Multilingual Retrieval on WSDM Cup 2026 (test)
0.403nDCG@20Low-Cost Multi-Stage Retrieval Pipeline
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Low-Cost Multi-Stage Retrieval PipelinePipeline=GRF expansion + BM25 + jina-embeddings-v4 + Qwen3-Reranker-4B2026.02 | 0.403 | 0.95 | |
| MILCORetrieval Type=Multilingual2026.02 | 0.395 | 0.93 | |
| JinaV3Retrieval Type=Multilingual2026.02 | 0.355 | 0.93 | |
| Arctic-Embed Large v2Retrieval Type=Dense2026.02 | 0.352 | 0.92 | |
| BM25Retrieval Type=Sparse2026.02 | 0.349 | 0.97 | |
| Qwen3 0.6B EmbedRetrieval Type=Dense2026.02 | 0.309 | 0.92 | |
| RepLlamaRetrieval Type=LLM-based2026.02 | 0.255 | 0.86 | |
| e5 LargeRetrieval Type=Dense2026.02 | 0.205 | 0.7 | |
| BGE-M3 SparseRetrieval Type=Sparse2026.02 | 0.054 | 0.3 |