TL;DR of the paper, generated automatically.

A probe finds LLMs lose track of cumulative updates, and output format shifts scores.

• Results: 84.9% on single-step math, 31.7% on core items after 3 to 7 updates.
• Long chains: no model exceeds 10% from 50 updates on.
• Format: 88.0% with written steps, mainly from the instruction changes.