SkillRet: leaderboard
Skill retrieval for LLM agents (Kang, Cho and Kim, 2026): for each of 4,392 long natural-language requests an embedding retriever must rank the one to three needed skills first among 6,006 public agent skills. NDCG@10 on a 0-100 scale on the audited evaluation set; higher is better.
Metric: NDCG@10 (0-100; first-stage dense retrieval of the gold skills for 4,392 LLM-agent requests from a pool of 6,006 skills, corrected evaluation set; binary relevance). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | SkillRet-Embedding-8B | 86.44 | |
| 2 | SkillRet-Embedding-0.6B | 81.12 | |
| 3 | SkillRouter-Embedding-0.6B | 73.54 | |
| 4 | harrier-oss-v1-0.6b | 70.26 | |
| 5 | Octen-Embedding-8B | 65.93 | |
| 6 | harrier-oss-v1-270m | 64.56 | |
| 7 | Qwen3-Embedding-8B | 63.64 | |
| 8 | jina-embeddings-v5-text-small | 62.96 | |
| 9 | Qwen3-Embedding-0.6B | 61.94 | |
| 10 | bge-large-en-v1.5 | 59 |
Interactive version: theaggregate.ai/benchmark?slug=skillret · How It Works · Data refreshed daily, snapshot 2026-11-12.