LM2LM preprint reports lower cost for training AI research proposal models

Researchers report that LM2LM trained Qwen3.8-27B to propose language-model improvements that outperformed GPT-6-sol in real-execution evaluations on held-out tasks, with 33% lower data-generation cost. The preprint tests eight development environments covering autoregressive and diffusion language models. The method uses reinforcement learning, rewarding proposals through a frozen language model that predicts experimental gains from records of actual experiments. Researchers periodically execute highly ranked proposals and add measured results to its context to address unreliable predictions as training changes the proposals. The trained system also generalized to a model family unseen during training. The comparison gave GPT-6-sol 300 real experimental records in context; the findings concern evaluated development tasks.