TurnSight: Turn-Level Hindsight Self-Distillation for Tool-Integrated Reasoning
TurnSight assigns learning signals at individual tool-use turns using execution hindsight, targeting the weak credit assignment of trajectory-level reinforcement learning.
TurnSight derives supervision from states an agent actually visited, builds hindsight views with **multiple lookahead horizons**, and selects signals when their directions agree. It evaluates the method on **three benchmarks**.
For teams training tool-using agents, preserve turn structure in the learning signal and use execution traces for credit assignment. Normalizing across sibling rollouts can adjust RL advantages without reversing their original direction.
TurnSight derives supervision from states an agent actually visited, builds hindsight views with **multiple lookahead horizons**, and selects signals when their directions agree. It evaluates the method on **three benchmarks**. For teams training tool-using agents, preserve turn structure in the learning signal and use execution traces for credit assignment. Normalizing across sibling rollouts can adjust RL advantages without reversing their original direction. The material reports effectiveness but gives no model names, score deltas, compute costs, or failure breakdowns. The released code enables inspection, but the abstract alone does not establish how well the method transfers to coding-agent workloads.
TurnSight makes execution traces more actionable for training by assigning supervision at the turn level and requiring agreement across hindsight horizons. It directly advances the unresolved self-distillation and credit-assignment problem in prior candidates, but does not yet show that gains survive the longer, less replayable, production coding workflows those candidates describe.