ICLR 2026 Papers

Quantization techniques are essential for the deployment of Large Language Models (LLMs) on edge devices. However, prevailing methods often rely on mixed-precision multiplication that lacks efficient hardware support, making it not feasible. Ternary weight quantization addresses this by constraining weights to {-1, 0, 1}, replacing expensive multiplications with hardware-efficient additions. However, such aggressive compression leads to significant accuracy degradation, even after costly quantization-aware training with massive data. We identify the core issue as _**deadzone trapping**: a large number of weights are trapped at the deadzone boundary._ This occurs because these weights receive only noisy, less informative gradients, preventing stable escape from the deadzone and severely impeding model capacity and optimization. To address this issue, we propose **Tequila**, a trapping-free quantization optimization method that reactivates deadzone-trapped weights by repurposing them as dynamic biases. This allows the repurposed weights to provide a continuous signal in the forward pass and, critically, receive direct, meaningful gradient signals during backpropagation, thereby enhancing model capacity and optimization with nearly _zero_ inference overhead. Extensive evaluations demonstrate that Tequila outperforms state-of-the-art (SOTA) ternary quantization methods across five benchmarks. Specifically, on the ARC benchmark, it achieves $>4$% accuracy gain over the SOTA baseline, nearly matching full-precision performance (within $<1$% gap) with an $3.0\times$ inference speedup. Consequently, Tequila offers a highly practical and efficient implementation for the deployment of advanced LLMs in resource-constrained environments. The code is available at \url{https://anonymous.4open.science/r/Tequila-2B5F/}

Existing mobile device control agents often perform poorly when solving complex tasks requiring long-horizon planning and precise operations, typically due to a lack of relevant task experience or unfamiliarity with skill execution. We propose $\textbf{K²-Agent}$, a hierarchical framework that models human-like cognition by separating and co-evolving declarative ("knowing what") and procedural ("knowing how") knowledge for planning and execution. K²-Agent’s high level reasoner is bootstrapped from a single demonstration per task and runs a Summarize–Reflect–Locate–Revise (SRLR) loop to distill and iteratively refine task-level declarative knowledge through self-evolution. The low-level executor is trained with our curriculum-guided Group Relative Policy Optimization (C-GRPO), which (i) constructs a balanced sample pool using decoupled reward signals and (ii) employs dynamic demonstration injection to guide the model in autonomously generating successful trajectories for training. On the challenging AndroidWorld benchmark, K$^2$-Agent achieves a new $\textbf{state of the art}$ with $\textbf{76.1\% success rate}$, ranking $\textbf{1st}$ among all methods $\textbf{using only raw screenshots and open-source backbones}$. Furthermore, K²-Agent shows powerful dual generalization: its high-level declarative knowledge transfers across diverse base models, while its low-level procedural skills achieve competitive performance on unseen tasks in ScreenSpot-v2 and Android-in-the-Wild (AitW).

Popular explanation methods often produce unreliable feature importance scores due to "missingness bias", a systematic distortion that arises when models are probed with ablated, out-of-distribution inputs. Existing solutions treat this as a deep representational flaw that requires expensive retraining or architectural modifications. In this work, we challenge this assumption and show that missingness bias can be effectively treated as a superficial artifact of the model's output space. We introduce MCal, a lightweight post-hoc method that corrects this bias by fine-tuning a simple linear head on the outputs of a frozen base model. Surprisingly, we find this simple correction consistently reduces missingness bias and is competitive with, or even outperforms, prior heavyweight approaches across diverse medical benchmarks spanning vision, language, and tabular domains.

Despite the importance of global contextual modeling in visual tasks, a rigorous mathematical definition remains absent, and the concept is still largely described in heuristic or empirical terms. Existing methods either rely on computationally expensive attention mechanisms or are constrained by the recursive modeling nature of State Space Models (SSMs), making it challenging to achieve both efficiency and true global perception. To address this, we first propose a mathematical definition of global modeling for visual images, providing a theoretical foundation for designing globally-aware and interpretable models. Based on in-depth analysis of SSMs and frequency-domain modeling principles, we construct a complete theoretical framework that overcomes the limitations imposed by SSMs' recursive modeling mechanism from a frequency perspective, thereby adapting SSMs for global perception in image modeling. Guided by this framework, we design the Global-aware SSM (GSSM) module and formally prove that it satisfies definitional requirements of global image modeling. GSSM leverages a Discrete Fourier Transform (DFT)-based modulation mechanism, providing precise front-end control over the SSM's modeling behavior, and enabling efficient global image modeling with linear-logarithmic complexity. Building upon GSSM, we develop GMamba, a plug-and-play module that can be seamlessly integrated at any stage of Convolutional Neural Networks (CNNs). Extensive experiments across multiple tasks, including object detection, semantic segmentation, and instance segmentation, across diverse model architectures, demonstrate that GMamba consistently outperforms existing global modeling modules, validating both the effectiveness of our theoretical framework and the rigor of proposed definition. Code is available at \url{https://github.com/Xinmu-Tantai/GMamba-GSSM}

