ICLR 2026 Papers

Multi-LLM systems harness the complementary strengths of diverse Large Language Models, achieving performance and efficiency gains unattainable by a single model. In existing designs, LLMs communicate through text, forcing internal representations to be transformed into output token sequences. This process both loses rich semantic information and incurs token-by-token generation latency. Motivated by these limitations, we ask: Can LLMs communicate beyond text? Oracle experiments show that enriching the KV-Cache semantics can improve response quality without increasing cache size, supporting KV-Cache as an effective medium for inter-model communication. Thus, we propose Cache-to-Cache (C2C), a new paradigm for direct semantic communication between LLMs. C2C uses a neural network to project and fuse the source model’s KV-cache with that of the target model to enable direct semantic transfer. A learnable gating mechanism selects the target layers that benefit from cache communication. Compared with text communication, C2C utilizes the deep, specialized semantics from both models, while avoiding explicit intermediate text generation. Experiments show that C2C achieves 6.4–14.2% higher average accuracy than individual models. It further outperforms the text communication paradigm by approximately 3.1–5.4%, while delivering an average 2.5× speedup in latency.

Vision-Language Models (VLMs) have achieved significant progress in multimodal understanding tasks, demonstrating strong capabilities particularly in general tasks such as image captioning and visual reasoning. However, when dealing with specialized cultural heritage domains like 3D vase artifacts, existing models face severe data scarcity issues and insufficient domain knowledge limitations. Due to the lack of targeted training data, current VLMs struggle to effectively handle such culturally significant specialized tasks. To address these challenges, we propose the VaseVQA-3D dataset, which serves as the first 3D visual question answering dataset for ancient Greek pottery analysis, collecting 664 ancient Greek vase 3D models with corresponding question-answer data and establishing a complete data construction pipeline. We further develop the VaseVLM model, enhancing model performance in vase artifact analysis through domain-adaptive training. Experimental results validate the effectiveness of our approach, where our VaseVLM-7B-RL achieves 12.8\% improvement in R@1 accuracy and 6.6\% improvement in lexical similarity compared to the strongest baselines on the VaseVQA-3D dataset, significantly improving the recognition and understanding of 3D vase artifacts, providing new technical pathways for digital heritage preservation research.

Multimodal representation learning aims to capture both shared and complementary semantic information across multiple modalities. However, the intrinsic heterogeneity of diverse modalities presents substantial challenges to achieve effective cross-modal collaboration and integration. To address this, we introduce DecAlign, a novel hierarchical cross-modal alignment framework designed to decouple multimodal representations into modality-unique (heterogeneous) and modality-common (homogeneous) features. For handling heterogeneity, we employ a prototype-guided optimal transport alignment strategy leveraging gaussian mixture modeling and multi-marginal transport plans, thus mitigating distribution discrepancies while preserving modality-unique characteristics. To reinforce homogeneity, we ensure semantic consistency across modalities by aligning latent distribution matching with Maximum Mean Discrepancy regularization. Furthermore, we incorporate a multimodal transformer to enhance high-level semantic feature fusion, thereby further reducing cross-modal inconsistencies. Our extensive experiments on four widely used multimodal benchmarks demonstrate that DecAlign consistently outperforms existing state-of-the-art methods across five metrics. These results highlight the efficacy of DecAlign in enhancing superior cross-modal alignment and semantic consistency while preserving modality-unique features, marking a significant advancement in multimodal representation learning scenarios.

Verification is a key bottleneck in improving inference speed while maintaining distribution fidelity in Speculative Decoding. Recent work has shown that sequence-level verification leads to a higher number of accepted tokens compared to token-wise verification. However, existing solutions often rely on surrogate approximations or are constrained by partial information, struggling with joint intractability. In this work, we propose Hierarchical Speculative Decoding (HSD), a provably lossless verification method that significantly boosts the expected number of accepted tokens and overcomes joint intractability by balancing excess and deficient mass across accessible branches. Through extensive large-scale experiments, we show that HSD consistently improves acceptance rates, especially with longer draft sequences. Its strong explainability and generality further highlight the potential for integration into a wide range of speculative decoding frameworks.

