File-level Retrieval on SWE-bench Verified (50 instances)
100Hit@1OracleGoldContext
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| OracleGoldContext2026.06 | 100 | 100 | 82.4 | 82.4 | 100 | 1 | |
| Qwen CausalRerankjudge=Qwen3.6-plus, rerank_candidates=top-502026.06 | 78 | 80 | 65.4 | 65.4 | 79 | 1.03 | |
| Qwen CausalHybridjudge=Qwen3.6-plus, rerank_candidates=top-502026.06 | 70 | 82 | 62.4 | 63.4 | 73.3 | 1.88 | |
| BM252026.06 | 58 | 80 | 55.4 | 63.4 | 63.4 | 2.43 | |
| HybridRAG2026.06 | 58 | 80 | 55.4 | 63.4 | 63.4 | 2.43 | |
| Deterministic CausalReranktype=older causal proxy2026.06 | 26 | 56 | 27.4 | 32.4 | 31.4 | 5.25 | |
| HashedEmbedding2026.06 | 18 | 30 | 16 | 16 | 19.2 | 5.07 | |
| Random2026.06 | 0 | 4 | 0 | 0.7 | 0.2 | 10.5 |