Reproduce “A Implies B: Circuit Analysis in LLMs for Propositional Logical Reasoning” 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

Guan Zhe Hong, Nishanth Dikkala, Enming Luo, Cyrus Rashtchian, Xin Wang, Rina Panigrahy

Causal mediation analysis reveals that large language models (LLMs) like Gemma-2-27B employ sparse, modular reasoning circuits for multi-step propositional logic, composed of distinct attention head families. These models exhibit lazy reasoning, activating specific sub-circuits only when a query is presented, and similar circuit patterns across different model scales.

Key insights

  • LLMs exhibit modular and localized reasoning for propositional logic problems, challenging the view of entirely opaque or distributed computation.
  • Reasoning often follows a 'lazy' strategy, where core circuits activate on-demand upon query rather than through eager pre-processing.
  • Mechanistic insights from smaller, task-specific models may not generalize to large, pre-trained LLMs, highlighting the need for direct study of frontier models.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.