PROOF-Gen Recovers Golden Trajectories from Failed Tool-Calling Trials
August 26, 2026
PROOF-Gen optimizes data distillation by using a reflector to analyze execution traces and feedback from failed teacher trials. Instead of discarding the 57% of failed attempts in tau2-bench, it generates corrective guidance to produce high-quality golden trajectories for student model training.
HOW THIS AFFECTS YOU
●
builderYou can reduce frontier-teacher costs and improve agent performance by recycling failed distillation attempts.
●
researcherThis provides a method for extracting signal from teacher-model failures during post-training.