Trace2Tower: Transition-Aware EigenTrace Induction... | AI Research

Trace2Tower: Transition-Aware EigenTrace Induction of Multi-Level Skills for LLM Agents
Large language model (LLM) agents often struggle to learn from past experiences because raw execution logs are messy, redundant, and filled with both successful strategies and failed attempts. Trace2Tower is a new framework designed to solve this by transforming these raw, chaotic trajectories into a clean, organized hierarchy of reusable skills. Instead of just storing past actions, the system identifies the underlying behavioral patterns that actually lead to success, allowing agents to perform complex tasks more efficiently. The ai search story also surfaces in Stanford AI discovery identifies natural weight..., adding another angle.

From Raw Logs to Structured Events

The process begins by breaking down long, complex interaction logs into "canonical events." The system identifies specific stages of an action—such as changes in subgoals or manipulated objects—and standardizes them. By replacing task-specific details with general labels, the framework can recognize when the same underlying operation is being performed across different tasks. This creates a foundation where the agent can learn from the structure of its past behavior rather than just memorizing specific, isolated instances.

Identifying Success-Aligned Patterns

A core innovation of Trace2Tower is its use of "contrastive spectral decomposition." The framework builds a graph that maps out how different events transition into one another. Crucially, it distinguishes between successful outcomes and failure-prone shortcuts. By mathematically contrasting these two types of data, the system isolates stable behavioral modes that reliably lead to success while suppressing patterns that typically result in failure. This ensures that the skills the agent learns are based on effective, proven strategies. The ai search story also surfaces in Google AI Releases TimesFM 3 for..., adding another angle.

Building a Hierarchical Skill Tower

Once these behavioral modes are identified, they are organized into a "skill tower" consisting of three levels: action templates, procedural routines, and overarching task strategies. This hierarchy allows the agent to retrieve the right level of guidance for any given situation. During deployment, the agent uses a verifier-guided feedback loop to continuously refine these skills. This means the agent doesn't just learn once; it constantly updates its library based on whether its chosen skills actually save time, reduce errors, or improve task outcomes.

Performance and Efficiency

In testing on the ALFWorld and WebShop benchmarks, Trace2Tower demonstrated significant improvements over existing methods. On ALFWorld, the framework achieved an 87.31% success rate while requiring fewer execution steps and significantly fewer invalid actions compared to baseline approaches. These results suggest that by focusing on the structural relationships between events and outcomes, agents can achieve higher mastery of complex tasks while reusing their past experiences more effectively. The ai search story also surfaces in Qwen Developers Open-Source Local-First Search Layer..., adding another angle. as detailed in the full paper on Arxiv