DENSE: Distilling Agent Trajectories into Evidence-Grounded Shortcut Trees for Self-Refinement
Siyuan Liu; Fan Yu; Dongyu Ru; Yizhu Liu; Yifan Yang; Xuezhi Cao; Xunliang Cai; Yixin Cao
- Summary
- DENSE distills evidence from nested subtask executions to convert redundant, incomplete, and failed agent trajectories into structured shortcut trees. Evaluated via REFIT on Terminal-Bench 2.1, it raises strict pass rates by 7.12–15.64 per…
- Method
- DENSE beroperasi pada himpunan trajektori eksekusi $\mathcal{T} = \{t_1, t_2, \dots, t_n\}$ yang dihasilkan oleh agen saat mencoba suatu tugas. Setiap trajektori didekomposisi men…
- Results
- Di antara metode umpan balik tanpa supervisi hasil eksternal, DENSE mencapai strict pass rate tertinggi pada empat model agen di **Terminal-Bench 2.1**. Metode ini meningkatkan hasil dibanding upaya …