Researchers have introduced COBALT, a novel method for training large language models (LLMs) in multi-turn code generation. This approach utilizes offline trajectories, which are more cost-effective and stable than traditional online reinforcement learning. COBALT treats multi-turn code generation as a one-step recoverable Markov decision process, training LLMs to complete partial trajectories through single-step code generation. Experiments on LiveCodeBench showed significant improvements for models like R1-Distill 8B and Qwen3 8B, matching the performance of existing online RL baselines. AI
IMPACT This method offers a more efficient and stable approach to training LLMs for complex code generation tasks.
RANK_REASON The cluster contains an academic paper detailing a new method for LLM training. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
