Reproduce “Reason in Style: Discovering and Controlling Style in Language Models” 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
Ioana Marinescu, Eric Karl Oermann, Kyunghyun Cho
The authors train Qwen3 students on math solutions by discovering solution styles, adding style tokens, and reweighting traces to balance styles per question. Across six math benchmarks, importance-weighted style-conditioned fine-tuning improves answer success over standard fine-tuning; for Qwen3 0.6B on AIME 2024, Pass@256 is 29.7% versus 9.7%. These results are limited to the math setting and corpus.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.