Trace2Tower: Transition-Aware EigenTrace Induction of Multi-Level Skills for LLM Agents

Jiazheng Sun; Boyu Yang; Binhao Yuan; Mingxuan Li; Xin Peng

arXiv 2026cs.AIcs.SE

Trace2Tower Transition-Aware EigenTrace Induction of Multi-Level Skills for LLM Agents

From execution traces to a success-aligned tower of reusable agent skills.

Jiazheng Sun^*,Boyu Yang^*,Binhao Yuan,Mingxuan Li,Xin Peng

College of Computer Science and Artificial Intelligence, Fudan University

^* Equal contribution

Paper Code arXiv

Overview

Structure is the missing layer in experience reuse.

Existing agent systems retrieve trajectories or summarize flat skills, but they rarely preserve which behaviors belong together, which transitions are executable, and which patterns consistently lead to success.

Trace2Tower turns successful and failed interaction histories into canonical events, builds an outcome-aware transition graph, and uses contrastive spectral decomposition to identify stable behavioral modes. The result is a dynamic hierarchy of action templates, reusable procedures, and task-level strategies that supports compact retrieval and verifier-guided refinement.

What is still missing?

Behavior is more than isolated text.

Relations

Which behaviors perform the same functional role?

Transitions

Which behaviors reliably follow one another?

Outcomes

Which patterns align with success instead of failure?

Composition

How do local skills become procedures and strategies?

Method

Trace, induce, compose.

Three stages preserve local semantics without losing temporal structure or outcome evidence.

01 / Raw traces to events

Segment interactions around meaningful state changes.

Benchmark adapters replace task-specific entities with typed arguments while retaining subgoals, action families, observed state changes, and feedback.

Preserves execution order
Shares behavior across objects and tasks
Retains successful and failed evidence

Event-level abstraction from successful and failed WebShop trajectories

Results

Higher success with less execution overhead.

Trace2Tower improves task mastery while reducing unnecessary steps, invalid actions, and injected skill context.

87.31%ALFWorld success+5.72 vs. ExpeL

50.67%WebShop exact successBest automatic method

10.35ALFWorld stepsShortest execution

0.26Invalid actionsLowest on ALFWorld

Structural ablation results showing the effect of transition, outcome, and contrastive evidence — **Transition evidence matters most.** Removing it drops ALFWorld success from 87.31% to 70.15% and sharply reduces the induced hierarchy.

Success and token cost trade-offs on ALFWorld and WebShop — **Structure improves the quality-cost trade-off.** High-only minimizes context while Full maximizes success.

ALFWorld success rates by task family and method — Performance remains strong across all six ALFWorld task families.

Deployment success before and after verifier-guided graph refinement — Verifier-guided graph edits raise pooled success from 80.42% to 84.58% without re-inducing the Tower.

Statistics describing procedure size, strategy reuse, and strategy path length — The hierarchy is compact: strategies typically compose only two to four procedures.

Skill Tower

Explore the induced hierarchy.

Strategy skill

Task-level successful compositions

ALFWorld

Maximal procedure paths supported by successful trajectories form compact plans for long-horizon execution.

Bind task entities and constraints
Plan a supported procedure path
Retrieve state-relevant procedures
Execute and refine from verifier feedback

Citation

Build on Trace2Tower.

Read as HTML

@misc{sun2026trace2tower,
  title         = {Trace2Tower: Transition-Aware EigenTrace Induction
                   of Multi-Level Skills for LLM Agents},
  author        = {Jiazheng Sun and Boyu Yang and Binhao Yuan and
                   Mingxuan Li and Xin Peng},
  year          = {2026},
  eprint        = {2609.05261},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2609.05261}
}