Reproduce “ARCHER: From Rationale Imitation to Executable Reasoning in Small 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
Dhruv Dixit, Khushi Patel, Congcong Xu, Min Song
ARCHER trains compact models to solve math problems through explicit reasoning, calculator calls, observations, verification, and answers, then refines their policies using executed trajectories. Its grouped rewards prioritize answer correctness and condition tool-use and trace-alignment feedback. Across five backbones and eight benchmarks, ARCHER averages 73.36% accuracy versus 67.85% for CoA; process evidence is limited to GSM-8K, not general reasoning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.