Advancing Multimodal Reasoning via Reinforcement Learning with Cold Start

Recent advancements in large language models (LLMs) have demonstrated impressive chain-of-thought reasoning capabilities, with reinforcement learning (RL) playing a crucial role in this progress. While "aha moment" patterns--where models exhibit self-correction through reflection--are often attributed to emergent properties from RL, we first demonstrate that these patterns exist in multimodal LLMs (MLLMs) prior to RL training but may not necessarily correlate with improved reasoning performance. Building on these insights, we present a comprehensive study on enhancing multimodal reasoning through a two-stage approach: (1) supervised fine-tuning (SFT) as a cold start with structured chain-of-thought reasoning patterns, followed by (2) reinforcement learning via GRPO to further refine these capabilities. Our extensive experiments show that this combined approach consistently outperforms both SFT-only and RL-only methods across challenging multimodal reasoning benchmarks. The resulting models achieve state-of-the-art performance among open-source MLLMs at both 3B and 7B scales, with our 7B model showing substantial improvements over base models (e.g., 66.3 %$\rightarrow$73.4 % on MathVista, 62.9 %$\rightarrow$70.4 % on We-Math) and our 3B model achieving performance competitive with several 7B models. Overall, this work provides practical guidance for building advanced multimodal reasoning models. Our code is available at https://github.com/waltonfuture/RL-with-Cold-Start.

Unsupervised Post-Training for Multi-Modal LLM Reasoning via GRPO

Improving Multi-modal Large Language Models (MLLMs) in the post-training stage typically relies on supervised fine-tuning (SFT) or reinforcement learning (RL). However, these supervised methods require expensive and manually annotated multi-modal data--an ultimately unsustainable resource. While recent efforts have explored unsupervised post-training, their methods are complex and difficult to iterate. In this work, we are the first to investigate the use of GRPO, a stable and scalable online RL algorithm, for enabling continual self-improvement without any external supervision. We propose MM-UPT, a simple yet effective framework for unsupervised post-training of MLLMs. MM-UPT builds upon GRPO, replacing traditional reward signals with a self-rewarding mechanism based on majority voting over multiple sampled responses. Our experiments demonstrate that MM-UPT significantly improves the reasoning ability of Qwen2.5-VL-7B (e.g., 66.3 %$\rightarrow$72.9 % on MathVista, 62.9 %$\rightarrow$68.7 % on We-Math), using standard dataset without ground truth labels. MM-UPT also outperforms prior unsupervised baselines and even approaches the results of supervised GRPO. Furthermore, we show that incorporating synthetic questions, generated solely by MLLM itself, can boost performance as well, highlighting a promising approach for scalable self-improvement. Overall, MM-UPT offers a new paradigm for continual, autonomous enhancement of MLLMs in the absence of external supervision. Our code is available at https://github.com/waltonfuture/MM-UPT.

Pessimism Principle Can Be Effective: Towards a Framework for Zero-Shot Transfer Reinforcement Learning

Transfer reinforcement learning aims to derive a near-optimal policy for a target environment with limited data by leveraging abundant data from related source domains. However, it faces two key challenges: the lack of performance guarantees for the transferred policy, which can lead to undesired actions, and the risk of negative transfer when multiple source domains are involved. We propose a novel framework based on the pessimism principle, which constructs and optimizes a conservative estimation of the target domain's performance. Our framework effectively addresses the two challenges by providing an optimized lower bound on target performance, ensuring safe and reliable decisions, and by exhibiting monotonic improvement with respect to the quality of the source domains, thereby avoiding negative transfer. We construct two types of conservative estimations, rigorously characterize their effectiveness, and develop efficient distributed algorithms with convergence guarantees. Our framework provides a theoretically sound and practically robust solution for transfer learning in reinforcement learning.

