New preprint: shared circuits predict LLM arithmetic generalization across formats

burny_tech · x · 2026-09-08

A new preprint (with Jacob Andreas and Evelina Fedorenko) examines whether LLMs generalize arithmetic reasoning across formats: humans who can solve 2+5 can solve "two plus five" for free, but LLMs are more brittle. The authors propose "shared circuits" as a predictor of cross-format generalization, offering a new lens on LLM internal representations and generalization.

Original post →

More from Research

Research channel →