Learning Globally Reusable Skills for Coding Agents¶
Core mechanism evidence. Paper and local evaluations are separate.
Paper information¶
| Field | Value |
|---|---|
| Paper | arXiv 2608.06153 |
| Affiliations | Tianjin University |
| First public release | 2026-08-06 (arXiv v1) |
| Authors' code | No public author code found as of 2026-08-09. |
| Adapter / method | gse |
| Local code | src/auto_research/agent_research/latest_20260809.py |
Original paper¶
Background and contribution¶
GSE maintains a Skill Relation Graph. Cluster local experience to consolidate skills and use replay verification to reject overfitting and regressions.
flowchart LR
E[Local skill evidence] --> G[Skill graph] --> C[Cluster and merge] --> R[Replay verification] --> S[Reusable skills]Key figure from the paper¶
Original Figure 1: Learning Globally Reusable Skills for Coding Agents. From the paper; copyright belongs to the authors. Not a local result.
Core mechanism¶
\(\max_{S,G}\;U(S)-\lambda\operatorname{Conflict}(G)-\mu\operatorname{Regression}(S)\).
Paper results¶
Paper test-generation precision/recall gains 6.1%–34.1% / 31.8%–180.0%; internal industrial agent F1 +61.4%.
Local reproduction¶
Compare with a same-suite control without the mechanism. Metrics; no committed checkpoint or raw run directory. Evolve discovery includes gse, but combinations remain constrained by public evaluation and budget.
Reproduction limits¶
Executes the skill evolution operator, not the paper model, private environment, or expensive judge.