LATIDIA · Investigación

Una descomposición en cuatro etapas de la resolución de problemas de palabras y la fragilidad mecánica en el razonamiento matemático de LLM

arXiv:2609.17804v1 Tipo de anuncio: nuevo Resumen: Los modelos de lenguaje grandes resuelven problemas matemáticos de la escuela primaria con alta precisión, pero una sola cláusula irrelevante insertada en el problema puede colapsarlo. Conciliamos estos

La noticia

arXiv:2609.17804v1 Announce Type: new 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