Conversational Retrieval on TopiOCQA (test)
50.1NDCG@3ConvSearch-R1
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| ConvSearch-R1Retriever=Dense (ANCE), Backbone=Qwen2.5-3B, NS=false, NCoT=false2026.01 | 50.1 | 72 | 50.5 | 86.3 | |
| ACQORetriever=Dense (ANCE), Backbone=Qwen2.5-3B, NS=true, NCoT=true2026.01 | 39.4 | 65.6 | 36.6 | 85.1 | |
| vanilla RLRetriever=Dense (ANCE), Backbone=Qwen2.5-3B2026.01 | 38.3 | 62.1 | 34.5 | 81.1 | |
| ACQORetriever=Sparse (BM25), Backbone=Qwen2.5-3B, NS=true, NCoT=true2026.01 | 37.7 | 62.6 | 34.9 | 83.2 | |
| ConvSearch-R1Retriever=Sparse (BM25), Backbone=Qwen2.5-3B, NS=false, NCoT=false2026.01 | 36.2 | 59.6 | 37.8 | 80.1 | |
| vanilla RLRetriever=Sparse (BM25), Backbone=Qwen2.5-3B2026.01 | 36.1 | 60.8 | 31.2 | 82.5 | |
| DeepSeek-V3.1Retriever=Dense (ANCE)2026.01 | 30.8 | 56.3 | 28.4 | 77.8 | |
| ADACQRRetriever=Sparse (BM25), Backbone=T5-base+LLaMA7B, NS=false, NCoT=true2026.01 | 26.5 | 48.9 | 28.3 | 71.2 | |
| RETPORetriever=Sparse (BM25), Backbone=LLaMA2-7B, NS=false, NCoT=true2026.01 | 26.5 | 48.3 | 28.3 | 73.1 | |
| LLM4CS-RARRetriever=Sparse (BM25), Backbone=ChatGPT, NS=true, NCoT=false2026.01 | 26.4 | 48.4 | 27.9 | 71.1 | |
| CHIQ-FusionRetriever=Sparse (BM25), Backbone=T5-base+LLaMA2-7B, NS=false, NCoT=true2026.01 | 23.5 | 44.7 | 25.6 | — | |
| AdaQRRetriever=Sparse (BM25), Backbone=T5-base, NS=false, NCoT=true2026.01 | 18 | 37.1 | 20.3 | 66.2 | |
| DeepSeek-V3.1Retriever=Sparse (BM25)2026.01 | 17 | 36.7 | 15.5 | 65.3 | |
| IterCQRRetriever=Sparse (BM25), Backbone=T5-base, NS=false, NCoT=true2026.01 | 14.9 | 29.3 | 16.5 | 54.1 | |
| UniConvCategory=CDR, Architecture=LLM-based Decoder-only Dense Retriever (7B)2025.07 | 0.426 | 67.4 | — | — | |
| ChatRetrieverCategory=CDR, Architecture=LLM-based Decoder-only Dense Retriever (7B)2025.07 | 0.401 | 63.7 | — | — | |
| QRACDRCategory=CDR, Architecture=SLM-based Encoder-only Dense Retriever (110M)2025.07 | 0.365 | 57.1 | — | — | |
| Conv-GRITCategory=CDR, Architecture=LLM-based Decoder-only Dense Retriever (7B)2025.07 | 0.36 | 54.2 | — | — | |
| CHIQCategory=CQR, Architecture=LLM-based Conversational Query Rewriter (7B) + Ad-hoc Dense Retriever (110M)2025.07 | 0.322 | 53 | — | — | |
| RETPOCategory=CQR, Architecture=LLM-based Conversational Query Rewriter (7B) + Ad-hoc Dense Retriever (110M), Configuration=fine-tuned2025.07 | 0.289 | 49.6 | — | — | |
| Conv-ANCECategory=CDR, Architecture=SLM-based Encoder-only Dense Retriever (110M)2025.07 | 0.285 | 52.6 | — | — | |
| LLM4CSCategory=CQR, Architecture=LLM-based Conversational Query Rewriter (7B) + Ad-hoc Dense Retriever (110M)2025.07 | 0.267 | 43.3 | — | — | |
| ConvDRCategory=CDR, Architecture=SLM-based Encoder-only Dense Retriever (110M)2025.07 | 0.264 | 43.5 | — | — | |
| GRITCategory=CDR, Architecture=LLM-based Decoder-only Dense Retriever (7B)2025.07 | 0.173 | 30.9 | — | — | |
| E5Category=CDR, Architecture=LLM-based Decoder-only Dense Retriever (7B)2025.07 | 0.169 | 28.7 | — | — | |
| RepLLaMACategory=CDR, Architecture=LLM-based Decoder-only Dense Retriever (7B)2025.07 | 0.15 | 27.2 | — | — |