Continual Generative Information Retrieval on Natural Questions
80.34D0 ScoreLoRA-SPQ
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| LoRA-SPQAdaptation Stage=Stage 1: Base adaptation, Tuning Strategy=LoRA, DocID Design=SPQ2026.04 | 80.34 | 33.76 | 49.31 | 50.5 | 51.15 | 50.78 | 8.76 | 33.02 | -35.63 | |
| Full FT-SPQAdaptation Stage=Stage 1: Base adaptation, Tuning Strategy=Full FT, DocID Design=SPQ2026.04 | 80.34 | 55.71 | 62.93 | 62.66 | 64.31 | 65.93 | 15.38 | 41.28 | -35.95 | |
| PAMT-LoRA-SPQAdaptation Stage=Stage 2: Post-Adaptation Memory Tuning, Tuning Strategy=LoRA, DocID Design=SPQ2026.04 | 80.34 | 31.82 | 46.9 | 48.75 | 49.62 | 48.96 | 19.47 | 31.24 | -13.56 | |
| PAMT-Full FT-SPQAdaptation Stage=Stage 2: Post-Adaptation Memory Tuning, Tuning Strategy=Full FT, DocID Design=SPQ2026.04 | 80.34 | 53.25 | 60.17 | 60.4 | 62.18 | 63.46 | 27.83 | 39.15 | -14.21 | |
| DPR-HNEvaluation Protocol=Expanded protocol, Decoding Strategy=constrained decoding2026.04 | 78.41 | 73.99 | 76.21 | 70.98 | 71.08 | 69.7 | 39.5 | 42.15 | -4 | |
| MixLoRA-DSICategory=PEFT-based continual GenIR baselines, Evaluation Protocol=Expanded protocol2026.04 | 77.89 | 48.57 | 58.62 | 57.96 | 59.23 | 58.31 | 24.56 | 41.87 | -18.43 | |
| DPREvaluation Protocol=Expanded protocol, Decoding Strategy=constrained decoding2026.04 | 77.21 | 72.79 | 75.01 | 69.78 | 69.88 | 68.5 | 38 | 40.75 | -4.2 | |
| CLEVERCategory=Non-PEFT continual GenIR baselines, Evaluation Protocol=Expanded protocol2026.04 | 75.42 | 35.84 | 48.97 | 49.75 | 50.9 | 49.87 | 14.67 | 32.54 | -24.16 | |
| LoRA-TUAdaptation Stage=Stage 1: Base adaptation, Tuning Strategy=LoRA, DocID Design=TU2026.04 | 74.28 | 52.92 | 64.31 | 63.41 | 63.64 | 63.97 | 21.36 | 50.34 | -32.9 | |
| Full FT-TUAdaptation Stage=Stage 1: Base adaptation, Tuning Strategy=Full FT, DocID Design=TU2026.04 | 74.28 | 63.96 | 70.24 | 66.67 | 67.44 | 66.84 | 31.59 | 55.64 | -25.93 | |
| PAMT-LoRA-TUAdaptation Stage=Stage 2: Post-Adaptation Memory Tuning, Tuning Strategy=LoRA, DocID Design=TU2026.04 | 74.28 | 50.39 | 61.55 | 60.9 | 61.28 | 61.45 | 31.28 | 47.62 | -10.84 | |
| PAMT-Full FT-TUAdaptation Stage=Stage 2: Post-Adaptation Memory Tuning, Tuning Strategy=Full FT, DocID Design=TU2026.04 | 74.28 | 61.69 | 67.48 | 64.29 | 65.38 | 64.97 | 42.31 | 52.87 | -9.16 | |
| BM25Evaluation Protocol=Expanded protocol, Decoding Strategy=constrained decoding2026.04 | 73.78 | 71.57 | 73.52 | 68.55 | 69.06 | 67.89 | 44.87 | 47.58 | -4.49 | |
| DSI++Category=Non-PEFT continual GenIR baselines, Evaluation Protocol=Expanded protocol2026.04 | 72.15 | 28.57 | 41.38 | 43.28 | 44.62 | 43.51 | 11.23 | 27.86 | -28.74 | |
| CorpusBrain++Category=PEFT-based continual GenIR baselines, Evaluation Protocol=Expanded protocol2026.04 | 70.58 | 31.17 | 44.14 | 45.27 | 46.15 | 45.33 | 12.84 | 29.43 | -26.58 | |
| PromptDSICategory=PEFT-based continual GenIR baselines, Evaluation Protocol=Expanded protocol2026.04 | 68.92 | 27.27 | 39.66 | 41.04 | 42.31 | 41.56 | 10.56 | 25.72 | -30.21 |