Trace2Tower: Transition-Aware EigenTrace Induction of Multi-Level Skills for LLM Agents
Jiazheng Sun, Boyu Yang, Binhao Yuan, Mingxuan Li, Xin Peng
cs.AI · cs.SE
Quick verdict
Claims transition-aware spectral skill induction beats baselines on ALFWorld/WebShop; novel and well-theorized, but main-table significance testing is absent.
Summary
The paper proposes Trace2Tower, a framework that converts raw LLM-agent execution traces into a three-level skill hierarchy (action templates, procedures, strategies). It first segments trajectories into canonical events with typed arguments, then builds a directed graph over event identities whose edges geometrically combine semantic compatibility, success/failure-conditioned transition counts, and smoothed outcome tendencies. A parameter-free contrastive transformation (A-bar = A+ squared over A+ plus A-) suppresses failure-correlated edges before component-wise spectral decomposition; an eigengap rule selects the number of procedure clusters, and strategy skills are maximal success-supported paths over procedure graphs. The authors prove a contrastive-stability lemma and a Davis-Kahan sin-Theta bound for the induced eigenspaces. With GPT-5.4 as Skill Author and DeepSeek-V4-Flash as Skill User, Trace2Tower reports 87.31 plus/minus 0.75 percent success on 134 ALFWorld valid_unseen tasks and 50.67 plus/minus 2.89 percent exact success on a 100-task WebShop manifest, using fewer steps, invalid actions, and injected context than ExpeL, SkillX, and Trace2Skill baselines. Ablations attribute the largest drop to removing transition evidence (70.15 percent), and verifier-guided graph edits lift a frozen tower from 80.42 to 84.58 percent on held-out sets. Main limitations: the headline WebShop margin over SkillX (1.34 points) sits within reported standard deviations with no significance test, ablations are single-run, the embedding model is not identifiable from the paper, and evaluation covers only two text-based benchmarks.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Integrating semantic, transition, and outcome evidence in a contrastive spectral decomposition induces success-aligned behavioral structure that content-centric experience or skill systems leave implicit.
- With GPT-5.4 as Skill Author and DeepSeek-V4-Flash as Skill User, Trace2Tower Full reaches 87.31 plus/minus 0.75 percent success on 134 ALFWorld valid_unseen tasks and 50.67 plus/minus 2.89 percent exact success on a 100-task WebShop manifest, with fewer steps, invalid actions, and injected context than ExpeL, SkillX, and Trace2Skill.
- The induced tower transfers across author/user models: a Flash-authored tower lifts both users about 12.7 points, and the GPT-authored tower beats it 88.06 vs 65.67 (McNemar p=6.04e-7).
- Verifier-guided graph edits improve a frozen tower from 80.42 to 84.58 percent (TF-IDF selection) across two held-out ALFWorld sets without re-induction.
- Transition evidence is the single most critical signal: removing it drops ALFWorld success to 70.15 percent and shrinks the hierarchy from 39/118 to 19/76.
Strengths
Principled, parameter-free graph construction
Eq. (3) geometrically integrates semantic compatibility, success/failure-conditioned transition ratios, and smoothed outcome compatibility, so every retained edge needs joint support; the observation mask M and evidence discount omega discount weakly observed transitions without hand-tuned mixing weights.
Real theory attached to the mechanism
Lemma 1 (Appendix A) proves the contrastive transform g(x,y)=x^2/(x+y) is bounded by x, monotone in each argument, and 1-Lipschitz per coordinate, so failure suppression does not amplify estimation error; Theorem 1 (Appendix C) gives a Davis-Kahan sin-Theta bound for the induced eigenspace under an explicit eigengap condition.
Ablations isolate every graph signal
Cross-model decoupling with a proper paired test
Efficiency claims are quantified
Weaknesses
Headline WebShop claim is statistically unsupported
Inconsistent ablation protocol
Reproducibility gaps the authors themselves flag
Formalized objective never solved; closest relatives not compared
Narrow evaluation scope
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Trace2Tower — Preliminary Review
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Trace2Tower?
Trace2Tower is a framework from Fudan University that converts LLM-agent execution trajectories into a three-level skill hierarchy (action templates, procedures, strategies). It abstracts trajectories into canonical events, builds a graph whose edges combine semantic similarity, success/failure transition statistics, and outcome evidence, then applies a parameter-free contrastive spectral decomposition to find success-aligned behavioral modes while suppressing failure-prone patterns.
How well does Trace2Tower perform?
With GPT-5.4 writing the skills and DeepSeek-V4-Flash executing them, the paper reports 87.31 percent success on 134 unseen ALFWorld household tasks (10.35 steps and 0.26 invalid actions per episode) and 50.67 percent exact success on a fixed 100-task WebShop shopping manifest, using fewer steps, fewer invalid actions, and less injected context than ExpeL, SkillX, and Trace2Skill baselines.
How is Trace2Tower different from ExpeL or SkillX?
ExpeL and SkillX store and retrieve experiences or skill artifacts selected mainly by contextual relevance. Trace2Tower induces behavioral structure before skills exist: it explicitly models which event transitions lead to success versus failure and clusters events with spectral graph theory, so procedures and strategies inherit observed execution order and success alignment rather than relations imposed over pre-existing skill items.
Is code available for Trace2Tower?
The paper links a GitHub repository (FudanSELab/Trace2Tower). Note that the pipeline depends on GPT-5.4 and DeepSeek-V4-Flash API access, and the paper does not name the embedding model used for event representations, which the authors themselves flag as unrecoverable from their records.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure-embedded numerals (per-family bars in Figure 5, axes in Figures 7-9) are interleaved with caption/axis text by PDF extraction; figure-level readings here rely on the surrounding prose, which is intact.
- The main text cites 'Appendix G' for detailed configurations, but the provided appendix ends at Section F.2 on a complete sentence; configuration details actually appear in Appendix E.2, so this reads as a stale in-paper cross-reference rather than extraction truncation.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.