Supervised fine-tuning on teacher-generated trajectories is the usual first stage for distilling tool-calling capabilities into deployable fashions. Publish-training pipelines that drive shipped tool-calling brokers re-run this stage on a each day or weekly cadence, paying the frontier-teacher value every cycle, but the mechanism is generate-and-filter (hold the trainer’s passing trajectories, discard the remaining) and every cycle leaves behind the identical laborious situations as a result of failures provide no sign. On τ 2-bench, 57% of trainer trials fail, two-thirds of them near-misses (most device calls appropriate, undone by one decisive error).
We introduce PROOF-Gen (Per-scenario Reflective Optimization to Overcome Failed Technology), which recovers golden trajectories from these failures through per-scenario immediate optimization. For every failed process, a reflector analyzes the execution hint and analysis suggestions, then writes corrective steering that steers the trainer to a passing trajectory. The steering is stripped earlier than coaching, so the coed learns from clear demonstrations with no task-specific scaffold.
On τ 2-bench, per-scenario optimization recovers 93% of failed situations. High-quality-tuned on the mixed knowledge, Qwen3-4B-Instruct-2507 improves from Passˆ1=0.132 to 0.529 and Gemma 4 E4B-it positive aspects +7.2pp on BFCL v4 multi-turn. In a deployed pipeline, the strategy lifts trajectory high quality by +6.3pp aim completion and transfers to a deployed on-device mannequin (+1.5pp aim completion; +1.7 to +5.0pp throughout response-quality metrics), with constructive switch in each locale (non-English common +1.48pp).
