Tsinghua's ActionOPD lifts Qwen3-4B agent success on τ²-bench from 29.2% to 40%

jiqizhixin · x · 2026-11-17

Tsinghua University introduces ActionOPD in an arXiv preprint: a teacher-distillation scheme for agents that scores teacher actions against student candidates at states the student actually visits — requiring no teacher token probabilities and no tokenizer matching. Tool calls are compared by operation and normalized arguments; user-facing replies are checked for intent and key facts; the best candidate advances the environment and the candidate group supplies the policy update. With a Qwen3-32B teacher, Qwen3-4B's average single-trial success on held-out τ²-bench tasks rises from 29.2% to 40.0% in matched controlled comparisons (trained variants at step 50), though offline distillation still retains higher performance.

Original post →

More from coding & agent

coding & agent channel →