Paper Overview
Field: NLP Authors: Feihao Fang, My T. Thai, Yuanyuan Lei Posted: 2026-04-21 arXiv: 2604.19716
Summary
Large Language Models (LLMs) still struggle with multi-step logical reasoning. Existing approaches either purely refine the reasoning chain in natural language form or attach a symbolic solver as an external module. In this work, the authors instead ask whether LLMs contain a shared internal logical subspace that simultaneously aligns natural-language and symbolic-language views of the reasoning process.
The hypothesis is that this logical subspace captures logical reasoning capabilities in LLMs that are shared across views while remaining independent of surface forms. To verify this, the authors employ Canonical Correlation Analysis (CCA) on the paired residual activations from natural-language and symbolic-language reasoning chains, learning a low-dimensional subspace with maximum cross-view correlation.
They further design a training-free method that steers the LLM's reasoning chain along this logical subspace, leveraging complementary reasoning signals from both the natural-language and symbolic views.
Results
- Experiments on four logical reasoning benchmarks demonstrate the method's effectiveness.
- Accuracy improvements of up to 11 percentage points.
- Strong generalization to out-of-domain questions.
Original Abstract (excerpt)
> Large Language Models (LLMs) still struggle with multi-step logical reasoning. Existing approaches either purely refine the reasoning chain in natural language form or attach a symbolic solver as an external module. In this work, we instead ask whether LLMs contain a shared internal logical subspace that simultaneously aligns natural-language and symbolic-language views of the reasoning process...
--- *Auto-collected on 2026-04-23.*