Temporal Preference Optimization for Unsupervised Retrieval

Abstract: 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 72.7x 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.

• The paper introduces TPOUR which integrates contrastive learning with time vector interpolation to align temporal signals in retrieval.
• It demonstrates significant gains in nDCG@5 over conventional models, ensuring temporal relevance in both implicit and explicit queries.
• The method enables continuous temporal generalization, allowing retrieval systems to interpolate and extrapolate to unseen time periods.

Temporal Preference Optimization for Unsupervised Retrieval: An Expert Analysis

Introduction

The paper "Temporal Preference Optimization for Unsupervised Retrieval" (2606.17664) addresses temporal misalignment in dense information retrieval by introducing TPOUR (Temporal Preference Optimization for Unsupervised Retriever). Temporal misalignment arises when retrievers that maximize semantic similarity retrieve documents that are not temporally aligned with the query intent, an especially acute problem for timestamped or time-sensitive document corpora. The authors propose a novel training methodology, Temporal Retrieval Preference Optimization (TRPO), that induces temporal awareness in unsupervised bi-encoder retrievers, enabling them to prefer temporally aligned documents even when queries contain implicit temporal signals.

Motivation and Limitations of Existing Methods

Contemporary unsupervised dense retrieval models, exemplified by Contriever, optimize for semantic correspondence but neglect temporal context, leading to retrieval of documents that are semantically but not temporally relevant. When queries are anchored in a specific temporal frame—explicitly ("Who was president in 2019?") or implicitly ("Who is the current president?")—retrieval performance suffers if the system is unaware of temporal alignment.

Supervised methods that encode timestamp constraints or leverage human labeling (e.g., BM25 or supervised bi-encoders) offer some temporal grounding but are constrained by the impracticality of large-scale human labeling and suffer from limited adaptability to queries outside their training distribution. Moreover, typical methods require discrete document-level timestamps or rely on explicit temporal annotations, which are often unavailable.

Methodology: Temporal Retrieval Preference Optimization (TRPO)

TPOUR's core methodological contribution is the TRPO objective, inspired by Direct Preference Optimization (DPO), but adapted to temporal signals present implicitly across temporally versioned corpora. It does this by constructing preference pairs $(D_t, D_{t'})$ for a given query $Q$, where $D_t$ and $D_{t'}$ are documents temporally aligned and misaligned, respectively, according to the timestamp of the corpus dump.

The total loss combines contrastive learning (semantic alignment) and TRPO (temporal preference):

$$L_{\text{total}} = \lambda L_{\text{CE}} + (1 - \lambda) L_{\text{TRPO}}$$

where $L_{\text{TRPO}}$ encourages the embedding similarity between a query and a temporally aligned document $D_t$ to exceed that for $D_{t'}$. This yields encoders that respect both semantic and temporal constraints without the need for explicit human annotation or timestamp supervision at the document level. The learning is enabled through versions of the corpus (e.g., Wikipedia dumps across years), using semantic drift as a surrogate temporal signal.

A momentum-based contrastive learning architecture (MoCo) is employed for stable large-batch negative construction, allowing efficient representation learning with limited compute.

Continuous Temporal Generalization: Time Vector Interpolation and Extrapolation

Discrete time period-specific retrievers are insufficient for continuous temporal modeling. TPOUR leverages time vector interpolation: for encoders fine-tuned on different temporal snapshots (e.g., 2018, 2021), "time vectors" are computed as the difference between their weights and the base encoder. An interpolated encoder for intermediate periods is synthesized via weighted combination of these time vectors, yielding temporal alignment for arbitrary (including unseen) time periods.

Notably, this enables TPOUR not only to interpolate between seen times but also to extrapolate toward future or unseen periods, by leveraging multi-period time vector arithmetic; performance for future periods can surpass retrievers trained solely on past data, provided sufficient continuous drift is present in the semantic distributions.

Timestamp Prediction via Ensemble of Temporally Specialized Encoders

A Mixture-of-TPOUR framework is constructed for timestamp inference, where a set of frozen TPOUR retrievers (each aligned to a distinct period) produce document embeddings, which are then concatenated and passed to a shared linear head. This method demonstrates superior timestamp classification accuracy compared to both parameter-matched baselines and much larger unsupervised methods, indicating that TPOUR embeddings capture strong temporal signatures.

