Recent advances in large reasoning models (LRMs) demonstrate that sophisticated behaviors such as multi-step reasoning and self-reflection can emerge via reinforcement learning with verifiable rewards~(\textit{RLVR}). However, existing \textit{RLVR} approaches are inherently ``on-policy'', limiting learning to a model's own outputs and failing to acquire reasoning abilities beyond its initial capabilities. To address this issue, we introduce \textbf{LUFFY} (\textbf{L}earning to reason \textbf{U}nder o\textbf{FF}-polic\textbf{Y} guidance), a framework that augments \textit{RLVR} with off-policy reasoning traces. LUFFY dynamically balances imitation and exploration by combining off-policy demonstrations with on-policy rollouts during training. Specifically, LUFFY combines the Mixed-Policy GRPO framework, which has a theoretically guaranteed convergence rate, alongside policy shaping via regularized importance sampling to avoid superficial and rigid imitation during mixed-policy training. Compared with previous RLVR methods, LUFFY achieves an over \textbf{+6.4} average gain across six math benchmarks and an advantage of over \textbf{+6.2} points in out-of-distribution tasks. Most significantly, we show that LUFFY successfully trains weak models in scenarios where on-policy RLVR completely fails. These results provide compelling evidence that LUFFY transcends the fundamental limitations of on-policy RLVR and demonstrates the great potential of utilizing off-policy guidance in RLVR.

We present the Multi-Agent Transformer World Model (MATWM), a novel transformer-based world model designed for multi-agent reinforcement learning in both vector- and image-based environments. MATWM combines a decentralized imagination framework with a semi-centralized critic and a teammate prediction module, enabling agents to model and anticipate the behavior of others under partial observability. To address non-stationarity, we incorporate a prioritized replay mechanism that trains the world model on recent experiences, allowing it to adapt to agents' evolving policies. We evaluated MATWM on a broad suite of benchmarks, including the StarCraft Multi-Agent Challenge, PettingZoo, and MeltingPot. MATWM achieves state-of-the-art performance, outperforming both model-free and prior world model approaches, while demonstrating strong sample efficiency, achieving near-optimal performance in as few as 50K environment interactions. Ablation studies confirm the impact of each component, with substantial gains in coordination-heavy tasks.

Large vision-language models (VLMs) for autonomous driving (AD) are evolving beyond perception and cognition tasks toward motion planning. However, we identify two critical challenges in this direction: (1) VLMs tend to learn shortcuts by relying heavily on history input information, achieving seemingly strong planning results without genuinely understanding the visual inputs; and (2) the chain-ofthought (COT) reasoning processes are always misaligned with the motion planning outcomes, and how to effectively leverage the complex reasoning capability to enhance planning remains largely underexplored. In this paper, we start from a small-scale domain-specific VLM and propose Drive-R1 designed to bridges the scenario reasoning and motion planning for AD. Drive-R1 first undergoes the supervised finetuning on a elaborate dataset containing both long and short COT data. Drive-R1 is encouraged to reason step-by-step from visual input to final planning decisions. Subsequently, Drive-R1 is trained within a reinforcement learning framework that incentivizes the discovery of reasoning paths that are more informative for planning, guided by rewards based on predicted trajectories and meta actions. Experimental evaluations on the nuScenes and DriveLM-nuScenes benchmarks demonstrate that Drive-R1 achieves superior performance compared to existing state-of-the-art VLMs. We believe that Drive-R1 presents a promising direction for bridging reasoning and planning in AD, offering methodological insights for future research and applications.