Understanding long videos requires Multimodal Large Language Models (MLLMs) to grasp multi-timescale information, often organized in hierarchies. However, current long-video understanding benchmarks either overlook multi-timescale design or distribute questions targeting different timescales across different videos. This approach entangles timescales with video content, thereby hindering a clear assessment of MLLM multi-timescale performance. To address this, we introduce ScaleLong, the first benchmark to disentangle these factors by embedding questions targeting four hierarchical timescales\textemdash clip (seconds), shot (tens of seconds), event (minutes), and story (hours)\textemdash all within the same video content. This ``within-content'' multi-timescale questioning design enables direct comparison of model performance across timescales on identical videos. ScaleLong features 269 videos (avg. 86 min) from 5 main categories and 36 sub-categories, with 4–8 carefully designed questions, with at least one question targeting each timescale. Evaluating 22 MLLMs reveals a distinct U-shaped performance trend: higher accuracy at the shortest (clip) and longest (story) timescales, with a dip at intermediate levels. Furthermore, ablation studies demonstrate that increased visual token capacity consistently enhances reasoning across all timescales. ScaleLong offers a crucial fine-grained, multi-timescale benchmark for advancing MLLM capabilities in long-video understanding. The code and dataset are available at \url{https://anonymous.4open.science/r/ScaleLong-7717}.

Anticipating traffic accidents in real time is critical for intelligent transportation systems, yet remains challenging under edge-device constraints. We propose a lightweight spatio-temporal framework that introduces a temporally shifted distillation strategy, enabling a student model to acquire predictive temporal dynamics from a frozen image-based teacher without requiring a video pre-trained teacher. The student combines a RepMixer spatial encoding with a RWKV-inspired recurrent module for efficient long-range temporal reasoning. To enhance robustness under partial observability, we design a masking memory strategy that leverages memory retention to reconstruct missing visual tokens, effectively simulating occlusions and future events. In addition, multi-modal vision-language supervision enriches semantic grounding. Our framework achieves state-of-the-art performance on multiple real-world dashcam benchmarks while sustaining real-time inference on resource-limited platforms such as the NVIDIA Jetson Orin Nano. Remarkably, it is 3-7$\times$ smaller than leading approaches yet delivers superior accuracy and earlier anticipation, underscoring its practicality for deployment in intelligent vehicles.

While significant research has focused on developing embodied reasoning capabilities using Vision-Language Models (VLMs) or integrating advanced VLMs into Vision-Language-Action (VLA) models for end-to-end robot control, few studies directly address the critical gap between upstream VLM-based reasoning and downstream VLA policy learning. In this work, we take an initial step toward bridging embodied reasoning with VLA policy learning by introducing **Vlaser** - a **V**ision-**L**anguage-**A**ction Model with **s**ynergistic **e**mbodied **r**easoning capability, which is a foundational vision-language model designed to integrate high-level reasoning with low-level control for embodied agents. Built upon the high-quality Vlaser-6M dataset, Vlaser achieves state-of-the-art performance across a range of embodied reasoning benchmarks—including spatial reasoning, embodied grounding, embodied QA, and task planning. Furthermore, we systematically examine how different VLM initializations affect supervised VLA fine-tuning, offering novel insights into mitigating the domain shift between internet-scale pre-training data and embodied-specific policy learning data. Based on these insights, our approach achieves state-of-the-art results on the WidowX benchmark and competitive performance on the Google Robot benchmark. We will open-source the model weights, data generation pipelines, and the full dataset to support future research.

