Temporal Preference Optimization for Unsupervised Retrieval

Unsupervised dense retrievers offer scalability by learning semantic similarity from unlabeled documents via contrastive learning, but they struggle to capture the temporal relevance, retrieving semantically related but temporally misaligned documents-an important aspect when a document collection spans multiple time periods (*e.g.,* retrieving documents from 2018-2025 for "Who is the president in 2019?" introduces temporal ambiguity). Existing methods rely on supervised training with explicit timestamps, which are not always feasible. We propose TPOUR (*Temporal Preference Optimization for Unsupervised Retriever*), which uses our novel training method *Temporal Retrieval Preference Optimization* (TRPO). TRPO reinterprets preference learning in the temporal dimension, guiding the retriever to favor temporally aligned documents. TPOUR further generalizes to unseen time periods via interpolation in a learned time embedding, enabling continuous temporal alignment. Experiments on temporal information retrieval (T-IR), TPOUR outperforms both unsupervised and supervised baselines. Compared to Qwen-Embedding-8B, despite being about 72.7$\times$ smaller, TPOUR Contriever improves average nDCG@5 by +4.04 (+12.15%) on explicit and +4.98 (+15.21%) on implicit queries. We provide our code at https://github.com/agwaBom/TPOUR.

Search systems are often asked questions whose answers depend on time, such as "Who was the president in 2019?" This is challenging because a document may be topically relevant but from the wrong time period, causing the system to return misleading results. Many existing methods address this by using supervised training data with explicit time labels, but such information is not always available for large document collections. We introduce TPOUR, a method for training search systems to better understand time without requiring labeled data. Our key idea is to teach the retriever to prefer documents from the correct time period over semantically similar but temporally mismatched documents. TPOUR can also adapt to intermediate or future time periods it has not directly seen during training by combining models trained for different time periods. In experiments on time-sensitive information retrieval, TPOUR retrieves more temporally appropriate documents than baselines. Even though our model is 72.7 times smaller than Qwen-Embedding-8B, it achieves stronger retrieval performance on both explicit and implicit temporal queries. We release our code to support future research on time-aware retrieval.