We propose that reinforcement learning (RL) from partial expert demonstrations is not merely a training heuristic, but a promising framework for solving complex sequence generation tasks. Supervised fine-tuning (SFT) relies on dense ground-truth labels, which become increasingly costly as sequence length grows. RL, on the other hand, struggles with sparse rewards and a combinatorially large output space. We address this by introducing adaptive backtracking (AdaBack), a per-sample curriculum learning algorithm that reveals only a partial prefix of the target output during training. The supervision length is adjusted dynamically for each sample based on the model's past reward signal, allowing it to incrementally learn to complete reasoning chains by conditioning on correct partial solutions. We investigate this intermediate regime between SFT and RL and argue that per-sample curriculum learning is more than a trade-off between efficiency and generality, it can succeed in tasks with long sequences of latent dependencies where SFT and RL both fail to generalize. Using a synthetic task with latent parity constraints, we show that our adaptive curriculum over partial answers reliably solves problems that are otherwise intractable. On mathematical reasoning benchmarks (MATH, GSM8k), we find that curriculum learning enables models to solve problems that RL alone cannot, acquiring new reasoning capabilities through incremental exposure to partial solutions.

Test-time compute has empowered multimodal large language models to generate extended reasoning chains, yielding strong performance on tasks such as multimodal math reasoning. However, this improved reasoning ability often comes with increased hallucination: as generations become longer, models tend to drift away from image-grounded content and rely more heavily on language priors. Attention analysis shows that longer reasoning chains lead to reduced focus on visual inputs, which contributes to hallucination. To systematically study this phenomenon, we introduce RH-AUC, a metric that quantifies how a model's perception accuracy changes with reasoning length, allowing us to evaluate whether the model preserves visual grounding during reasoning. We also release RH-Bench, a diagnostic benchmark that spans a variety of multimodal tasks, designed to assess the trade-off between reasoning ability and hallucination. Our analysis reveals that (i) larger models typically achieve a better balance between reasoning and perception, and (ii) this balance is influenced more by the types and domains of training data than by its overall volume. These findings underscore the importance of evaluation frameworks that jointly consider both reasoning quality and perceptual fidelity.

As AI capabilities advance toward and potentially beyond human-level performance, a natural transition emerges where AI-driven development becomes more efficient than human-centric approaches. A promising pathway toward this transition lies in AI-for-AI (AI4AI), which leverages AI techniques to automate and optimize the design, training, and deployment of AI systems themselves. While LLM-based agents have shown the potential to realize AI4AI, they are often unable to fully leverage the experience accumulated by agents during the exploration of solutions in the reasoning process, leading to inefficiencies and suboptimal performance. To address this limitation, we propose ML-Master, a novel AI4AI agent that seamlessly integrates exploration and reasoning by employing a selectively scoped memory mechanism. This approach allows ML-Master to efficiently combine diverse insights from parallel solution trajectories with analytical reasoning, guiding further exploration without overwhelming the agent with excessive context. We evaluate ML-Master on the MLE-Bench, where it achieves a 29.3% average medal rate, significantly surpassing existing methods, particularly in medium-complexity tasks, while accomplishing this superior performance within a strict 12-hour time constraint-half the 24-hour limit used by previous baselines. These results demonstrate ML-Master's potential as a powerful tool for advancing AI4AI.

Small language models (SLMs) struggle to learn complex reasoning behaviors, especially when high-quality traces are scarce or difficult to learn from. The standard training approach combines a supervised fine-tuning (SFT) stage, often to distill capabilities of a larger model, followed by a reinforcement learning (RL)stage such as Group Relative Policy Optimization (GRPO). In this paper, we investigate the fundamental limitations of this SFT + RL paradigm and propose methods to overcome them. Under a suitable theoretical model, we demonstrate that the SFT + RL strategy can fail completely when (1) the expert's traces are too difficult for the small model to express, or (2) the small model's initialization has exponentially small likelihood of success. To address these, we introduce BREAD: a GRPO variant that unifies the SFT and RL stages via partial expert guidance and branched rollouts. When self-generated traces fail, BREAD adaptively inserts short expert prefixes/hints, allowing the small model to complete the rest of the reasoning path, and ensuring that each update includes at least one successful trace. This mechanism both densifies the reward signal and induces a natural learning curriculum. BREAD requires fewer than 40% of ground-truth traces, consistently outperforming standard GRPO while speeding up the training by about 3 times. Importantly, we demonstrate that BREAD helps the model solve problems that are otherwise unsolvable by the SFT + RL strategy, highlighting how branched rollouts and expert guidance can substantially boost SLM reasoning.