Reinforcement learning with verifiable rewards (RLVR) has significantly improved reasoning in large language models (LLMs), yet the token-level mechanisms through which they reshape model behavior remain unclear. We present a systematic empirical study of RLVR’s distributional effects across three complementary axes: (1) token-level distributional shifts, (2) functional validation via cross-sampling interventions, and (3) exploratory investigations of advantage signal modulation based on token divergence. We find that RL fine-tuning induces sparse, targeted changes, with only a small fraction of tokens exhibiting significant distributional divergence, and we further analyze the nature of these shifts. These divergent distributions are not uniformly predicted by entropy, indicating that RLVR can modify both initially high and low entropy distributions under different settings. Cross-sampling experiments reveal that inserting just a small fraction of RL-sampled tokens into base model generations recovers most RL performance gains, while injecting a small portion of base-sampled tokens into RL generations collapses performance to base levels, functionally isolating the critical role of divergent tokens. Finally, we explore divergence-weighted variants of the advantage signal, finding that they can amplify improvements in baselines. Our work sheds light on the distributional changes induced by RLVR and provides a granular, token-level lens for understanding and improving RL fine-tuning in LLMs.

Low-bit quantization has achieved significant progress in image super-resolution. However, existing quantization methods show evident limitations in handling the heterogeneity of different components. Particularly under extreme low-bit compression, the issue of information loss becomes especially pronounced. In this work, we present a novel low-bit post-training quantization method, namely static priority-based rectifier routing quantization (SPR$^2$Q). The starting point of this work is to attempt to inject rich and comprehensive compensation information into the model before the quantization , thereby enhancing the model's inference performance after quantization. Firstly, we constructed a low-rank rectifier group and embedded it into the model's fine-tuning process. By integrating weight increments learned from each rectifier, the model enhances the backbone network while minimizing information loss during the lightweighting process. Furthermore, we introduce a static rectifier priority routing mechanism that evaluates the offline capability of each rectifier and generates a fixed routing table. During quantisation, it updates weights based on each rectifier's priority, enhancing the model's capacity and representational power without introducing additional overhead during inference. Extensive experiments demonstrate that the proposed SPR$^2$Q significantly outperforms the state-of-the-arts in five benchmark datasets, achieving PSNR improvements of 0.55 and 1.31 dB on the Set5($\times 2$) dataset under 4-bit and 2-bit settings, respectively.

Current visual reasoning methods mainly focus on exploring specific reasoning modes. Although improvements can be achieved in particular domains, they struggle to develop general reasoning capabilities. Inspired by this, we propose a novel adaptive reasoning paradigm, $\underline{\text{M}}$ixture-$\underline{\text{o}}$f-$\underline{\text{V}}$isual-$\underline{\text{T}}$houghts (**MoVT**), which unifies different reasoning modes within a single model and guides it to select the appropriate mode based on context. To achieve this, we introduce **AdaVaR**, a two-stage $\underline{\text{Ada}}$ptive $\underline{\text{V}}$isu$\underline{\text{a}}$l $\underline{\text{R}}$easoning learning framework: different modes are unified and learned during the supervised cold-start stage, and the mode selection capability is induced via an RL process with a carefully designed AdaGRPO algorithm. Extensive experiments show that AdaVaR effectively guides the model to learn and differentiate multiple modes and perform context-adaptive mode selection, achieving consistent improvement across various scenarios, highlighting MoVT as an effective solution for building general visual reasoning models.

Next-token prediction serves as the foundational learning task that enables reasoning in LLMs. But what should the learning task be when aiming to equip MLLMs with temporal reasoning capabilities over video inputs? Existing tasks such as video captioning primarily promote modality alignment, while video question answering typically relies on annotations from humans or much stronger MLLMs. To address this gap, we propose next-event prediction (NEP), a learning task that harnesses future video segments as a rich, self-supervised signal to foster temporal reasoning. We segment each video into past and future frames: the MLLM takes the past frames as input and predicts events in the future, thereby encouraging the model to reason temporally in order to complete the task. To study this learning task, we curate V1-33K, a dataset comprising 33,000 automatically extracted videos spanning diverse real-world scenarios. Using the same videos, we further explore a range of video instruction-tuning tasks data to provide controlled comparisons and isolate the effect of NEP. To evaluate progress, we introduce FutureBench to assess coherence in predicting unseen future events. Experiments validate that NEP offers a scalable and effective training task for fostering temporal reasoning in MLLMs.

Web agents such as deep research have demonstrated superhuman cognitive abilities, capable of solving highly challenging information-seeking problems. However, most research remains largely text-centric, overlooking visual information in the real world. This makes multimodal deep research highly challenging, as such agents require much stronger perceptual, logical, and knowledge-based reasoning abilities, as well as proficiency in more sophisticated tools. To address this limitation, we introduce WebWatcher, a multimodal agent for deep research with joint reasoning ability across both visual and textual modalities. It uses high-quality synthetic trajectories for efficient cold start training, utilizes various tools for deep reasoning, and further enhances generalization through reinforcement learning. To better evaluate the capabilities of multimodal agents, we propose BrowseComp-VL, a benchmark with the style of BrowseComp that requires complex information retrieval involving both visual and textual information. Experimental results show that WebWatcher outperforms the prompt-based workflow and open-source agents on HLE and BrowseComp-VL, and demonstrates its perception, multimodal reasoning, and searching capabilities across the other three benchmarks, respectively.

