ICLR 2026 Papers

Audio-visual speech separation (AVSS) methods leverage visual cues to extract target speech and have demonstrated strong separation quality in noisy acoustic environments. However, these methods usually involve a large number of parameters and require high computational cost, which is unacceptable in many applications where speech separation serves as only a preprocessing step for further speech processing. To address this issue, we propose an efficient AVSS method, named **Dolphin**. For visual feature extraction, we develop **DP‑LipCoder**, a dual‑path lightweight video encoder that transforms lip‑motion into discrete audio‑aligned semantic tokens. For audio separation, we construct a lightweight encoder–decoder separator, in which each layer incorporates a global–local attention (GLA) block to efficiently capture multi-scale dependencies. Experiments on three benchmark datasets showed that Dolphin not only surpassed the current state-of-the-art (SOTA) model in separation quality but also achieved remarkable improvements in efficiency: over 50\% fewer parameters, more than 2.4$\times$ reduction in MACs, and over 6$\times$ faster GPU inference speed. These results indicate that Dolphin offers a practical and deployable solution for high-performance AVSS in real-world scenarios. Our code and demo page are publicly available at https://dolphin-avss.github.io/Dolphin.

Large reasoning models (LRMs) have shown strong performance in complex mathematical reasoning when optimized via reinforcement learning (RL). However, conventional outcome-only reward provides sparse feedback, leading to inefficient optimization. In this work, we investigate whether generative process reward models (GenPRMs) can accelerate RL training of LRMs by improving the utilization of reasoning trajectories. We first analyze critical limitations in existing GenPRMs, including their heavy reliance on reasoning ability during correctness judgment, and suppression of exploration as well as vulnerability to reward hacking during reward assignment. To address these limitations, we first propose a novel \textbf{intrinsic-signal-driven evaluation} mechanism, which judges reasoning steps using semantic cues from the solution, thus mitigating extensive dependence on GenPRM. Furthermore, we (i) adopt \textbf{thought-level rewarding granularity} to alleviate over-dense step rewards, and (ii) design a \textbf{difficulty-aware reward formulation} that dynamically balances exploration and exploitation and keeping the optimization target of key tokens to mitigate reward hacking. We integrate these innovations into the process reward-based GRPO, resulting in the proposed \textbf{TP-GRPO} algorithm. Experiments on LRMs with 1.5B and 7B parameters show that TP-GRPO achieves higher improvements while using significantly fewer training samples, and more analyses further confirm the effectiveness of our proposed process evaluation mechanism.

Human interaction is inherently multimodal and full-duplex: we listen while watching, speak while acting, and fluidly adapt to turn-taking and interruptions. Realizing these capabilities is essential for building models simulating humans. We present ELLSA (End-to-end Listen, Look, Speak and Act), which, to our knowledge, is the first full-duplex, end-to-end model that simultaneously perceives and generates across vision, text, speech, and action within a single architecture, enabling interaction patterns previously out of reach, yielding more natural, human-like behaviors. At its core is a novel SA-MoE architecture (Self-Attention Mixture-of-Experts) that routes each modality to specialized experts and fuses them through a unified attention backbone. This provides a generalizable solution for joint multimodal perception and concurrent generation, leveraging strong pre-trained components while enabling efficient modality integration and mitigating modality interference. On speech-interaction and robot-manipulation benchmarks, ELLSA matches modality-specific baselines, while uniquely supporting advanced multimodal and full-duplex behaviors such as dialogue and action turn-taking, defective instruction rejection, speaking-while-acting, context-grounded visual question answering, and action barge-ins. We contend that ELLSA represents a step toward more natural and general interactive intelligence, contributing to the broader pursuit of artificial general intelligence. A demonstration is available at https://anonymous.4open.science/r/LLSA-E821.

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.

Full-image relighting remains a challenging problem due to the difficulty of collecting large-scale structured paired data, the difficulty of maintaining physical plausibility, and the limited generalizability imposed by data-driven priors. Existing attempts to bridge the synthetic-to-real gap for full-scene relighting remain suboptimal. To tackle these challenges, we introduce **P**hysics-**I**nspired diffusion for full-image re**Light** ($\pi$-Light, or PI-Light), a two-stage framework that leverages physics-inspired diffusion models. Our design incorporates (i) batch-aware attention, which improves the consistency of intrinsic predictions across a collection of images, (ii) a physics-guided neural rendering module that enforces physically plausible light transport, (iii) physics-inspired losses that regularize training dynamics toward a physically meaningful landscape, thereby enhancing generalizability to real-world image editing, and (iv) a carefully curated dataset of diverse objects and scenes captured under controlled lighting conditions. Together, these components enable efficient finetuning of pretrained diffusion models while also providing a solid benchmark for downstream evaluation. Experiments demonstrate that $\pi$-Light synthesizes specular highlights and diffuse reflections across a wide variety of materials, achieving superior generalization to real-world scenes compared with prior approaches.