Experimental Evaluation

Temporal QA and Retrieval Benchmarks

Evaluations cover SituatedQA (yearly temporal QA, 2018–2021), RealTimeQA (monthly, 2023), and the BEIR general retrieval benchmark (18 diverse datasets spanning multiple domains and creation years).

Key findings include:

• Substantial gains in temporal alignment: On SituatedQA (2018), TPOUR Contriever achieves nDCG@5 of 44.11 (implicit) vs. 29.89 from vanilla Contriever. Gains persist regardless of explicitness of temporal cues (+7.36 nDCG@5 over the strongest baseline in "timeless" queries).
• Parameter efficiency: TPOUR Contriever with 110M parameters outperforms Qwen-Embedding-8B (8B parameters) and all other large embedding models by significant margins, notably +3.01 to +14.99 nDCG@5 in various settings.
• Interpolation enables robust generalization: Interpolated encoders match or surpass period-specific fine-tuned models at intermediate time points, demonstrating efficient and flexible deployment across continuous time (e.g., +13.4 nDCG@5 over earliest retrainer for 2019 test set).
• Extrapolation to the future shows clear improvement compared to outdated checkpoints: On RealTimeQA (2023, Dec) the extrapolated TPOUR retriever (using only 2018/2021/2022 checkpoints) yields nDCG@5 of 30.00 vs. 21.45 for 2018 and 27.78 for 2022 retrainers.
• Temporal awareness in general IR: On BEIR, optimal interpolation weights for different datasets correlate with their publication years, demonstrating that even nominally "timeless" tasks benefit from temporal-aware retrievers when underlying distributions drift.

Ablations and Analysis

The ablation on the loss weight $\lambda$ reveals the necessity of a moderate balance: relying solely on semantic or temporal signals reduces performance, with optimal results for $\lambda$ in $Q$0. Larger contrastive queues confer mild benefit up to medium sizes, after which returns saturate.

Qualitative and distributional analyses confirm that TPOUR learns implicit temporal signals from semantic drift rather than explicit year mentions; the vast majority of training documents ($Q$197%) do not mention their year, yet the model robustly infers and aligns temporal context.

Implications and Future Directions

The results establish that temporal preference optimization is feasible and effective in dense unsupervised retrieval, addressing a long-standing deficiency regarding time sensitivity. TPOUR’s architecture demonstrates that document-level supervision or explicit time annotation is not a strict requirement—a significant advantage for web-scale or historic corpora.

Theoretical implications include positioning time as a latent axis in embedding and parameter space. Interpolations in weight space (a strategy previously used for task/domain adaptation and model averaging) are validated as effective for temporally continuous generalization as well.

Practically, deploying a Mixture-of-TPOUR ensemble or applying time vector interpolation as a retrieval-time operation allows retrieval systems to flexibly tune their temporal alignment per-query, contingent on inferred or specified temporal intent. This modularity enables time-sensitive applications (news, law, QA, time-aware search) to maintain accuracy despite data drift and evolving corpora.

Potential future developments:

• Extending TPOUR to less regularly timestamped document streams (forums, user-generated content) via coarse temporal supervision.
• Automated selection or prediction of optimal interpolation weights $Q$2 for implicit temporal queries using intent classification.
• Integrating TPOUR into LLM-based or RAG systems for stronger temporal grounding.
• Further investigation of time vector extrapolation for robust generalization into future periods, possibly leveraging domain-adaptive techniques.

Conclusion

Temporal Preference Optimization for Unsupervised Retrieval is a rigorous and comprehensive contribution to time-aware information retrieval. By reframing preference optimization in the temporal domain and leveraging unsupervised signals via corpus drift, TPOUR enables dense retrievers to capture both semantic and temporal relevance at scale. The demonstrated improvements in nDCG, timestamp prediction accuracy, and cross-temporal generalization establish TPOUR as a new paradigm for unsupervised, temporally grounded retrieval, with substantial applicability both in time-sensitive QA and broader information retrieval contexts (2606.17664).