Reproduce “TraceForge: Distilling with Your Own Thinking” 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
Yan Wang
TraceForge generates student-family reasoning traces from a teacher’s context and answer, rather than recovering hidden reasoning. In a controlled comparison, Qwen3.5-9B-Base students trained with answers fixed raise pass@1 on HumanEval+ from 64.63% to 68.29% and MBPP+ from 52.38% to 59.52% using family-native traces instead of Fable 5 traces. Evidence is limited to coding benchmarks and one run per condition.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.