ActionOPD Distills Agent Decisions, Not Tokens

ActionOPD targets a real weakness in conventional distillation: agents need to make the right tool-use decision, not reproduce identical wording. Its results are promising, but semantic agreement remains a proxy for correctness rather than a guarantee.

• –Uses on-policy student states, reducing train–inference distribution mismatch
• –Supports cross-family teacher–student pairs without tokenizer alignment
• –Improves four-benchmark averages by 3.3–6.9 percentage points
• –Open-source implementation includes training workflows for tool-use environments
• –Stronger single-trial results come with weaker repeated-success reliability than some baselines