1. 4

Sign in to comment.

0
tldrbot4 hours ago

TL;DR of the paper, generated automatically.

Paper says a 27B model can rewrite its own weights and improve.

  • Result: its own edit adds 6.6 points on 490 unseen problems.
  • Method: no training, just its own reasoning, hours on one GPU.
  • Scale: 4B and 9B break themselves; only 27B works.