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

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

cs.AISep 4, 2026

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

Authors: Jiazheng Sun, Boyu Yang, Binhao Yuan, Mingxuan Li, Xin Peng

Organizations: College of Computer Science and Artificial Intelligence, Fudan University Shanghai 200433, China

Abstract

Large language model agents increasingly rely on execution traces to master complex interactive tasks. However, current paradigms are bottlenecked by shallow trajectory retrieval and flat skill summarization, fundamentally ignoring the temporal dependencies and outcome-conditioned topology of agent behavior. We introduce Trace2Tower, a transition-aware EigenTrace framework that distills raw trajectories into a robust skill hierarchy. Trace2Tower abstracts step-level interactions into canonical events, constructing a unified graph governed by semantic compatibility, transition dynamics, and outcome evidence. Through a novel contrastive spectral decomposition, it isolates stable, success-aligned behavioral modes while rigorously suppressing failure-prone shortcuts. These modes organically populate a dynamic skill tower of action templates, procedural routines, and overarching task strategies, continuously refined via verifier-guided feedback. On ALFWorld, Trace2Tower achieves 87.31% success requiring only 10.35 steps and 0.26 invalid actions; on WebShop, it reaches 50.67% exact success. Across both benchmarks, Trace2Tower significantly outperforms existing baselines in task mastery and context-efficient experience reuse.

Figures & tables

Figure 1: From experience reuse to behavioral structure induction. Existing methods transform trajectories into reusable insights or skills, but the underlying behavioral relations remain implicit. Trace2Tower recovers success-aligned behavioral structures through semantic, transition, and outcome-aware EigenTrace induction.

Figure 2: Event-level abstraction from raw agent trajectories. Trace2Tower converts successful and failed step-level executions into ordered canonical event sequences for subsequent behavioral structure induction.

Figure 3: Construction of the transition-aware EigenTrace graph. The outcome-conditioned edge affinities $A_{uv}^{\pm}$ geometrically integrate semantic compatibility $s_{uv}$ , transition dependency $t_{uv}^{\pm}$ , and outcome compatibility $o_{uv}^{\pm}$ , which are further gated by an edge mask $M_{uv}$ and scaled by evidence strength $\omega_{uv}$ to yield robust behavioral topologies.

Figure 4: From success-aligned EigenTrace decomposition to hierarchical skill tower induction.

Method	ALFWorld					WebShop
	Success % $\uparrow$	Steps $\downarrow$	Invalid $\downarrow$	Input Tokens $\downarrow$	Skill Context Chars $\downarrow$	Exact Success % $\uparrow$	Reward $\uparrow$	Invalid $\downarrow$
No-Skill	$46.27\pm 5.83$	15.58	0.58	49,052	0	$40.67\pm 0.58$	$0.6776\pm 0.0044$	$0.68\pm 0.03$
Expert-Crafted	$71.39\pm 4.11$	12.01	0.38	41,502	3,286	$48.00\pm 0.00$	$0.7183\pm 0.0000$	$0.07\pm 0.02$
Trace2Skill +Combined	$61.44\pm 2.15$	13.91	0.28	67,186	9,508	$40.67\pm 0.58$	$0.5580\pm 0.0084$	$2.39\pm 0.07$
Trace2Skill +Error	$61.19\pm 1.29$	14.19	0.37	73,447	11,447	$44.00\pm 0.00$	$0.5892\pm 0.0055$	$1.81\pm 0.02$
SkillX	$78.61\pm 3.53$	12.06	0.40	64,229	10,039	$\underline{49.33\pm 2.08}$	$0.6922\pm 0.0102$	$\underline{0.32\pm 0.04}$

Table 1: Overall performance and inference efficiency on ALFWorld and WebShop across three independent runs. Bold and underline denote the best and second-best automatic methods.

Figure 5: ALFWorld success rates (%) by task family and method. Outlined columns denote the two Trace2Tower deployment policies.

Figure 6: Structural ablation on ALFWorld: (a) success rate and (b) induced procedure and strategy skills.

Figure 7: Success–cost trade-offs on ALFWorld and WebShop over three runs.

Experience	Skill User
Experience	DeepSeek-V4 Flash	DeepSeek-V4 Pro
No-Skill	52.99	66.42
GPT-5.4 Tower	88.06	85.82
DeepSeek-V4 Flash Tower	65.67	79.10

Table 2: ALFWorld success (%) by experience condition and Skill User.

Figure 8: Deployment success and observed range over two held-out ALFWorld sets.

Figure 9: Hierarchy organization: (a) supporting event segments per procedure, (b) strategy-path reuse per procedure, and (c) procedures per strategy path.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Stage	Input	Main operation	Output
Event abstraction	Raw trajectory	Segment and canonicalize	$\mathcal{E}$
Graph induction	Canonical events	Semantic + transition + outcome evidence	$A^{+},A^{-}$
Contrastive graph	$A^{+},A^{-}$	Failure-aware transformation	$\bar{A}$
Procedure induction	$\bar{A}$	Component-wise spectral clustering	$\mathcal{S}^{(2)}$
Strategy induction	Procedure sequences	Directed graph + SCC paths	$\mathcal{S}^{(3)}$
Deployment	Skill tower	High plan + optional Mid support	Retrieved context

Appendix

Table 3: Trace2Tower representations and their roles.

