TurnSight: Turn-Level Hindsight Self-Distillation for Tool Use
August 3, 2026
TurnSight improves Tool-Integrated Reasoning (TIR) by applying hindsight self-distillation at the turn level rather than the trajectory level. This provides denser supervision signals for agents performing long-horizon tasks through iterative tool interactions.
HOW THIS AFFECTS YOU
●
builderYou can build more reliable tool-using agents by training them on execution-conditioned hindsight.
●
researcherYou can improve fine-grained credit assignment in tool-using agents through turn-level distillation.