Large Language Models (LLMs) currently respond to every prompt. However, they can produce incorrect answers when they lack knowledge or capability -- a problem known as hallucination. We instead propose post-training an LLM to generate content only when confident in its correctness and to otherwise (partially) abstain. Specifically, our method, HALT, produces capability-aligned post-training data that encodes what the model can and cannot reliably generate. We generate this data by splitting responses of the pretrained LLM into factual fragments (atomic statements or reasoning steps), and use ground truth information to identify incorrect fragments. We achieve capability-aligned finetuning responses by either removing incorrect fragments or replacing them with "Unsure from Here" -- according to a tunable threshold that allows practitioners to trade off response completeness and mean correctness of the response's fragments. We finetune four open-source models for biography writing, mathematics, coding, and medicine with HALT for three different trade-off thresholds. HALT effectively trades off response completeness for correctness, increasing the mean correctness of response fragments by 15% on average, while resulting in a 4% improvement in the F1 score (mean of completeness and correctness of the response) compared to the relevant baselines. By tuning HALT for highest correctness, we train a single reliable Llama3-70B model with correctness increased from 51% to 87% across all four domains while maintaining 53% of the response completeness achieved with standard finetuning.

We introduce FLARE, a family of vision language models (VLMs) with a fully vision-language alignment and integration paradigm. Unlike existing approaches that rely on single MLP projectors for modality alignment and defer cross-modal interaction to LLM decoding, FLARE achieves deep, dynamic integration throughout the pipeline. Our key contributions include: (1) Text-Guided Vision Encoding that incorporates textual information during vision encoding to achieve pixel-level alignment; (2) Context-Aware Alignment Decoding that aggregates visual features conditioned on textual context during decoding for query-level integration; (3) Dual-Semantic Mapping Loss to supervise feature mapping from both modalities and enable modality-level bridging; and (4) Text-Driven VQA Synthesis that leverages high-quality text to generate VQA pairs and synthesize corresponding images, enabling data-level optimization. We train FLARE at 3B and 8B scales under both fixed and dynamic resolution settings, demonstrating that our full-modality alignment significantly outperforms existing methods while maintaining strong generalizability. FLARE 3B surpasses Cambrian-1 8B and Florence-VL 8B using only 630 vision tokens. Ablation studies reveal that FLARE achieves superior performance over existing methods with minimal computational cost. Even without dynamic resolution, FLARE outperforms LLaVA-NeXT, validating the effectiveness of our approach.

Methods based on weight compensation, which iteratively apply quantization and weight compensation to minimize the output error, have recently demonstrated remarkable success in quantizing Large Language Models (LLMs). The representative work, GPTQ, introduces several key techniques that make such iterative methods practical for LLMs with billions of parameters. GPTAQ extends this approach by introducing an asymmetric calibration process that aligns the output of each quantized layer with its full-precision counterpart, incorporating a residual error into the weight compensation framework. In this work, we revisit the formulation of the residual error. We identify a sub-optimal calibration objective in existing methods: during the intra-layer calibration process, they align the quantized output with the output from compensated weights, rather than the true output from the original full-precision model. This leads to a sub-optimal calibration objective. Therefore, we redefine the objective to precisely align the quantized model's output with the original output of the full-precision model at each step. We then reveal that the residual error originates not only from the output difference of the preceding layer but also from the discrepancy between the compensated and original weights within each layer, which we name the 'compensation-aware error'. By inheriting the neuron decomposition technique from GPTAQ, we can efficiently incorporate this compensation-aware error into the weight update process. Extensive experiments on various LLMs and quantization settings demonstrate that our proposed enhancements integrate seamlessly with both GPTQ and GPTAQ, significantly improving their quantization performance.

Universal photometric stereo (PS) is defined by two factors: it must (i) operate under arbitrary, unknown lighting conditions and (ii) avoid reliance on specific illumination models. Despite progress (e.g., SDM UniPS), two challenges remain. First, current encoders cannot guarantee that illumination and normal information are decoupled. To enforce decoupling, we introduce LINO UniPS with two key components: (i) Light Register Tokens with light alignment supervision to aggregate point, direction, and environment lights; (ii) Interleaved Attention Block featuring global cross-image attention that takes all lighting conditions together so the encoder can factor out lighting while retaining normal-related evidence. Second, high-frequency geometric details are easily lost. We address this with (i) a Wavelet-based Dual-branch Architecture and (ii) a Normal-gradient Perception Loss. These techniques yield a \textbf{unified} feature space in which lighting is explicitly represented by register tokens, while normal details are preserved via wavelet branch. We further introduce PS-Verse, a large-scale synthetic dataset graded by geometric complexity and lighting diversity, and adopt curriculum training from simple to complex scenes. Extensive experiments show new state-of-the-art results on public benchmarks (e.g., DiLiGenT, Luces), stronger generalization to real materials, and improved efficiency; ablations confirm that Light Register Tokens + Interleaved Attention Block drive better feature decoupling, while Wavelet-based Dual-branch Architecture + Normal-gradient Perception Loss recover finer details.

