2026-09-06
5 papers covered
RoboTok leverages internet-scale video collections to train dexterous robot policies by matching 3D hand-pose trajectories across varying viewpoints and scenes. By converting web videos into a continuously growing source of supervision, it significantly improves downstream robot manipulation performance. To fix the lack of standardized benchmarks in speech brain-computer interfaces, researchers developed Open-Vocabulary Mutual Information (OVMI). This unified metric enables fair comparisons across heterogeneous systems and optimizes vocabulary selection for up to a 16.3% accuracy boost. VeriPhy evaluates physical realism in generated videos by translating text prompts into clear, auditable physical rules checked by specialized tools. Instead of returning a vague quality score, it provides traceable evidence records that pin down exact physical failures in space and time. Reinforcement learning with verifiable rewards boosts single-sample accuracy but severely restricts solution diversity by locking model reasoning at the very first step. Researchers found that alternative paths remain fully executable, and targeted early-layer interventions can restore solution coverage by 37% without sacrificing performance. DRACO solves the credit assignment problem in long-horizon AI agent training where automatic programmatic verifiers are unavailable. By dynamically creating multi-criteria rubrics and redistributing final outcomes into step-by-step rewards, it outperforms traditional ground-truth training on complex agent benchmarks.