LM2LM uses a research world model to train improvement proposals
Original title: Language Models Learn to Improve Language Models
Held-out runs show Qwen3.8-27B beating GPT-6-sol given 300 records, at 33% lower data-generation cost.
LM2LM trains LLMs with reinforcement learning to propose language-model improvements, using a frozen research world model as a cheaper reward surrogate. The world model predicts proposal gains from experimental records and is refreshed with measured outcomes from highly ranked current-policy proposals. Across eight autoregressive and diffusion environments, the paper reports gains on held-out tasks and generalization to an unseen model family.
Unconfirmed: Outperformance and cost figures are the authors' held-out evaluations, not independently confirmed.
- Real RL feedback is described as too costly, potentially hundreds of thousands of GPU-hours.
- Periodic real executions are added to the world model's context to limit overestimation.
- The method generalized to a model family not seen in training.
- Products
- LM2LM
- AI models
- Qwen3.8-27BGPT-6-sol
From the original article
1 min read at huggingface.co
Large language models (LLMs) are increasingly used to automate AI research, including language-model (LM) development, offering a potential path toward recursive self-improvement. However, turning experimental feedback into a learned ability to propose effective improvements across models faces two challenges.
By Minhua Lin, Zijun Wang, Zewen Liu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.