Spatial embodied intelligence often operates under partial observability, where agents must act to acquire missing information rather than passively consume complete observations. In such settings, progress depends on actively selecting informative actions that reduce uncertainty and support the construction of spatial understanding. While multimodal foundation models have shown strong performance on passive multimodal perception and reasoning tasks, their ability to support active, self-directed exploration under partial observability has not been systematically studied. In particular, it remains unclear whether and how these models can decide what to observe next in order to build and maintain a coherent spatial belief over time. We therefore propose \tos, defined as an agent’s ability to actively acquire information through self-directed, active exploration and to construct, revise, and exploit a spatial belief from sequential, partial observations. We implement \tos using a benchmark with textual and visual environments. Rather than solving specific tasks, the goal is curiosity-driven exploration to build a complete, accurate spatial belief. A core innovation is spatial belief probing: we prompt it to reveal its internal spatial belief as a cognitive map at each step, letting us measure the quality of its underlying spatial model. Our evaluation of state-of-the-art models on a suite of downstream tasks reveals critical bottlenecks: (1) \textbf{The Active-Passive Gap}: Performance degrades when agents must autonomously gather information (e.g., \textsc{GPT-5.2}: $0.57{\to}0.46$); (2) \textbf{Inefficiency}: Models explore in an unsystematic way and with high redundancy, failing to match the efficiency of program-based proxies while producing no better results. Through belief probing, we diagnose that perception acts as an initial bottleneck, yet global beliefs suffer further from \textbf{instability} that causes spatial knowledge to degrade over time. Finally, using a false belief paradigm to test belief revision, we uncover \textbf{Belief Inertia} where agents fail to overwrite obsolete priors. This issue exists in text agents but is notably severe in vision-based models.

In this work, we propose FFDP, a set of IO-aware non-GEMM fused kernels supplemented with a distributed framework for image registration at unprecedented scales. Image registration is an inverse problem fundamental to biomedical and life sciences, but algorithms have not scaled in tandem with image acquisition capabilities. Our framework complements existing model parallelism techniques proposed for large-scale transformer training by optimizing non-GEMM bottlenecks and enabling convolution-aware tensor sharding. We demonstrate unprecedented capabilities by performing multimodal registration of a 100μm ex-vivo human brain MRI volume at native resolution – an inverse problem more than 570× larger than a standard clinical datum in about a minute using only 8 A6000 GPUs. FFDP accelerates existing state-of-the-art optimization and deep learning registration pipelines by upto 6 − 7× while reducing peak memory consumption by 20 − 59%. Comparative analysis on a 250μm dataset shows that FFDP can fit upto 64× larger problems than existing SOTA on a single GPU, and highlights both the performance and efficiency gains of FFDP compared to SOTA image registration methods.

Despite rapid progress in multimodal large language models (MLLMs), their capability for deep emotional understanding remains limited. We argue that genuine affective intelligence requires explicit modeling of Theory of Mind (ToM), the cognitive substrate from which emotions arise. To this end, we introduce HitEmotion, a ToM-grounded hierarchical benchmark that diagnoses capability breakpoints across increasing levels of cognitive depth. Second, we propose a ToM-guided reasoning chain that tracks mental states and calibrates cross-modal evidence to achieve faithful emotional reasoning. We further introduce TMPO, a reinforcement learning method that uses intermediate mental states as process-level supervision to guide and strengthen model reasoning. Extensive experiments show that HitEmotion exposes deep emotional reasoning deficits in state-of-the-art models, especially on cognitively demanding tasks. In evaluation, the ToM-guided reasoning chain and TMPO improve end-task accuracy and yield more faithful, more coherent rationales. In conclusion, our work provides the research community with a practical toolkit for evaluating and enhancing the cognition-based emotional understanding capabilities of MLLMs.