Stage	Time	Working memory
Event abstraction	$O(S)$	$O(n)$
Embedding / quotienting	$O(nd)$	$O(nd)$
Graph statistics	$O(n+ed)$	$O(m+e)$
Component eigensystems	$O(\sum_{c}m_{c}^{3})$	$O(\max_{c}m_{c}^{2})$
Procedure clustering	$O(\sum_{c}m_{c}r_{c}^{2})$	$O(\max_{c}m_{c}r_{c})$
Strategy graph / SCCs	$O(P+E_{R}d+R+E_{R})$	$O(Rd+E_{R})$

Appendix

Table 4: Dominant time and working-memory costs of the reported pipeline.

Benchmark	Train tasks	Rollouts/task	Trajectories	Test tasks
ALFWorld	310	4	1,240	134
WebShop	100	4	400	100

Appendix

Table 5: Construction and evaluation pools. Both benchmarks use a 20-step interaction horizon.

Setting	Value
Outcome pseudo-counts	$1$ success + $1$ failure
Minimum nontrivial component	3 nodes
Zero-degree tolerance	$10^{-12}$
$K$ -means initializations	20
$K$ -means maximum iterations	300
Random seed	42

Appendix

Table 6: Essential graph-induction and deployment settings.

Variant	Retained signals	Mid/High	Success
Full	S+T+O+C	39/118	$87.31\pm 0.75$
No Transition	S+O+C	19/76	70.15
No Outcome	S+T+C	39/106	73.88
No Contrastive	S+T+O	10/44	73.88

Appendix

Table 7: ALFWorld structural ablations. S, T, O, and C denote semantic, transition, outcome, and contrastive evidence. Each ablation is one 134-task evaluation; Full is the three-run mean.

Edit	Target	Structural effect
Split	Heterogeneous procedure	Creates evidence-specific children while retaining the parent
Merge	Compatible procedures	Contracts evidence and supported relations into one procedure
Promote	Supported Mid path	Adds a reusable strategy motif
Downweight	Harmful High skill	Adds a reversible retrieval penalty without deleting the skill

Appendix

Table 8: Verifier-guided structural edit operators.

LLM Agent Skill Learning

Large language model agent skill learning studies how agents acquire, represent, reuse, and refine procedures for multi-step tasks.

See all 398 papers

Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills
Mar 26, 2026Jingwei Ni, Yihao Liu, Xinpeng Liu +7LLM Agent Skill Learning LLM Agent Self-Improvement
ETH Zürich · University of Zurich · Peking University +2
SKILL-DISCO: Distilling and Compiling Agent Traces into Reusable Procedural Skills
Jun 25, 2026Zhongxin Guo, Danrui Qi, Hanwen Gu +2LLM Agent Skill Learning Agent Skill Learning
Microsoft Research · Beijing Foreign Studies University
HiSkill: Empowering LLM Agents with Hierarchical Skill Graphs
Jul 28, 2026Yu Hao, Jinxuan Cai, Qi Zhang +4LLM Agent Skill Learning LLM Agent Skill Retrieval
Beijing University of Posts and Telecommunications · China Mobile Group Shaanxi Co., Ltd · Ant Group

Mar 26, 2026J/K move · Enter open · S save

Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills

Jingwei Ni, Yihao Liu, Xinpeng Liu, Yutao Sun+6

ETH Zürich · University of Zurich · Peking University+2

LLM Agent Skill Learning LLM Agent Self-Improvement

Open paper AskPDF arXiv

Large Language Model (LLM) agents increasingly rely on domain-specific skills, yet manually authoring such skills does not scale, and skills generated purely from parametric knowledge often miss critical operational pitfalls. We introduce Trace2Skill, a framework that consolidates broad execution trajectories in parallel into a unified skill directory through inductive reasoning over agent experience. Trace2Skill supports both deepening existing human-written skills and creating useful skills from weak LLM-generated drafts. Experiments demonstrate the effectiveness of Trace2Skill across diverse domains, including office workflows, math reasoning, and vision QA. Importantly, the evolved skills are not merely memorized artifacts of the trajectories used to create them: they often transfer across model scales, across model families, and to out-of-distribution settings. For example, skills evolved from Qwen3.5-35B trajectories improve a Qwen3.5-122B agent by up to

57.65

percentage points on WikiTableQuestions. Further analyses show that Trace2Skill outperforms sequential skill editing and ReasoningBank-style retrieval memories, compresses recurring failures and workarounds into standard operating procedures (SoPs), and yields portable skills that can be reused without parameter updates or test-time retrieval.

Abstract

LLM Agent Skill Learning

Explore similar work

Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills

SKILL-DISCO: Distilling and Compiling Agent Traces into Reusable Procedural Skills

HiSkill: Empowering LLM Agents with Hierarchical Skill Graphs