Procedural Knowledge In Pretraining Drives Reasoning In Large Language Models 2024 11 20

A study summarized by Cohere finds evidence that language models draw on learned solution methods when answering some mathematical questions. The publisher abstract describes tracing influential training documents for two models across three simple mathematical reasoning tasks. Different questions within a task often depended on similar documents containing formulas or code that demonstrated a procedure. Factual questions tended to depend on more distinct documents, often containing the answer itself. The authors interpret this pattern as evidence of reusable procedural knowledge. Its scope is those models and tasks, rather than a general demonstration that model reasoning works like human reasoning.