Reproduce “Inductive Feedback for Mixed-Policy Distillation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Amir Moeini, Huaijiang Zhu, Daniel Havir, Shangtong Zhang
The method trains a student from verbal feedback by measuring how the same teacher’s next-token probabilities change with and without feedback. It uses those changes to reweight the student’s predictions and supervises student-generated and feedback-guided continuations. On EmbodiedEval, Qwen3.5-4B reaches an 82.1% ± 3.0 mean pass rate, versus 73.7% ± 1.0 for SDPO; these are benchmark results, not evidence of physical deployment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.