U2-BENCH: Benchmarking Large Vision-Language Models on Ultrasound Understanding

Ultrasound is a widely-used imaging modality critical to global healthcare, yet its interpretation remains challenging due to its varying image quality on operators, noises, and anatomical structures. Although large vision-language models (LVLMs) have demonstrated impressive multimodal capabilities across natural and medical domains, their performance on ultrasound remains largely unexplored. We introduce U2-BENCH, the first comprehensive benchmark to evaluate LVLMs on ultrasound understanding across classification, detection, regression, and text generation tasks. U2-BENCH aggregates 7,241 cases spanning 15 anatomical regions and defines 8 clinically inspired tasks, such as diagnosis, view recognition, lesion localization, clinical value estimation, and report generation, across 50 ultrasound application scenarios. We evaluate 20 state-of-the-art LVLMs, both open- and closed-source, general-purpose and medical-specific. Our results reveal strong performance on image-level classification, but persistent challenges in spatial reasoning and clinical language generation. U2-BENCH establishes a rigorous and unified testbed to assess and accelerate LVLM research in the uniquely multimodal domain of medical ultrasound imaging.

Two Experts Are All You Need for Steering Thinking: Reinforcing Cognitive Effort in MoE Reasoning Models Without Additional Training

Mixture-of-Experts (MoE) architectures within Large Reasoning Models (LRMs) have achieved impressive reasoning capabilities by selectively activating experts to facilitate structured cognitive processes. Despite notable advances, existing reasoning models often suffer from cognitive inefficiencies like overthinking and underthinking. To address these limitations, we introduce a novel inference-time steering methodology called Reinforcing Cognitive Experts (RICE), designed to improve reasoning performance without additional training or complex heuristics. Leveraging normalized Pointwise Mutual Information (nPMI), we systematically identify specialized experts, termed ''cognitive experts'' that orchestrate meta-level reasoning operations characterized by tokens like ''<think>''. Empirical evaluations with leading MoE-based LRMs (DeepSeek-R1 and Qwen3-235B) on rigorous quantitative and scientific reasoning benchmarks demonstrate noticeable and consistent improvements in reasoning accuracy, cognitive efficiency, and cross-domain generalization. Crucially, our lightweight approach substantially outperforms prevalent reasoning-steering techniques, such as prompt design and decoding constraints, while preserving the model's general instruction-following skills. These results highlight reinforcing cognitive experts as a promising, practical, and interpretable direction to enhance cognitive efficiency within advanced reasoning models.

DGRO: Enhancing LLM Reasoning via Exploration-Exploitation Control and Reward Variance Management

Inference scaling further accelerates Large Language Models (LLMs) toward Artificial General Intelligence (AGI), with large-scale Reinforcement Learning (RL) to unleash long Chain-of-Thought reasoning. Most contemporary reasoning approaches usually rely on handcrafted rule-based reward functions. However, the tarde-offs of exploration and exploitation in RL algorithms involves multiple complex considerations, and the theoretical and empirical impacts of manually designed reward functions remain insufficiently explored. In this paper, we propose Decoupled Group Reward Optimization (DGRO), a general RL algorithm for LLM reasoning. On the one hand, DGRO decouples the traditional regularization coefficient into two independent hyperparameters: one scales the policy gradient term, and the other regulates the distance from the sampling policy. This decoupling not only enables precise control over balancing exploration and exploitation, but also can be seamlessly extended to Online Policy Mirror Descent (OPMD) algorithms in Kimi k1.5 and Direct Reward Optimization. On the other hand, we observe that reward variance significantly affects both convergence speed and final model performance. We conduct both theoretical analysis and extensive empirical validation to assess DGRO, including a detailed ablation study that investigates its performance and optimization dynamics. Experimental results show that DGRO achieves state-of-the-art performance on the Logic dataset with an average accuracy of 96.9\%, and demonstrates strong generalization across mathematical benchmarks.

