ASCENT:通过验证经验自蒸馏实现长程智能体在线测试时训练
原标题:Paper page - ASCENT: Online Test-Time Training of Long-Horizon Agents via Self-Distillation of Verified Experience
AI 摘要
该论文提出 ASCENT 方法,用于长程 LLM 智能体的在线测试时训练(OaTTT):智能体在部署中每个任务只执行一次,仅凭该次轨迹及其验证结果作为学习信号。ASCENT 通过让冻结的初始模型副本以验证轨迹作为特权信息进行自蒸馏,将预测分布蒸馏进持久的 LoRA 快速权重,从而避免直接模仿或强化单次尝试导致的策略不稳定。在 ALFWorld、WebShop 和 AppWorld 上,ASCENT 随经验积累提升任务成功率与交互效率,超过在线适应方法,并可迁移到未见场景;在 ALFWorld 和 WebShop 上两种模型规模下精确成功率提升超过 22 个百分点。
正文节选
ASCENT: Online Test-Time Training of Long-Horizon Agents via Self-Distillation of Verified Experience Abstract A large language model (LLM) agent solves long-horizon tasks through many reasoning-action turns, with one verification signal at termination. Deployed agents face streams of related tasks, making their trajectories a natural resource for improvement. In-context adaptation agents store reflections, memories, or skills as text, so reuse depends on retrieving the right experience and on a