Reproduce “Distillation, Self-Training, and the Fragility of Learned Chemical Reasoning” 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
Anagha Aneesh, Sebastian Ibarraran, Frank Hu, Jason Liu, Grant Rotskoff
The authors train open models to predict reaction products by distilling product-visible traces from a stronger teacher, then training on student outputs verified against recorded products. On USPTO-50k, Resonance-8B exact-match accuracy rises from 1.2% to 42.8% after three stages. Yet the 8B model often abandons correct answers when told they are wrong; these results do not establish robust chemical reasoning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.