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.

While multilingual pretrained language models (LMs) fine-tuned on a single language have shown substantial cross-lingual task transfer capabilities, there is still a wide performance gap in semantic parsing tasks when target language supervision is available. In this paper, we propose a novel Translate-and-Fill (TaF) method to produce silver training data for a multilingual semantic parser. This method simplifies the popular Translate-Align-Project (TAP) pipeline and consists of a sequence-to-sequence filler model that constructs a full parse conditioned on an utterance and a view of the same parse. Our filler is trained on English data only but can accurately complete instances in other languages (i.e., translations of the English training utterances), in a zero-shot fashion. Experimental results on three multilingual semantic parsing datasets show that data augmentation with TaF reaches accuracies competitive with similar systems which rely on traditional alignment techniques.