Report
CriticOPD uses teacher critiques to help AI students rewrite incorrect solutions
The researchers report a 4.57-point gain in average accuracy over on-policy distillation for Qwen3-1.7B across six math benchmarks.
TLDR
In on-policy distillation, a student model learns from an AI teacher as it generates solutions. The CriticOPD researchers say the teacher can follow a student’s mistake even when it recognizes the error. Their approach has the teacher critique the solution, then lets the student rewrite from its last error. The authors report average accuracy gains over standard on-policy distillation of 4.57 percentage points for Qwen3-1.7B and 4.06 for Qwen3-0.6B across six math benchmarks.
Combined views
—
1 Source, first seen ago
