ActionOPD: Distilling Decisions, Not Tokens
ActionOPD reports average gains of 3.3–6.9 percentage points across four out-of-distribution benchmarks for smaller tool-using agents.
The method rewards student actions for semantic agreement with actions sampled from a teacher at states the student encounters. It needs neither teacher token probabilities nor matching tokenizers, and does not require students to reproduce the teacher’s written reasoning. The authors trained Qwen3-4B and Qwen3-8B with two teachers on curated tool-use tasks. On τ²-bench, they report competitive performance against offline distillation and token-level on-policy distillation.
HF Daily Papers' note
The method rewards student actions for semantic agreement with actions sampled from a teacher at states the student encounters. It needs neither teacher token probabilities nor matching tokenizers, and does not require students to reproduce the teacher’s written reasoning. The authors trained Qwen3-4B and Qwen3-8B with two teachers on curated tool-use tasks. On τ²-bench, they report competitive performance against offline distillation and token-level on-policy distillation.
HF Daily Papers' note
score 5