Reasoning-based large language models have excelled in mathematics and programming, yet their potential in knowledge-intensive medical question answering remains underexplored and insufficiently validated in clinical contexts. To bridge this gap, we introduce ReasonMed, the largest medical reasoning dataset to date, comprising 370k high-quality examples distilled from 1.75 million initial reasoning paths generated by complementary LLMs and curated through a cost-efficient easy-medium-difficult (EMD) pipeline. ReasonMed is built through a multi-agent generation, verification, and refinement process, in which an Error Refiner improves reasoning paths by correcting error-prone steps identified by a verifier. Using ReasonMed, we investigate effective strategies for training medical reasoning models and find that integrating detailed CoT reasoning with concise answer summaries yields the most robust fine-tuning results. Models trained on ReasonMed set a new benchmark: ReasonMed-7B surpasses the prior best sub-10B models by 4.17% and even exceeds LLaMA3.1-70B on PubMedQA by 4.60%. When scaled to ReasonMed-14B, it remains highly competitive, underscoring consistent scaling potential. The codes and datasets are available at https://github.com/YuSun-Work/ReasonMed.

Modern neural models capture rich priors and have complementary knowledge over shared data domains, e.g., images and videos. Integrating diverse knowledge from multiple sources -- including visual generative models, visual language models, and sources with human-crafted knowledge such as graphics engines and physics simulators -- remains under-explored. We propose a Product of Experts (PoE) framework that performs inference-time knowledge composition from heterogeneous models. This training-free approach samples from the product distribution across experts via Annealed Importance Sampling (AIS). Our framework shows practical benefits in image and video synthesis tasks, yielding better controllability than monolithic methods and additionally providing flexible user interfaces for specifying visual generation goals.

Developing reasoning capabilities in multimodal large language models (MLLMs) remains challenging. Motivated by literature suggesting that gameplay promotes transferable reasoning skills, we propose a novel post-training method, Visual Game Learning (ViGaL), where MLLMs develop generalizable reasoning skills through playing arcade-like games. Specifically, we show that training a 7B-parameter MLLM via reinforcement learning (RL) on simple games like Snake significantly enhances the downstream performance on multimodal math benchmarks like MathVista, on multi-discipline questions like MMMU and on 3D spatial reasoning benchmarks like VSI-Bench, without seeing any worked solutions, equations, or diagrams during RL. Remarkably, our model outperforms specialist models post-trained on benchmark-oriented multimodal reasoning data, while preserving the model's performance on general visual benchmarks, a challenge where specialist models often fall short. Our findings suggest that multimodal reasoning can emerge from gameplay, pointing to a promising strategy of designing surrogate tasks for RL post-training.

The ability to infer what others are looking at is a critical component of a theory of mind that underpins natural human-AI interaction. We characterized this skill in 111 Vision Language Models (VLMs) and human participants (N = 65) using photos taken with manipulated difficulty and variability. We found that 94 of the 111 VLMs were not better than random guessing, while humans achieved near-ceiling accuracy. VLMs respond with each choice almost equally frequently. Are they randomly guessing? At least for five top-tier VLMs, their performance was above chance, declined with increasing task difficulty, but barely varied across different prompts and scene objects. These behavioral patterns cannot be explained by considering VLMs as random guessers. Instead, they likely utilize head orientation but not eye appearance to infer gaze direction, such that their performance is imperfect, subject to the task difficulty, but robust to superficial perceptual variations. This suggests that VLMs, lacking effective gaze inference skills, have yet to become technologies that can naturally interact with humans, but the potential remains.

Instance segmentation demands costly per-pixel annotations and computationally expensive models. We introduce CAST, a semi-supervised knowledge distillation (SSKD) framework that compresses pre-trained vision foundation models (VFM) into compact experts using limited labeled and abundant unlabeled data. CAST unfolds in three stages: (1) domain adaptation of the VFM(s) via self-training with contrastive calibration, (2) knowledge transfer through a unified multi-objective loss, and (3) student refinement to mitigate residual pseudo-label bias. Central to CAST is an \emph{instance-aware pixel-wise contrastive loss} that fuses mask and class scores to extract informative negatives and enforce clear inter-instance margins. By maintaining this contrastive signal across both adaptation and distillation, we align teacher and student embeddings and fully leverage unlabeled images. On Cityscapes and ADE20K, our ~11x smaller student improves over its zero-shot VFM teacher(s) by +8.5 and +7.1 AP, surpasses adapted teacher(s) by +3.4 and +1.5 AP, and further outperforms state-of-the-art SSKD methods on both benchmarks.

Recent successes of reinforcement learning (RL) in training large reasoning models motivate the question of whether self-training - the process where a model learns from its own judgments - can be sustained within RL. In this work, we study this question using majority voting as a simple self-feedback mechanism. On a comprehensive set of experiments on both synthetic and real reasoning tasks, we find that this basic approach improves not only the model's reasoning performance, but also its capability of generating better quality feedback for the next RL iteration, driving further model improvement. Yet our analysis also reveals a critical limitation of such a self-training paradigm - prolonged RL with self-reward leads to reward hacking where models learn to maximize training (pseudo-)reward, resulting in sudden and complete performance collapse. Together, these results highlight feedback design as the central challenge and call for future research on mechanisms to enable prolonged self-improvement.

With the surging growth of model parameters, foundation models pose unprecedented challenges to traditional computational infrastructures.

Recent advancements in large language models (LLMs) focus on aligning to heterogeneous human expectations and values via multi-objective preference alignment.