Reproduce “Privileged Context as Drift in 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
Ravenor Davion, Nick Rui
The authors compare privileged-context choices in on-policy self-distillation, where a frozen teacher conditions on extra information while a student trains on its own generated responses. Across nine conditions for one Qwen2.5-7B-Instruct model, content choices produced 5.1 times wider median per-token reverse-KL ranges than source choices. These one-seed comparisons describe this grid, not universal effects, and drift did not predict retention.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.