Reproduce “Predicting Privileged Information Effectiveness for On-Policy Self-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
Kaijian Zou, Emiliano Penaloza, MohammadHossein Rezaei, Anisha Gunjal, Mohit Raghavendra, Lu Wang, Yunzhong He, Anas Mahmoud
The authors compare 16 privileged-information configurations for on-policy self-distillation across ALFWorld, AppWorld, and τ²-retail, using three Qwen3 students. They find that error-step KL contrast correlates positively with post-training performance in all nine student-benchmark combinations and can help select information when candidates vary. This evidence is limited to tested models, tasks, and candidate pools.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.