HuggingFace Daily Papers· Fangcong Yin, Ming Yin, Tanvi Bhandarkar, Aashu Singh, Eryk Helenowski, Shlok Kumar Mishra, Saghar Hosseini, Karthik Abinav Sankararaman, Greg Durrett, Dinkar Juyal·· 6 天前AI 评分55
Self-Credit: 通过组内重缩放为长时程智能体RL分配信用
Self-Credit: Credit Assignment for Long-Horizon Agentic RL via Intra-group Rescaling
AI 导读
研究人员提出Self-Credit方法,通过组内重缩放机制为长时程智能体强化学习分配信用。该方法在BrowseComp-Plus上比基线模型提升37.0和36.2分,在GAIA和xBench-DeepSearch上也分别超越最强基线4.9和4.7分。
来源:HuggingFace Daily Papers · arxiv.org