MedGR$^2$: Breaking the Data Barrier for Medical Reasoning via Generative Reward Learning

The recent introduction of TrajMoE, a scene-adaptive trajectory planning framework, leverages a Mixture of Experts (MoE) architecture combined with Reinforcement Learning to enhance trajectory evaluation in autonomous driving. This approach addresses the variability of trajectory priors across different driving scenarios and improves the scoring mechanism through policy-driven refinement.

A new framework named BiMi has been introduced to enhance the detection and localization of bilingual multimodal misinformation, particularly in news media where images are often paired with bilingual subtitles. This framework addresses the challenges posed by localized image edits and cross-lingual inconsistencies that can distort meaning while appearing plausible.

A new framework called Speculative Verdict (SV) has been introduced to enhance the reasoning capabilities of Vision-Language Models (VLMs) when dealing with complex, information-rich images. SV operates in two stages: the draft stage, where small VLMs generate diverse reasoning paths, and the verdict stage, where a stronger VLM synthesizes these paths to produce accurate answers efficiently.

A novel reward mechanism named COMPASS has been introduced to enhance test-time reinforcement learning (RL) for large language models (LLMs). This mechanism allows models to autonomously learn from unlabeled data, addressing the scalability challenges faced by traditional RL methods that rely heavily on human-curated data for reward modeling.

Recent advancements in Vision-Language Models (VLMs) have led to the development of Training-free Dual Hyperbolic Adapters (T-DHA), a novel adaptation method that enhances cross-modal reasoning without requiring extensive training resources. This method utilizes hyperbolic space to better represent hierarchical relationships between semantic concepts, improving both representation and discrimination capabilities.

A new benchmark called Tri-Bench has been introduced to assess the reliability of Vision-Language Models (VLMs) in spatial reasoning tasks, particularly under conditions of camera tilt and object interference. The benchmark evaluates four recent VLMs using a fixed prompt and measures their accuracy against 3D ground truth, revealing an average accuracy of approximately 69%.

The introduction of OS-Sentinel marks a significant advancement in enhancing the safety of mobile GUI agents powered by Vision-Language Models (VLMs). This framework aims to address critical safety concerns, such as system compromise and privacy leakage, by utilizing a hybrid validation approach within a dynamic sandbox environment called MobileRisk-Live, which includes realistic operational trajectories with detailed annotations.

Recent research has identified a significant flaw in Large Vision-Language Models (LVLMs), revealing that these models often reach correct answers through incorrect reasoning paths. This issue stems from a path selection bias within the reasoning search space, leading to unreliable outcomes despite the models' knowledge of the correct answers. The proposed Path-Select Optimization (PSO) framework aims to enhance reasoning performance and stability in LVLMs.