As textual reasoning with large language models (LLMs) has advanced significantly, there has been growing interest in enhancing the multimodal reasoning capabilities of large vision-language models (LVLMs). However, existing methods primarily approach multimodal reasoning in a straightforward, text-centric manner, where both reasoning and answer derivation are conducted purely through text, with the only difference being the presence of multimodal input. As a result, these methods often encounter fundamental limitations in spatial reasoning tasks that demand precise geometric understanding and continuous spatial tracking-capabilities that humans achieve through mental visualization and manipulation. To address the limitations, we propose drawing to reason in space, a novel paradigm that enables LVLMs to reason through elementary drawing operations in the visual space. By equipping models with basic drawing operations, including annotating bounding boxes and drawing auxiliary lines, we empower them to express and analyze spatial relationships through direct visual manipulation, meanwhile avoiding the performance ceiling imposed by specialized perception tools in previous tool-integrated reasoning approaches. To cultivate this capability, we develop a three-stage training framework: cold-start training with synthetic data to establish basic drawing abilities, reflective rejection sampling to enhance self-reflection behaviors, and reinforcement learning to directly optimize for target rewards. Extensive experiments demonstrate that our model, named VILASR, consistently outperforms existing methods across diverse spatial reasoning benchmarks, involving maze navigation, static spatial reasoning, video-based reasoning, and multi-view-based reasoning tasks, with an average improvement of 18.4%.

Vision-Language Models (VLMs) acquire real-world knowledge and general reasoning ability through Internet-scale image-text corpora. They can augment robotic systems with scene understanding and task planning, and assist visuomotor policies that are trained on robot trajectory data. We explore the reverse paradigm - using rich, real, multi-modal robot trajectory data to enhance and evaluate VLMs. In this paper, we present Robo2VLM, a Visual Question Answering (VQA) dataset generation framework for VLMs. Given a human tele-operated robot trajectory, Robo2VLM derives ground-truth from non-visual and non-descriptive sensory modalities, such as end-effector pose, gripper aperture, and force sensing. Based on these modalities, it segments the robot trajectory into a sequence of manipulation phases. At each phase, Robo2VLM uses scene and interaction understanding to identify 3D properties of the robot, task goal, and the target object. The properties are used to generate representative VQA queries - images with textural multiple-choice questions - based on spatial, goal-conditioned, and interaction reasoning question templates. We curate Robo2VLM-1, a large-scale in-the-wild dataset with 684,710 questions covering 463 distinct scenes and 3,396 robotic manipulation tasks from 176k real robot trajectories. Results suggest that Robo2VLM-1 can benchmark and improve VLM capabilities in spatial and interaction reasoning.

Large language models (LLMs) have demonstrated impressive mathematical reasoning capabilities but remain susceptible to hallucinations--plausible yet incorrect statements--particularly in complex domains requiring rigorous logical deduction. Current approaches to improve reliability often neglect the logical consistency of intermediate reasoning steps, focusing primarily on final answer verification. We propose a structured self-consistency (SC) framework that systematically evaluates factual concordance across both intermediate reasoning steps and final outputs, thereby creating a hierarchical verification mechanism for mathematical reasoning. Our framework employs a probabilistic formulation that quantifies consistency through ensemble agreement, entropy minimization, and structural isomorphism detection in reasoning graphs. We evaluate our approach on three fundamental mathematical tasks: formal theorem proving, symbolic transformation, and numerical computation. Experimental results demonstrate that our method achieves significant improvements over baseline approaches: proof validity increases by 8.3% ( p < 0. 01), symbolic reasoning accuracy by 9.6%, and numerical stability by 42.8% while reducing computational overhead by 56.3%. Further analysis reveals that our structured SC framework exhibits strong correlation with human expert evaluation ( ρ = 0 .87),