Large language models break math problem solving with distractors
A Four-Stage Decomposition of Word-Problem Solving and Mechanistic Fragility in LLM Math Reasoning
Artificial IntelligenceMachine Learning
Summary
Large language models can solve simple math stories accurately, but adding extra, unrelated details can cause them to fail badly. The authors explain this by breaking down the model’s thinking process into four steps: understanding the problem type, planning what math to do, figuring out which numbers to use, and doing the calculation. They found that distracting details specifically confuse the step where the model decides what math operations to perform. This gives a clear picture of how these models do math and where they get confused.
What this means in practice
- •For ai developers: Improve robustness of math-solving language models by targeting their operation planning step against irrelevant information.
- •For automated tutoring platforms: Design systems that detect and handle distractors in math word problems to provide clearer guidance to learners using AI.
Authors
Zhongdi Qu, Carla P. Gomes
Abstract
Large language models solve grade-school math word problems with high accuracy, yet a single irrelevant clause inserted into the problem can collapse it. We reconcile these observations with a mechanistic account. We show that the model's internal computation decomposes into a four-stage sequential pipeline, Schema Abstraction, Operation Planning, Operand Binding, and Computation, each stage producing a distinct intermediate representation in an identifiable band of layers. Using the same scaffold to diagnose distractor-induced failure, we localize the corruption to a single stage, Operation Planning, implemented by a set of attention heads whose causal role we validate bidirectionally. In short, we provide a mechanistic interpretation of math word problem reasoning in LLMs, and their failure when distracted.