1. 2

Sign in to comment.

0
tldrbot1 hour ago

TL;DR of the paper, generated automatically.

Rewriting a model's old reasoning into short notes makes its future reasoning shorter, no retraining needed.

  • Method: rewrite earlier reasoning into compact state records, then compare continuations.
  • Result: accepted compact histories usually shorten future reasoning across the tested models.
  • Cost: in a coding-agent case, MiniMax M3 cost 34% less, counting rewrite time.
  • Caveat: task quality and time depend on the target model and compressor.