Unified multimodal models (UMMs) have emerged as a powerful paradigm in fundamental cross-modality research, demonstrating significant potential in both image understanding and generation. However, existing research in the face domain primarily faces two challenges: **(1) fragmentation development**, with existing methods failing to unify understanding and generation into a single one, hindering the way to artificial general intelligence. **(2) lack of fine-grained facial attributes**, which are crucial for high-fidelity applications. To handle those issues, we propose UniF$^2$ace, the first UMM specifically tailored for fine-grained face understanding and generation. **First**, we introduce a novel theoretical framework with a Dual Discrete Diffusion (D3Diff) loss, unifying masked generative models with discrete score matching diffusion and leading to a more precise approximation of the negative log-likelihood. Moreover, this D3Diff significantly enhances the model's ability to synthesize high-fidelity facial details aligned with text input. **Second**, we propose a multi-level grouped Mixture-of-Experts architecture, adaptively incorporating the semantic and identity facial embeddings to complement the attribute forgotten phenomenon in representation evolvement. **Finally**, to this end, we construct UniF$^2$aceD-1M, a large-scale dataset comprising *130K* fine-grained image-caption pairs and *1M* visual question-answering pairs, spanning a much wider range of facial attributes than existing datasets. Extensive experiments demonstrate that UniF$^2$ace outperforms existing models with a similar scale in both understanding and generation tasks, with 7.1% higher Desc-GPT and 6.6% higher VQA-score, respectively. Code is available in the supplementary materials.

Ensuring Large Language Model (LLM) safety remains challenging due to the absence of universal standards and reliable content validators, making it difficult to obtain effective training signals. We discover that aligned models already possess robust internal safety beliefs: they consistently produce high-confidence refusals to harmful requests while exhibiting high entropy when generating potentially dangerous content. This entropy gap reveals an untapped signal—models intrinsically "know" when to refuse. We introduce Safety Instincts Reinforcement Learning (*SIRL*), which transforms this internal confidence into a self-generated reward signal, eliminating dependence on external validators or human annotations. *SIRL* teaches models to trust their safety instincts by reinforcing low-entropy refusal behaviors. Evaluated on Llama and Qwen models, *SIRL* maintains 89\%+ Defense Success Rates (DSRs) against 20+ jailbreak methods, from static prompts to automated attacks. Using only 15,000 unlabeled prompts, *SIRL* surpasses resource-intensive supervised methods while preserving performance on mathematics, coding, and conversation benchmarks. Our work demonstrates that effective alignment can emerge from within, paving the way for more autonomous and robust AI safety mechanisms that scale without extensive human oversight.

Modern large-scale recommendation systems rely heavily on user interaction history sequences to enhance the model performance. The advent of large language models and sequential modeling techniques, particularly transformer architectures, has led to significant advancements (e.g., HSTU, SIM, and TWIN models). While scaling to ultra-long user histories (10k to 100k items) generally improves model performance, it also creates significant challenges on latency, queries per second (QPS) and GPU cost in industry-scale recommendation systems. Existing models do not adequately address these industrial scalability issues. In this paper, we propose a novel two-stage modeling framework, namely \emph{VIrtual Sequential Target Attention} (VISTA), which decomposes traditional target attention from a candidate item to user history items into two distinct stages: (1) user history summarization into a few hundred tokens; followed by (2) candidate item attention to those tokens. These summarization token embeddings are then cached in storage system and then utilized as sequence features for downstream model training and inference. This novel design for scalability enables VISTA to scale to lifelong user histories (up to one million items) while keeping downstream training and inference costs fixed, which is essential in industry. Our approach achieves significant improvements in offline and online metrics and has been successfully deployed on an industrial platform serving billions of users.

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.

Reasoning before action and imagining potential outcomes (i.e., world models) are essential for embodied agents operating in complex open-world environments. Yet, prior work either incorporates only one of these abilities in an end-to-end agent or integrates multiple specialized models into an agent system, limiting the learning efficiency and generalization of the policy. Thus, this paper makes the first attempt to synergize Reasoning and Imagination in an end-to-end Generalist policy, termed RIG. To train RIG in an end-to-end manner, we construct a data pipeline that progressively integrates and enriches the content of imagination and reasoning in the trajectories collected from existing agents. The joint learning of reasoning and next image generation explicitly models the inherent correlation between reasoning, action, and dynamics of environments. It thus exhibits more than $17\times$ sample efficiency improvements and generalization in comparison with previous works. During inference, RIG first reasons about the next action, produces potential action, and then predicts the action outcomes, which offers the agent a chance to review and self-correct based on the imagination before taking real actions. Experimental results show that the synergy of reasoning and imagination not only improves the robustness, generalization, and interoperability of generalist policy but also enables test-time scaling to enhance overall performance.

