To keep long interaction contexts tractable, stepwise group-relative training for multi-turn LLM agents materializes each source trajectory into multiple res…
机构:联想
来源:arXiv 2611.16816 | AI4Papers 论文推荐平台