PROOF-Gen: From Optimized Data to Better Distillation

Apple researchers introduced PROOF-Gen, a method that recovers golden trajectories from failed teacher generations via per-scenario prompt optimization, improving tool-calling distillation.

On τ2-bench it recovered 93% of failed scenarios; fine-tuning Qwen3-4B-Instruct-2507 improved Pass@1 from 0.132 to 0.529, and Gemma 4 E4B-it gained +7.2pp on BFCL v4 multi-turn.

In a deployed pipeline, the method lifted goal completion by +6.3pp and transferred to an on-device model with positive results across locales.