Apple proposes PROOF-Gen to improve distillation by optimizing data

Apple ML Research · rss · 2026-08-26

Apple introduces PROOF-Gen, a method to improve the distillation of tool-calling capabilities from teacher models to deployable models. Traditional generate-and-filter mechanisms discard failed trajectories, leaving hard scenarios unresolved in each cycle. PROOF-Gen optimizes data by leveraging near-misses to provide learning signals, enhancing distillation performance on benchmarks like τ2-bench.

Original post →

More from Research

Research channel →