Large language model agents increasingly rely on execution traces to handle complex interactive tasks, yet current approaches only perform shallow trajectory retrieval and flat skill summarization, overlooking temporal dependencies and outcome‑conditioned topology. Trace2Tower introduces a transition‑aware EigenTrace framework that distills raw trajectories into a robust skill hierarchy. It first abstracts step‑level interactions into canonical events and builds a unified graph constrained by semantic compatibility, transition dynamics, and outcome evidence. A novel contrastive spectral decomposition then isolates stable, success‑aligned behavioral modes while rigorously suppressing failure‑prone shortcuts. These modes organically populate a dynamic skill tower comprising action templates, procedural routines, and overarching task strategies, continuously refined through verifier‑guided feedback. On ALFWorld, Trace2Tower attains 87.31% success with only 10.35 steps and 0.26 invalid actions; on WebShop it reaches 50.67% exact success. Across both benchmarks, it markedly outperforms existing baselines in task mastery and context‑efficient experience reuse.
Review