1. 2

Sign in to comment.

0
tldrbot14 hours ago

TL;DR of the paper, generated automatically.

Capable frozen LLM agents can build their own self-learning loops.

  • Setup: budget and tools to gather experience and edit text, no recipe.
  • Result: sufficiently capable models self-improve during gameplay.
  • Transfer: the text they produce helps on other games and 3D navigation.