TPOUR: 72.7x smaller model beats Qwen-Embedding-8B on temporal retrieval

Kim et al. introduce TPOUR, a novel unsupervised dense retriever that tackles temporal relevance—a critical blind spot for existing models that retrieve semantically similar but temporally misaligned documents (e.g., returning 2020 articles for a query about the 2019 president). Their core innovation, Temporal Retrieval Preference Optimization (TRPO), reinterprets preference learning in the time dimension, guiding the retriever to favor temporally aligned documents without requiring explicit timestamp supervision. TPOUR further generalizes to unseen time periods via interpolation within a learned time embedding, enabling continuous temporal alignment across arbitrary date ranges.

In evaluations on temporal information retrieval (T-IR) benchmarks, TPOUR Contriever—despite being 72.7× smaller than Qwen-Embedding-8B—achieves a +4.04 gain in average nDCG@5 (+12.15%) on explicit queries and +4.98 (+15.21%) on implicit queries. The method outperforms both unsupervised and supervised baselines, demonstrating that temporal awareness can be learned without costly timestamped training data. The code has been open-sourced, and the paper was accepted at ICML 2026, signaling strong peer validation.