment. We develop a synthetic data platform to construct a training dataset and a benchmark for trajectory field estimation. Experiments show that Trace Anything surpasses existing methods or performs competitively on the new benchmark and established point tracking benchmarks, with significant efficiency gains. Moreover, it facilitates downstream applications such as goal-conditioned manipulation, simple motion extrapolation, and spatio-temporal fusion. We will release the code, the model weights, and the data platform.

Open-set text classification (OSTC) requires models to correctly classify in-distribution (ID) samples while reliably rejecting out-of-distribution (OOD) inputs—an essential capability for real-world NLP systems. Most OSTC methods train on ID data under the closed assumption that all outputs belong to the known label space and then perform OOD detection with the biased representations, which inherently lack awareness of unknowns and thus yield overconfident predictions on OOD inputs. In this work, we present UnLLM, an Unknown-aware Large Language Model for OSTC. Instead of fixing classification to the entire known label space, we reformulate it into a subset-conditioned text generation task: the LLM is prompted with sampled subsets of known labels, and any instance outside the candidate set is explicitly assigned as “unknown”. This reformulation transforms OOD detection from a post-hoc procedure into an intrinsic modeling capability. More importantly, our approach is the first to explicitly incorporate the unknown into classification, enabling systematic modeling of unknowns through a unified representation–logits–inference optimization, which progressively strengthens the model’s capacity to capture open-set risk. Extensive experiments across six benchmarks show that UnLLM consistently outperforms state-of-the-art (SOTA) baselines. Code is available in an anonymous repository: https://anonymous.4open.science/r/UnLLM-03C2.