Reproduce “ProtoReasoning: Prototypes as the Foundation for Generalizable Reasoning in LLMs” on OpenResearch

About this paper

Feng He, Zijun Chen, Xinnian Liang, Tingting Ma, Yunqi Qiu, Shuangzhi Wu, Junchi Yan

The ProtoReasoning framework enhances Large Language Models' (LLMs) generalizable reasoning by training them on abstract, verifiable prototypes derived from formal logic (Prolog) and planning (PDDL). This approach led to performance improvements of up to 6.4% on planning tasks and 4.7% on logical reasoning, demonstrating that learning structural reasoning patterns facilitates cross-domain transfer.

Key insights

• Cross-domain generalization in Large Language Models (LLMs) originates from shared abstract reasoning prototypes, which represent fundamental, domain-agnostic reasoning patterns.
• Training LLMs on formal prototype representations (e.g., Prolog, PDDL) is as effective as direct natural language training for enhancing generalizable reasoning on structurally similar problems.
• Effective generalization through prototype training requires explicit Chain-of-Thought (CoT) reasoning, as it homogenizes and reinforces reasoning paths.

Summary by alphaXiv.