Computational imaging methods increasingly rely on powerful generative diffusion models to tackle challenging image restoration tasks. In particular, state-of-the-art zero-shot image inverse solvers leverage distilled text-to-image latent diffusion models (LDMs) to achieve unprecedented accuracy and perceptual quality with high computational efficiency. However, extending these advances to high-definition video restoration remains a significant challenge, due to the need to recover fine spatial detail while capturing subtle temporal dependencies. Consequently, methods that naively apply image-based LDM priors on a frame-by-frame basis often result in temporally inconsistent reconstructions. We address this challenge by leveraging recent advances in Video Consistency Models (VCMs), which distill video latent diffusion models into fast generators that explicitly capture temporal causality. Building on this foundation, we propose LVTINO, the first zero-shot or plug-and-play inverse solver for high definition video restoration with priors encoded by VCMs. Our conditioning mechanism bypasses the need for automatic differentiation and achieves state-of-the-art video reconstruction quality with only a few neural function evaluations, while ensuring strong measurement consistency and smooth temporal transitions across frames. Extensive experiments on a diverse set of video inverse problems show significant perceptual improvements over current state-of-the-art methods that apply image LDMs frame by frame, establishing a new benchmark in both reconstruction fidelity and computational efficiency.

The Mixture-of-Experts (MoE) architecture has become a predominant paradigm for scaling large language models (LLMs). Despite offering strong performance and computational efficiency, large MoE-based LLMs like DeepSeek-V3-0324 and Kimi-K2-Instruct present serious challenges due to substantial memory requirements in deployment. While recent works have explored MoE compression to address this issue, existing methods often suffer from considerable accuracy drops (e.g., 7-14% relatively) even at modest compression rates. This paper introduces a novel Mixture-of-Basis-Experts (MoBE) method that achieves model compression while incurring minimal accuracy drops. Specifically, each up/gate matrix in an expert is decomposed via a rank decomposition as W = AB, where matrix A is unique to each expert. The relatively larger matrix B is further reparameterized as a linear combination of basis matrices {Bi} shared across all experts within a given MoE layer. The factorization is learned by minimizing the reconstruction error relative to the original weight matrices. Experiments demonstrate that MoBE achieves notably lower accuracy drops compared to prior works. For instance, MoBE can reduce the parameter counts of Qwen3-235BA22B-2507, DeepSeek-V3-0324 (671B) and Kimi-K2-Instruct (1T) by 24%-30% with only 1%-2% accuracy drop (about 2% drops when measured relatively).

Context compression presents a promising approach for accelerating large language model (LLM) inference by compressing long contexts into compact representations.Current context compression methods predominantly rely on autoencoding tasks to train context-agnostic compression tokens to compress contextual semantics.While autoencoding tasks enable compression tokens to acquire compression capabilities, compression via autoencoding tasks creates a fundamental mismatch: the models are optimized for reconstruction that diverge from actual downstream tasks, thereby weakening the features more beneficial for real-world usage.We propose Semantic-Anchor Compression (SAC), a novel method that shifts from autoencoding task based compression to an architecture that is equipped with this compression capability $\textit{a priori}$. Instead of training models to compress contexts through autoencoding tasks, SAC directly selects so-called anchor tokens from the original context and aggregates contextual information into their key-value (KV) representations. By deriving representations directly from the contextual tokens, SAC eliminates the need for autoencoding training. To ensure compression performance while directly leveraging anchor tokens directly, SAC incorporates two key designs: (1) anchor embeddings that enable the compressor to identify critical tokens, and (2) bidirectional attention modification that allows anchor tokens to capture information from the entire context.Experimental results demonstrate that SAC consistently outperforms existing context compression methods across various compression ratios. On out-of-distribution evaluation using MRQA, SAC achieves 1 EM improvement at 5x compression over strong baselines,with increasing advantages at higher compression ratios.

Large Language Models (LLMs) have shown great potential in reasoning tasks through test-time scaling methods like self-consistency with majority voting. However, this approach often leads to diminishing returns in accuracy and high computational overhead. To address these challenges, we introduce Deep Think with Confidence (DeepConf), a simple yet powerful method that enhances both reasoning efficiency and performance at test time. DeepConf leverages model-internal confidence signals to dynamically filter out low-quality reasoning traces during or after generation. It requires no additional model training or hyperparameter tuning and can be seamlessly integrated into existing serving frameworks. We evaluate DeepConf across a variety of tasks and the latest open-source models, including Qwen3 and GPT-OSS series. Notably, on challenging benchmarks such as AIME 2025, DeepConf@512 achieves up to 99.9\% accuracy and reduces generated tokens by up to 84.7\% compared to full parallel thinking. Anonymous code is available at https://anonymous.4open.science/r/deepconf-anonymous-E4E3