Vision-Language Navigation (VLN) requires an agent to navigate 3D environments following natural language instructions. During navigation, existing agents commonly encounter perceptual uncertainty, such as insufficient evidence for reliable grounding or ambiguity in interpreting spatial cues, yet they typically ignore such information when predicting actions. In this work, we explicitly model three forms of perceptual uncertainty (i.e., geometric, semantic, and appearance uncertainty) and integrate them into the agent’s observation space to enable informed decision-making. Concretely, our agent first constructs a Semantic Gaussian Map (SGM), composed of differentiable 3D Gaussian primitives initialized from panoramic observations, that encodes both the geometric structure and semantic content of the environment. On top of SGM, geometric uncertainty is estimated through variational perturbations of Gaussian position and scale to assess structural reliability; semantic uncertainty is captured by perturbing Gaussian semantic attributes to reveal ambiguous interpretations; and appearance uncertainty is characterized by Fisher Information, which measures the sensitivity of rendered observations to Gaussian-level variations. These uncertainties are incorporated into SGM, extending it into a unified 3D Value Map, which grounds them as affordances and constraints that support reliable navigation. Comprehensive evaluations across multiple VLN benchmarks (i.e., R2R, RxR, REVERIE) demonstrate the effectiveness of our agent. The code will be released.

The past year has witnessed the significant advancement of video-based large language models. However, the challenge of developing a unified model for both short and long video understanding remains unresolved. Most existing video LLMs cannot handle hour-long videos, while methods custom for long videos tend to be ineffective for shorter videos and images. In this paper, we identify the key issue as the redundant content in videos. To address this, we propose a novel pooling strategy that simultaneously achieves token compression and instruction-aware visual feature aggregation. Our model is termed Prompt-guided Pooling LLaVA, or PPLLaVA for short. Specifically, PPLLaVA consists of three core components: the CLIP-based visual-prompt alignment that extracts visual information relevant to the user's instructions, the prompt-guided pooling that compresses the visual sequence to arbitrary scales using convolution-style pooling, and the clip context extension designed for lengthy prompt common in visual dialogue. Extensive experiments have validated the performance of our model. With superior throughput, PPLLaVA achieves better results on image benchmarks as a video LLM, while achieving state-of-the-art performance across various video benchmarks, excelling in tasks ranging from caption generation to multiple-choice questions, and handling video lengths from seconds to hours.

In this paper, we introduce SemHiTok, a unified image Tokenizer via Semantic-Guided Hierarchical codebook (SGHC) that provides consistent discrete representations for multimodal understanding and generation. Recently, unified image tokenizers have sparked exploration within the research community, which is designed to capture high-level semantic features for understanding and retaining low-level pixel features for generation. Previous works attempt to train a unified image tokenizer by combining loss for semantic distillation and pixel reconstruction. However, due to the differing levels of features prioritized by multimodal understanding and generation, joint training methods face significant challenges in achieving a good trade-off. SemHiTok addresses this challenge through a novel semantic-guided hierarchical codebook, which builds pixel sub-codebooks on a pretrained semantic codebook. This design decouples semantic and pixel in both terms of structure and training strategy, enabling the tokenizer to capture pixel features while retaining its ability to comprehend high-level semantic information. Our experiments demonstrate that SemHiTok achieves SOTA performance in image reconstruction and multimodal understanding under the LLaVA-v1.5 setting. Further, we develop a unified MLLM with SemHiTok, which exhibits superior performance across multimodal understanding and generation tasks. For understanding, SemHiTok achieves impressive performance on most benchmarks. For generation, our model achieves SOTA performance on MJHQ30K in unified MLLMs. Our code and models will be open source.

