Shared Logical Subspace Discovered: Aligning Natural Language and Symbolic Views Steers LLM Multi-Step Reasoning
arXiv·medium signal
Fang, Thai, and Lei find that LLMs encode a shared internal subspace where natural language reasoning and symbolic logic converge, and steering activations toward this subspace measurably improves multi-step logical reasoning. Unlike approaches that bolt on external symbolic solvers or purely refine chain-of-thought in natural language, this method operates on the model's internal representations. Practically relevant for anyone fine-tuning or prompting LLMs for structured reasoning tasks.