A Finite-Sample Analysis of Distributionally Robust Average-Reward Reinforcement Learning

Robust reinforcement learning (RL) under the average-reward criterion is crucial for long-term decision making under potential environment mismatches, yet its finite-sample complexity study remains largely unexplored. Existing works offer algorithms with asymptotic guarantees, but the absence of finite-sample analysis hinders its principled understanding and practical deployment, especially in data-limited settings. We close this gap by proposing Robust Halpern Iteration (RHI), the first algorithm with provable finite-sample complexity guarantee. Under standard uncertainty sets -- including contamination sets and $\ell_p$-norm balls -- RHI attains an $\epsilon$-optimal policy with near-optimal sample complexity of $\tilde{\mathcal O}\left(\frac{SA\mathcal H^{2}}{\epsilon^{2}}\right)$, where $S$ and $A$ denote the numbers of states and actions, and $\mathcal H$ is the robust optimal bias span. This result gives the first polynomial sample complexity guarantee for robust average-reward RL. Moreover, our RHI's independence from prior knowledge distinguishes it from many previous average-reward RL studies. Our work thus constitutes a significant advancement in enhancing the practical applicability of robust average-reward methods to complex, real-world problems.

MSCI: Addressing CLIP's Inherent Limitations for Compositional Zero-Shot Learning

Compositional Zero-Shot Learning (CZSL) aims to recognize unseen state-object combinations by leveraging known combinations. Existing studies basically rely on the cross-modal alignment capabilities of CLIP but tend to overlook its limitations in capturing fine-grained local features, which arise from its architectural and training paradigm. To address this issue, we propose a Multi-Stage Cross-modal Interaction (MSCI) model that effectively explores and utilizes intermediate-layer information from CLIP's visual encoder. Specifically, we design two self-adaptive aggregators to extract local information from low-level visual features and integrate global information from high-level visual features, respectively. These key information are progressively incorporated into textual representations through a stage-by-stage interaction mechanism, significantly enhancing the model's perception capability for fine-grained local visual information. Additionally, MSCI dynamically adjusts the attention weights between global and local visual information based on different combinations, as well as different elements within the same combination, allowing it to flexibly adapt to diverse scenarios. Experiments on three widely used datasets fully validate the effectiveness and superiority of the proposed model. Data and code are available at https://github.com/ltpwy/MSCI.

ManipBench: Benchmarking Vision-Language Models for Low-Level Robot Manipulation

Vision-Language Models (VLMs) have revolutionized artificial intelligence and robotics due to their commonsense reasoning capabilities. In robotic manipulation, VLMs are used primarily as high-level planners, but recent work has also studied their lower-level reasoning ability, which refers to making decisions about precise robot movements. However, the community currently lacks a clear and common benchmark that can evaluate how well VLMs can aid low-level reasoning in robotics. Consequently, we propose a novel benchmark, ManipBench, to evaluate the low-level robot manipulation reasoning capabilities of VLMs across various dimensions, including how well they understand object-object interactions and deformable object manipulation. We extensively test 33 representative VLMs across 10 model families on our benchmark, including variants to test different model sizes. Our evaluation shows that the performance of VLMs significantly varies across tasks, and there is a strong correlation between this performance and trends in our real-world manipulation tasks. It also shows that there remains a significant gap between these models and human-level understanding. See our website at: https://manipbench.github.io.

HMCF: A Human-in-the-loop Multi-Robot Collaboration Framework Based on Large Language Models

Rapid advancements in artificial intelligence (AI) have enabled robots to performcomplex tasks autonomously with increasing precision. However, multi-robot systems (MRSs) face challenges in generalization, heterogeneity, and safety, especially when scaling to large-scale deployments like disaster response. Traditional approaches often lack generalization, requiring extensive engineering for new tasks and scenarios, and struggle with managing diverse robots. To overcome these limitations, we propose a Human-in-the-loop Multi-Robot Collaboration Framework (HMCF) powered by large language models (LLMs). LLMs enhance adaptability by reasoning over diverse tasks and robot capabilities, while human oversight ensures safety and reliability, intervening only when necessary. Our framework seamlessly integrates human oversight, LLM agents, and heterogeneous robots to optimize task allocation and execution. Each robot is equipped with an LLM agent capable of understanding its capabilities, converting tasks into executable instructions, and reducing hallucinations through task verification and human supervision. Simulation results show that our framework outperforms state-of-the-art task planning methods, achieving higher task success rates with an improvement of 4.76%. Real-world tests demonstrate its robust zero-shot generalization feature and ability to handle diverse tasks and environments with minimal human intervention.