Molecular dynamics (MD) simulations are essential tools in computational chemistry and drug discovery, offering crucial insights into dynamic molecular behavior. However, their utility is significantly limited by substantial computational costs, which severely restrict accessible timescales for many biologically relevant processes. Despite the encouraging performance of existing machine learning (ML) methods, they struggle to generate extended biomolecular system trajectories, primarily due to the lack of MD datasets and the large computational demands of modeling long historical trajectories. Here, we introduce BioMD, the first all-atom generative model to simulate long-timescale protein-ligand dynamics using a hierarchical framework of forecasting and interpolation. We demonstrate the effectiveness and versatility of BioMD on the DD-13M (ligand unbinding) and MISATO datasets. For both datasets, BioMD generates highly realistic conformations, showing high physical plausibility and low reconstruction errors. Besides, BioMD successfully generates ligand unbinding paths for 97.1% of the protein-ligand systems within ten attempts, demonstrating its ability to explore critical unbinding pathways. Collectively, these results establish BioMD as a tool for simulating complex biomolecular processes, offering broad applicability for computational chemistry and drug discovery.

Language model (LM)-based embodied agents are increasingly deployed in real-world settings. Yet, their adaptability remains limited in dynamic environments, where constructing accurate and flexible world models is crucial for effective reasoning and decision-making. To address this challenge, we extend the Mixture-of-Experts (MoE) paradigm to embodied agents. While conventional MoE architectures modularize knowledge into expert components with pre-trained routing, they remain rigid once deployed, making them less effective for adapting to unseen domains in dynamic environments. We therefore propose Test-time Mixture of World Models (TMoW), a framework that enhances adaptability to unseen and evolving domains. TMoW updates its routing function over world models at test time, unlike conventional MoE where the function remains fixed, enabling agents to recombine existing models and integrate new ones for continual adaptation. It achieves this through (i) multi-granular prototype-based routing, which adapts mixtures across object- to scene-level similarities, (ii) test-time refinement that aligns unseen domain features with prototypes during inference, and (iii) distilled mixture-based augmentation, which efficiently constructs new models from few-shot data and existing prototypes. We evaluate TMoW on VirtualHome, ALFWorld, and RLBench benchmarks, demonstrating strong performance in both zero-shot adaptation and few-shot expansion scenarios, and showing that it enables embodied agents to operate effectively in dynamic environments.

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.

We propose ImitSAT, a branching policy for conflict-driven clause learning (CDCL) solvers based on imitation learning for the Boolean satisfiability problem (SAT). Unlike previous methods that predict instance-level signals to improve CDCL branching indirectly, or rely on reinforcement learning and insufficient CDCL information to enhance branching, ImitSAT learns from expert KeyTrace that collapses a full run into the sequence of surviving decisions. Replaying a KeyTrace on the same instance is nearly conflict-free, providing dense decision- level supervision and directly reducing propagations—the dominant contributor to wall-clock time. This prefix-conditioned supervision enables ImitSAT to reproduce high-quality branches without exploration, yielding faster convergence, stable training, and seamless integration into CDCL. Extensive experiments demonstrate that ImitSAT reduces propagation counts and runtime, outperforming state-of-the-art learned approaches. We released the source code and trained model at https://github.com/zewei-Zhang/ImitSAT.

UV unwrapping is an essential task in computer graphics, enabling various visual editing operations in rendering pipelines. However, existing UV unwrapping methods struggle with time-consuming, fragmentation, lack of semanticity, and irregular UV islands, limiting their practical use. An artist-style UV map must not only satisfy fundamental criteria, such as overlap-free mapping and minimal distortion, but also uphold higher-level standards, including clean boundaries, efficient space utilization, and semantic coherence. We introduce ArtUV, a fully automated, end-to-end method for generating artist-style UV unwrapping. We simulates the professional UV mapping process by dividing it into two stages: surface seam prediction and artist-style UV parameterization. In the seam prediction stage, SeamGPT is used to generate semantically meaningful cutting seams. Then, in the parameterization stage, a rough UV obtained from an optimization-based method, along with the mesh, is fed into an Auto-Encoder, which refines it into an artist-style UV map. Our method ensures semantic consistency and preserves topological structure, making the UV map ready for 2D editing. We evaluate ArtUV across multiple benchmarks and show that it serves as a versatile solution, functioning seamlessly as either a plug-in for professional rendering tools or as a standalone system for rapid, high-quality UV generation.

Information processing in neural populations is inherently constrained by metabolic resource limits and noise properties, with dynamics that are not accurately described by existing mathematical models. Recent data, for example, shows that neurons in mouse visual cortex go into a "low power mode" in which they maintain firing rate homeostasis while expending less energy. This adaptation leads to increased neuronal noise and tuning curve flattening in response to metabolic stress. We have developed a theoretical population coding framework that captures this behavior using two novel, surprisingly simple constraints: an approximation of firing rate homeostasis and an energy limit tied to noise levels via biophysical simulation. A key feature of our contribution is an energy budget model directly connecting adenosine triphosphate (ATP) use in cells to a fully explainable mathematical framework that generalizes existing optimal population codes. Specifically, our simulation provides an energy-dependent dispersed Poisson noise model, based on the assumption that the cell will follow an optimal decay path to produce the least-noisy spike rate that is possible at a given cellular energy budget. Each state along this optimal path is associated with properties (resting potential and leak conductance) which can be measured in electrophysiology experiments and have been shown to change under prolonged caloric deprivation. We analytically derive the optimal coding strategy for neurons under varying energy budgets and coding goals, and show how our method uniquely captures how populations of tuning curves adapt while maintaining homeostasis, as has been observed empirically.