Trace2Tower: Transition-Aware EigenTrace Induction of Multi-Level Skills for LLM Agents
Trace2Tower is a framework that distills raw execution traces into a skill hierarchy for LLM agents. It abstracts step-level interactions into canonical events and constructs a unified graph based on semantic compatibility, transition dynamics, and outcome evidence. Using contrastive spectral decomposition, it isolates stable, success-aligned behavioral modes and suppresses failure-prone shortcuts. On ALFWorld, it achieves 87.31% success with 10.35 steps and 0.26 invalid actions; on WebShop, it reaches 50.67% exact success.
Trace2Tower introduces a transition-aware EigenTrace framework to improve LLM agent performance on interactive tasks. It builds a skill tower of action templates, procedural routines, and task strategies, refined via verifier-guided feedback. The method outperforms existing baselines on ALFWorld and WebShop benchmarks.
The approach uses contrastive spectral decomposition on a graph of canonical events to extract success-aligned behavioral modes, which are then organized into a multi-level skill hierarchy. This enables more efficient task execution with fewer steps and invalid actions.
Improving agent reliability and efficiency on interactive benchmarks like ALFWorld and WebShop suggests potential for more robust autonomous agents in real-world applications, reducing error rates and operational costs.
Higher success rates with fewer steps and invalid actions can lead to cost savings and better user experience in agent-based products, making LLM agents more viable for commercial deployment.
Next signals include peer validation, open-source code release, and application to more complex or real-world environments. Further improvements may focus on scaling the skill tower and integrating with larger language models.