Research map / Label-efficient and robust learning / Recognition and long-tail
Temporal Preference Optimization for Unsupervised Retrieval
HyunJin Kim, Jaejun Shim, Young Jin Kim, JinYeong Bak
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train with TRPO preference loss on temporally aligned vs misaligned documents | Contriever · SituatedQA 2018 explicit | nDCG@5: 29.30 -> 43.93 | Improved |
| train with TRPO preference loss | Contriever · SituatedQA 2018 implicit | nDCG@5: 29.89 -> 44.11 | Improved |
| use TRPO preference loss instead of temporal contrastive loss (ablation) | Contriever · SituatedQA 2018 explicit | nDCG@5: 35.00 -> 43.93 | Improved |
| TPOUR Contriever versus 72.7x larger Qwen3-Embedding-8B with time-aware instruction | TPOUR Contriever · SituatedQA 2021 explicit | nDCG@5: 38.51 -> 40.21 | Improved |
| add query rewriting to Qwen3-Embedding-8B (ablation) | Qwen3-Embedding-8B · SituatedQA explicit 2018 vs 2021 | nDCG@5: 2018: 30.45 -> 40.92; 2021: 35.17 -> 34.46 | Mixed |
| add time-aware instruction to Qwen3-Embedding-8B (ablation) | Qwen3-Embedding-8B · SituatedQA explicit | nDCG@5: 2018: 30.45 -> 32.69; 2021: 35.17 -> 38.51 | Improved |
| interpolate time vectors at alpha 0.5 between 2018 and 2021 retrievers (ablation) | TPOUR Contriever · SituatedQA | nDCG@5: 29.05 / 31.72 -> 35.71 | Improved |
| use best interpolation alpha instead of directly fine-tuned retriever (ablation) | TPOUR Contriever · RealTimeQA | nDCG@5: 37.30 -> 38.77 | Improved |
| add time-vector extrapolation from 2018/2021/2022 vectors | TPOUR Contriever · RealTimeQA December | nDCG@5: 27.78 -> 30.00 | Improved |
| use mixture of TPOUR encoders for timestamp prediction | TPOUR · Year classification | accuracy (%): 50.18 -> 76.56 | Improved |
| use 10 encoders instead of 2 in mixture-of-TPOUR (ablation) | Mixture-of-TPOUR · Year classification | accuracy (%): 74.56 -> 76.56 | Improved |
| set loss weight lambda to moderate values 0.7-0.85 (ablation) | TPOUR · SituatedQA/RealTimeQA | nDCG | Improved |
Closest papers by meaning
- NUDGE: Lightweight Non-Parametric Fine-Tuning of Embeddings for Retrieval (ICLR 2025 · similarity 0.83)
- Benchmarking and Building Long-Context Retrieval Models with LoCo and M2-BERT (ICML 2024 · similarity 0.83)
- Semantic Feature Learning for Universal Unsupervised Cross-Domain Retrieval (NeurIPS 2024 · similarity 0.82)
- Contextual Document Embeddings (ICLR 2025 · similarity 0.82)
- Mitigating Semantic Collapse in Partially Relevant Video Retrieval (NeurIPS 2025 · similarity 0.82)
- MVR-cache: Optimizing Semantic Caching via Multi-Vector Retrieval and Learned Prompt Segmentation (ICML 2026 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
