Daily digest
4 items · ~4 min · Week 2026-W33
Worth knowing (3)
AgentOPSD: Recursive Self-Distillation for Agentic Reinforcement Learning
Tsinghua UniversityIntroduces a critic-free RL method for multi-step agentic tasks that aggregates teacher-student probability differences into turn-level indicators and maintains a Bayesian belief state to assign credit to individual decision turns, turning sparse outcome rewards into granular signals.
OSReward: Instituting Standardized Evaluation for Cross-Platform Computer-Use Reward Models
NLP Group of The University of Hong KongProposes a benchmark for judging computer-using agent trajectories with vision-language models, finds even state-of-the-art models show a systematic leniency bias that mislabels failures as successes, and releases the OS-Shepherd-100K dataset plus 9B/35B open reward models matching commercial alternatives at lower cost.
WorldClaw: Agentic 3D Open-World Generation at Scale
Tencent HunyuanPresents a coarse-to-fine agentic framework where planning agents convert text prompts into structured specifications, then build large explorable 3D scenes with coherent terrain, reusable assets, and editable instance-level meshes, refined further by render-based agents.
For reference (1)
GST-Bench: Can VLMs Develop Global Spatial Awareness from Video?
ByteDance SeedIntroduces a video question-answering benchmark built from nearly 6,800 minutes of synthetic video to test whether vision-language models can integrate spatial observations across a video into a coherent global representation, testing 22 state-of-the-art models.