Do LLM Agents Still Need to Be Told How to Self-Improve?
Capable agents improved at gameplay by devising their own learning routines while their model weights stayed frozen.
The researchers gave agents a compute budget, tools to collect experience and edit text documents, and a general self-improvement goal. They imposed no learning loop and provided no access to real-world data. The agents developed strategies resembling self-regulated learning and self-generated curricula. Their learned text artifacts also improved performance on external benchmarks spanning different game horizons, mechanics, and 3D navigation.
ArXiv · AI/CL/LG's note
The researchers gave agents a compute budget, tools to collect experience and edit text documents, and a general self-improvement goal. They imposed no learning loop and provided no access to real-world data. The agents developed strategies resembling self-regulated learning and self-generated curricula. Their learned text artifacts also improved performance on external benchmarks spanning different game horizons, mechanics, and 3D navigation.
ArXiv · AI/CL/LG's note
score 5