RESEARCH · RESEARCH · #532
Four-stage decomposition of LLM math reasoning; distractor fragility localized to Operation Planning
The paper proposes that LLMs solve grade‑school math word problems via a four-stage pipeline—Schema Abstraction, Operation Planning, Operand Binding, and Computation—each producing distinct intermediate representations in specific layer bands. Using the same scaffold to analyze failure modes, the authors trace collapse caused by a single irrelevant clause to corruption of the Operation Planning stage, implemented by a set of attention heads whose causal role is validated bidirectionally.
KEY POINTS
- The paper proposes that LLMs solve grade‑school math word problems via a four-stage pipeline—Schema Abstraction, Operation Planning, Operand Binding, and Computation—each producing distinct intermediate representations in specific layer bands.
- Using the same scaffold to analyze failure modes, the authors trace collapse caused by a single irrelevant clause to corruption of the Operation Planning stage, implemented by a set of attention heads whose causal role is validated bidirectionally.
- Provides a concrete mechanistic account that pinpoints which internal stage and attention heads cause distractor-induced failures, enabling targeted interpretability and robustness interventions for LLM math reasoning.
WHY IT MATTERS
Provides a concrete mechanistic account that pinpoints which internal stage and attention heads cause distractor-induced failures, enabling targeted interpretability and robustness interventions for LLM math reasoning.