Code data has been shown to enhance the reasoning capabilities of large language models (LLMs), but it remains unclear which aspects of code are most responsible. We investigate this question with a systematic, data-centric framework. We construct parallel instruction datasets in ten programming languages and apply controlled perturbations that selectively disrupt structural or semantic properties of code. We then finetune LLMs from five model families and eight scales on each variant and evaluate their performance on natural language, math, and code tasks. Across 3,331 experiments, our results show that LLMs are more vulnerable to structural perturbations than semantic ones, particularly on math and code tasks. Appropriate abstractions like pseudocode and flowcharts can be as effective as code, while encoding the same information with fewer tokens without adhering to original syntax can often retain or even improve performance. Remarkably, even corrupted code with misleading signals remains competitive when surface-level regularities persist. Finally, syntactic styles also shape task-specific gains with Python favoring natural language reasoning and lower-level languages such as Java and Rust favoring math. Through our systematic framework, we aim to provide insight into how different properties of code influence reasoning and inform the design of training data for enhancing LLM reasoning capabilities.

Recent benchmarks and datasets have been proposed to improve spatial reasoning in vision-language models (VLMs), yet existing open resources remain limited in scale, visual diversity, and instruction expressiveness. In this work, we introduce InternSpatial, the largest open-source dataset for spatial reasoning in VLMs, along with InternSpatial-Bench, a corresponding evaluation benchmark designed to assess spatial understanding under diverse instruction formats. InternSpatial comprises 12 million QA pairs spanning both single-view and multi-view settings, drawn from diverse visual environments and supporting 19 instruction formats that reflect varied query styles. For evaluation, we propose InternSpatial-Bench for single-view tasks and expand multi-view reasoning by introducing a novel rotation angle prediction task that has not been explored in prior work. Experimental results show that models trained on InternSpatial achieve 12.1% improvement on InternSpatial-Bench and 10.7% on VSI-Bench, while maintaining strong performance on general-purpose benchmarks. We hope these resources will support the development of spatially capable VLMs in practical applications such as robotics and embodied AI.

Diffusion-based large language models (dLLMs) are trained to model extreme flexibility/dependence in the data-distribution; however, how to best utilize this at inference time remains an open problem. In this work, we uncover an interesting property of these models: dLLMs {trained on textual data} implicitly learn a mixture of semi-autoregressive experts, where different generation orders reveal different specialized behaviors. We show that committing to any single, fixed inference time schedule, a common practice, collapses performance by failing to leverage this latent ensemble. To address this, we introduce HEX (Hidden semi-autoregressive EXperts for test-time scaling), a training-free inference method that ensembles across heterogeneous block schedules. By doing a majority vote over diverse block-sized generation paths, HEX robustly avoids failure modes associated with any single fixed schedule. On reasoning benchmarks such as GSM8K, it boosts accuracy by up to 3.56× (from 24.72\% to 88.10\%), outperforming top-K margin inference and specialized fine-tuned methods like GRPO, without additional training. HEX even yields significant gains on MATH benchmark from 16.40\% to 40.00\%, scientific reasoning on ARC-C from 54.18\% to 87.80\%, and TruthfulQA from 28.36\% to 57.46\%. Our results establish test-time scaling as a powerful principle for dLLMs, showing that the sequence in which masking is done can play a significant role in test-time scaling/inferencing of dLLMs.

Chain-of-Thought (CoT) prompting has emerged as a powerful approach to enhancing the reasoning capabilities of Large Language Models (LLMs). However, existing implementations, such as in-context learning and fine-tuning, remain costly and inefficient. To improve CoT reasoning at a lower cost, and inspired by the task vector paradigm, we introduce CoT Vectors, compact representations that encode task-general, multi-step reasoning knowledge. Through experiments with Extracted CoT Vectors, we observe pronounced layer-wise instability, manifesting as a U-shaped performance curve that reflects a systematic three-stage reasoning process in LLMs. To address this limitation, we propose Learnable CoT Vectors, optimized under a teacher–student framework to provide more stable and robust guidance. Extensive evaluations across diverse benchmarks and models demonstrate that CoT Vectors not only outperform existing baselines but also achieve performance comparable to parameter-efficient fine-tuning methods, while requiring fewer trainable parameters. Moreover, by treating CoT Vectors as a probe, we uncover how their effectiveness varies due to latent space structure, information density, acquisition mechanisms, and pre-training differences, offering new insights into the functional organization of multi-step reasoning in LLMs. The source code will be released.

