8-27B) using reinforcement learning to propose modifications to language model architectures (autoregressive and diffusion).
Language models learn to propose improvements for other language models via RL with a learned world model
LM2LM uses a frozen LLM to predict experimental outcomes, periodically grounding its feedback in real runs to maintain reliability.
Big Tech
Minhua Lin · Zijun Wang · Zewen Liu · Zhaotian Weng · Bing He · Zhan Shi · +4 more
The Pennsylvania State University · UC Santa Cruz · Emory University · UC Santa Barbara · Amazon
Research Digest··2 min read
The authors propose LM2LM, a framework that trains a language model via reinforcement learning to generate effective proposals for improving other language models.
§