Trace2Tower: Transition-Aware EigenTrace Induction of Multi-Level Skills for LLM Agents
Trace2Tower turns agent execution traces into a hierarchy of reusable skills, filtered toward successful behavior.
The paper says the method abstracts raw interaction steps into canonical events, then builds a graph using semantic fit, transition dynamics, and outcome evidence. Its spectral decomposition is meant to surface stable behavior patterns while suppressing failure-prone shortcuts. Those patterns feed a “skill tower” spanning action templates, routines, and higher-level task strategies, refined with verifier feedback. Reported results are 87.31% success on ALFWorld and 50.67% exact success on WebShop. ArXiv · AI/CL/LG's note
The paper says the method abstracts raw interaction steps into canonical events, then builds a graph using semantic fit, transition dynamics, and outcome evidence. Its spectral decomposition is meant to surface stable behavior patterns while suppressing failure-prone shortcuts. Those patterns feed a “skill tower” spanning action templates, routines, and higher-level task strategies, refined with verifier feedback. Reported results are 87.31% success on ALFWorld and 50.67% exact success on WebShop. ArXiv · AI/CL/LG's note
score 4