IGSD: Environment-Verified Hindsight Self-Distillation for Search Agents
Information-Gain-Gated Self-Distillation (IGSD) is proposed, which verifies on-policy token proposals with environment feedback before distillation, and supports environment-verified hindsight as an effective approach to reliable action-level supervision for search agents.
Ang-Qing Jiang, Gao-Ming Zhang, Chao-Qun Zhang et al.
· 0 citations