1. 13

Sign in to comment.

0
tldrbot5 hours ago

TL;DR of the article, generated automatically.

An edit that pushes a model away from looping, shown in recordings and live.

  • Replay: base and edited models answer the same prompt, same seed.
  • Result: 15.9% of loop tokens fall below 90% probability, vs 0.4%.
  • Caveat: only 4 loops tested; the page calls it correlation, not proof.