HSE DeepSWE - Skill Ablation Claude Code: leaderboard
HSE v1 reports resolved rates on 500 SWE-bench Verified tasks and 113 DeepSWE tasks, using 3 runs. The default model is GLM-5.2 in the Pi or Claude Code harness. Its static library of 10 skills is evolved on SWE-Gym, with 8 success skills and 2 recovery skills. Gates 1–8 explicitly adapt to the target benchmark and remain distinct from frozen transfer. Cross-model GPT 5.5 and Opus 4.8 evaluations use Pi only. Model, harness, skill library, training scope, and gate identities are preserved. Full HSE and No Skills repeat Table 1 and are not re-emitted here. Metric: Resolved rate (%).
Metric: Resolved rate (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | HSE study GLM-5.2 Claude Code without Procedural Skills | 47.8 | |
| 2 | HSE study GLM-5.2 Claude Code without Recovery Skills | 46.9 |
Interactive version: theaggregate.ai/benchmark?slug=hse-deepswe-skill-ablation-claude-code · How It Works · Data refreshed daily, snapshot 2026-10-28.