Reproduce “Where Reasoning Breaks: Logic-Aware Path Selection by Controlling Logical Connectives in LLMs Reasoning Chains” on OpenResearch

About this paper

Seunghyun Park, Yuanyuan Lei

This research identifies logical connectives as critical, high-entropy pivot points in LLM reasoning chains where models frequently struggle to make correct logical transitions. A multi-layered framework, comprising gradient-based steering, localized branching, and targeted transition preference optimization (TTPO), is introduced to intervene at these junctures, improving logical reasoning performance on benchmarks like ZebraLogic and ProntoQA while maintaining computational efficiency comparable to greedy decoding.

Key insights

• Logical connectives are disproportionately causal elements in LLM reasoning, with single-token perturbations at these points significantly altering final outcomes.
• LLM uncertainty in multi-step reasoning is concentrated at connective positions, where models often consider multiple, divergent logical relations.
• Successful reasoning repairs at connective pivots involve shifts between different types of logical relations (e.g., causal to instantiation), rather than mere stylistic changes.

Summary by alphaXiv.