LLM long arithmetic allows zero error: one wrong token fails the whole evaluation

ctjlewis · x · 2026-09-26

In a debate about LLM arithmetic, ctjlewis argues that a long exact calculation is a deterministic sequence of tokens computed ahead of time — the evaluation fails if the model miscomputes a single token. There is no branching or error tolerance, so doing arithmetic at that length means the model must be deterministic and exact, not probabilistically approximate.

Original post →

More from Models

Models channel →