EVENT DOSSIER
On September 7, 2026, arXiv published the paper “Trace2Tower: Transition-Aware EigenTrace Induction of Multi-Level Skills for LLM Agents”, which introduced a new reinforcement learning framework. This framework uses contrastive谱 decomposition to distill the original interaction trajectories into robust skill levels and continuously optimizes them through feedback from validators. Its core mechanism involves constructing a unified graph structure based on semantic compatibility, transition dynamics, and outcome evidence, thereby isolating successful behavior patterns and suppressing failed shortcuts. In the ALFWorld benchmark, Trace2Tower required only 10.35 steps with an ineffective action ratio of 0.26%, achieving a success rate of 87.31%; on the WebShop benchmark, it reached a precise success rate of 50.67%. These results are significantly better than those of existing baseline models.
arXiv:2609.05261v1 Trace2Tower 提出基于谱分解的技能塔框架
arXiv:2609.05261v1 提出 Trace2Tower，这是一种通过对比谱分解将原始轨迹蒸馏为稳健技能层级、利用验证器反馈持续优化的过渡感知 EigenTrace 框架。该框架将步骤级交互抽象为规范事件，构建由语义兼容性、转换动力学和结果证据统一管理的图结构，隔离成功对齐的行为模式并抑制失败捷径。在 ALFWorld 基准上，Trace2Tower 以仅需 10.35 步和 0.26 个无效动作实现了 87.31% 的成功率；在 WebShop 基准上达到 50.67% 的精确成功率，显著优于现有基线。