Reproduce “Reasoning Circuits in Language Models: A Mechanistic Interpretation of Syllogistic Inference” 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
Geonhee Kim, Marco Valentino, André Freitas
This research identifies a multi-stage "error-correction" reasoning circuit within language models for syllogistic inference. The study demonstrates this circuit's transferability across various logical forms and models, while also revealing its susceptibility to contamination by belief biases derived from world knowledge acquired during pre-training.
Key insights
- Language models perform syllogistic inference through a four-phase "error-correction" mechanism, including long induction, duplication, middle-term suppression, and term movement.
- The content-independent reasoning circuit, while necessary for both belief-consistent and inconsistent arguments, is only sufficient for belief-consistent inputs, indicating contamination by world knowledge.
- Core reasoning mechanisms generalize across a range of unconditionally valid syllogistic schemes, different model sizes, and diverse Transformer architectures, emerging during pre-training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.