Composed video retrieval presents a complex challenge: retrieving a target video based on a source video and a textual modification instruction. This task demands fine-grained reasoning over multimodal transformations. However, existing benchmarks predominantly focus on vision–text alignment, largely overlooking the rich semantic signals embedded in audio—such as speech, music, and environmental sounds—which are often decisive for comprehensive video understanding. To bridge this gap, we introduce **OmniCVR**, a large-scale benchmark for omni-composed video retrieval that establishes vision, audio, and text as first-class modalities. OmniCVR is constructed via a scalable, automated pipeline integrating content-aware segmentation, omni-modal annotation, and a rigorous dual-validation protocol involving both large language models and human experts. The benchmark comprises vision-centric, audio-centric, and integrated queries, with the latter forming the majority to accurately reflect real-world multimodal complexity. Furthermore, we propose **AudioVLM2Vec**, an audio-aware extension of VLM2Vec. By incorporating explicit audio semantics, AudioVLM2Vec achieves state-of-the-art performance, highlighting fundamental limitations in the audio reasoning capabilities of current multimodal retrieval systems.

Multimodal Large Language Models (MLLMs) pose critical safety challenges, as they are susceptible not only to adversarial attacks such as jailbreaking but also to inadvertently generating harmful content for benign users. While internal safety alignment via Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) is a primary mitigation strategy, current methods often face a safety-utility trade-off: they either refuse benign queries out of excessive caution or overlook latent risks in cross-modal interactions. To resolve this, we introduce Pragma-VL, an end-to-end alignment algorithm that enables MLLMs to pragmatically arbitrate between safety and helpfulness. First, we enhance visual risk perception with a novel cold-start SFT stage. This is achieved by applying risk-aware clustering to the visual encoder and using an interleaved dataset of risk descriptions and high-quality data. Second, we introduce a theoretically-guaranteed reward model that leverages synergistic learning. We train it with a novel data augmentation method that assigns dynamic weights based on the queries, enabling contextual arbitration between safety and helpfulness. Extensive experiments show that Pragma-VL effectively balances safety and helpfulness, outperforming baselines by 5% to 20% on most multimodal safety benchmarks while preserving its general capabilities in areas such as mathematics and knowledge reasoning.

Different time-series tasks benefit from distinct cues at various spaces and abstractions, yet existing time-series pre-trained models entangle these signals within large, monolithic embeddings, limiting transferability and zero-shot usability. Moreover, massive model sizes demand heavy compute, restricting practical deployments and real-time applications. To address this, we propose TSPulse, an ultra-light pre-trained model (1M parameters) that performs disentangled masked reconstruction across spaces and abstraction levels, explicitly learning three disentangled views: temporal embeddings for fine-grained time analysis, spectral embeddings for frequency-aware fidelity, and semantic embeddings for high-level task understanding. A hybrid masking scheme further randomizes mask style and span length to avoid pre-training bias and improve robustness. Despite its compact size, TSPulse achieves strong gains across four time-series tasks: +20\% and rank-1 on TSB-AD leaderboard benchmark for reliable anomaly detection through multi-head triangulation, which correlates complementary cues across disentangled views; +25\% in similarity search as the disentangled semantic embedding remain invariant to time, scale and noise shifts, making retrieval more robust; +50\% improvement in imputation since hybrid masking exposes the model to diverse real-world corruption patterns; and +5–16\% gains in multivariate classification with TSLens, a lightweight module that selectively attends to the most informative signals across variates. Overall, TSPulse outperform models that are 10–100× larger on 75+ datasets across tasks, while delivering state-of-the-art zero-shot results with GPU-free support and efficient fine-tuning. Models and source code will be open-sourced and currently shared in the supplementary material.

Vision transformers have demonstrated remarkable success in classification by leveraging global self-attention to capture long-range dependencies. However, this same mechanism can obscure fine-grained spatial details crucial for tasks such as segmentation. In this work, we seek to enhance the segmentation performance of vision transformers after being trained using the usual image-level classification objective. More specifically, we present a simple yet effective add-on for vision transformers that improve their performance on segmentation tasks while retaining their image-level recognition capabilities. In our approach, we modulate the self-attention with a learnable Gaussian kernel that biases the attention toward neighboring patches. We further refine the patch representations to learn better embeddings at patch positions. These modifications ensure meaningful representations at spatial positions and encourage tokens to focus on local surroundings, while still preserving the model's ability to incorporate global information. Experiments demonstrate the effectiveness of our modifications, evidenced by substantial segmentation gains on three benchmarks (e.g., over 6% and 4% on ADE20K for ViT Tiny and Base), without changing the training regime or sacrificing classification performance. The code is available at https://anonymous.4open.science/r/LocAtViTRepo/.