Trace2Tower

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

Paper · Project Page · Website Source

Trace2Tower induces a hierarchy of action, procedure, and strategy skills from successful and failed agent trajectories. It combines semantic compatibility, transition structure, and outcome evidence through contrastive spectral decomposition.

Results

Benchmark	Success	Avg. steps	Invalid actions
ALFWorld	87.31%	10.35	0.26
WebShop	50.67% exact success	—	—

Setup

Requires Python 3.11–3.13 and an OpenAI-compatible chat/embedding provider.

git clone https://github.com/FudanSELab/Trace2Tower.git
cd Trace2Tower
uv sync --extra dev
cp .env.example .env

Quick reproduction

uv run python examples/webshop_real_sample/run.py \
  --dataset-root /path/to/webshop-data \
  --webshop-source-root /path/to/WebShop \
  --output-dir work/webshop-real-sample \
  --concurrency 4 \
  --max-steps 15

ALFWorld and WebShop construction/evaluation configs are under configs/. Released skill catalogs are under skills/.

Website

cd website
npm ci
npm run dev

Citation

@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}
}