Supervised fine-tuning on teacher-generated trajectories is the standard first stage for distilling tool-calling capabilities into deployable models. Post-training pipelines that ship tool-calling agents re-run this stage on a daily or weekly cadence, paying the frontier-teacher cost each cycle, yet the mechanism is generate-and-filter: keep the teacher’s passing trajectories and discard the rest. Failures supply no signal, so the same hard scenarios are left behind every cycle. On τ2-bench, 57% of teacher trials fail, two-thirds of them near-misses where most tool calls are correct and a single decisive error breaks the run.

PROOF-Gen (Per-scenario Reflective Optimization to Overcome Failed Generation) recovers golden trajectories from these failures via per-scenario prompt optimization. For each failed task, a reflector analyzes the execution trace and evaluation feedback, then writes corrective guidance that steers the teacher toward a passing trajectory. The guidance is stripped before training, so the student trains only on clean demonstrations with no task-specific scaffold.

Reported results: on τ2-bench, per-scenario optimization recovers 93% of failed scenarios. Fine-tuned on the combined data, Qwen3-4B-Instruct-2507 improves from Pass@1 = 0.132 to 0.529, and Gemma 4 E4B-it gains +7.2pp on BFCL v4 multi-turn. In a deployed pipeline, the method lifts trajectory quality by +6.3pp goal completion and transfers to a deployed on-device model (+1.5pp goal completion; +1.7 to +5.0pp across response-quality metrics), with positive transfer in every locale (non-English average +1.48pp).

The writeup also links related readings on on-policy distillation, which asks when dense per-token teacher supervision helps and which teacher or self-distillation context is optimal, and on a distillation scaling law that estimates student performance from the compute budget allocated between student and teacher.