Reproduce “Solving Math Word Problems via Cooperative Reasoning induced 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
Xinyu Zhu, Junjie Wang, Lin Zhang, Yuxiang Zhang, Ruyi Gan, Jiaxing Zhang, Yujiu Yang
The Cooperative Reasoning (CoRe) framework enhances Language Models' ability to solve Math Word Problems by mimicking human dual-process cognition with a Generator and two Verifiers. This approach allows CoRe, using approximately 12.4 billion parameters, to achieve superior zero-shot performance on benchmarks like SingleEq (79.5%) and MultiArith (97.5%), outperforming much larger models such as Instruct GPT-3 (175B) and PaLM (540B).
Key insights
- Explicitly modeling human dual-process cognition (fast generation and deliberate verification) allows LMs to achieve more robust, adaptive, and human-like reasoning.
- Integrating timely, fine-grained feedback from both token-level and path-level verifiers directly into the reasoning process through MCTS significantly enhances the quality and correctness of generated solutions.
- A self-training strategy, coupled with quality filtering, can effectively boost an LM's intrinsic reasoning and generalization abilities by leveraging self-generated data, especially in data-scarce scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.