PROOF-Gen: From Optimized Data to Better Distillation
2026-08-26 08:00Models🔥 28.9 heat score
1sources
1days unfolding
28.9heat score
1mentions
SummaryAI generated
Apple ML Research released PROOF-Gen on August 26, 2026, aiming to optimize the distillation of models and the ability to call tools. The existing standard approach involves supervised fine-tuning on trajectories generated by teachers, but the generation and filtering mechanisms result in the same difficult scenarios being left over from each training cycle, as failures do not provide signals. In the τ 2-bench benchmark, 57% of teacher attempts failed, two-thirds of which were near misses (most tool calls were correct but not executed).
PROOF-Gen proposes a new mechanism to optimize the ability of distillation tools to call functions. The current standard approach involves supervised fine-tuning on trajectories generated by teachers, but the existing process uses a generation and filtering mechanism, resulting in the same difficult scenarios being left behind in each training cycle because failures do not provide signals. In the τ 2-bench benchmark, 57% of teacher attempts failed, two-thirds of which were near misses (most tools called correctly but did not execute).