Skill retrieval on LogicBench (Recall@1, Recall@10)
31.4Recall@1Qwen3-32B
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Qwen3-32BRetrieval Stage=LLM-based reranker, Candidate Pool=top-50 candidates retrieved by BM252026.04 | 31.4 | 55.3 | |
| Qwen3-235BRetrieval Stage=LLM-based reranker, Candidate Pool=top-50 candidates retrieved by BM252026.04 | 30.9 | 56.4 | |
| Llama-3.3-70BRetrieval Stage=LLM-based reranker, Candidate Pool=top-50 candidates retrieved by BM252026.04 | 27.4 | 55.4 | |
| Mistral3.1-24BRetrieval Stage=LLM-based reranker, Candidate Pool=top-50 candidates retrieved by BM252026.04 | 26.7 | 53.2 | |
| Qwen3-4BRetrieval Stage=LLM-based reranker, Candidate Pool=top-50 candidates retrieved by BM252026.04 | 21.8 | 43.3 | |
| Llama-3.1-8BRetrieval Stage=LLM-based reranker, Candidate Pool=top-50 candidates retrieved by BM252026.04 | 15.9 | 42 | |
| BM25Retrieval Stage=First-stage2026.04 | 12 | 36.1 | |
| HybridRetrieval Stage=First-stage2026.04 | 12 | 33.6 | |
| ContrieverRetrieval Stage=First-stage2026.04 | 5.5 | 18.4 | |
| BGERetrieval Stage=First-stage2026.04 | 4.1 | 20.5 | |
| TF-IDFRetrieval Stage=First-stage2026.04 | 1.8 | 18.2 |