Technical Summary: Temporal Preference Optimization for Unsupervised Retrieval (TPOUR)

1. Problem Statement

Current unsupervised dense retrievers, such as Contriever, excel at learning semantic similarity from unlabeled documents via contrastive learning. However, they suffer from temporal misalignment. When a document collection spans multiple time periods, these models often retrieve documents that are semantically relevant but temporally incorrect. For instance, a query like "Who is the president in 2019?" or an implicit query like "Who is the current president?" may retrieve documents from 2025 or 2015 if the model relies solely on semantic similarity.

Existing solutions to this problem typically rely on supervised training with explicit timestamps or labeled query-document pairs. This approach is often impractical at scale due to the scarcity of high-quality labeled data and the difficulty of obtaining explicit temporal annotations for large, unlabeled corpora. Furthermore, purely unsupervised methods generally ignore temporal signals, optimizing only for semantic relevance.

2. Methodology

The authors propose TPOUR (Temporal Preference Optimization for Unsupervised Retriever), a framework that integrates temporal preferences into unsupervised retrieval without requiring explicit document-level timestamps or labeled query pairs. The core of TPOUR is a novel training method called Temporal Retrieval Preference Optimization (TRPO).

2.1. Temporal Retrieval Preference Optimization (TRPO)

TRPO reinterprets preference learning within the temporal dimension. Instead of relying on human-labeled preferences (as in Direct Preference Optimization for LLMs), TRPO constructs preference pairs from document corpora collected at different time periods.

• Input: A query $Q_i$, a temporally aligned document $D^t_i$ (from the target time period), and a temporally misaligned document $D^{t'}_i$ (from a different period).
• Mechanism: The model is trained to maximize the similarity between the query and the aligned document while minimizing similarity to the misaligned one.
• Loss Function: The total training objective combines a standard contrastive loss ($L_{CE}$) with the TRPO loss ($L_{TRPO}$):
  $$L_{total} = \lambda L_{CE} + (1 - \lambda) L_{TRPO}$$
  Where $L_{TRPO}$ is defined based on the Bradley-Terry model, encouraging the model to prefer the aligned document over the misaligned one by aligning the score gap between the current model and a reference model (maintained via momentum).
• Data Source: The method utilizes temporally distinct document collections (e.g., Wikipedia dumps from 2018 and 2021). Crucially, the training data contains minimal explicit temporal markers (less than 2.5% of documents contain the target year), forcing the model to learn temporal preferences from semantic shifts in the content itself.

2.2. Continuous Temporal Representation via Interpolation

To address the limitation of discrete time periods and enable retrieval for unseen intermediate times without retraining, TPOUR employs time vector interpolation.

• Time Vectors: Defined as the difference between the weights of a model fine-tuned on a specific time period ($\theta_t$) and a base pre-trained model ($\theta_{base}$): $\tau_t = \theta_t - \theta_{base}$.
• Interpolation: To create a retriever for an intermediate time $t_{mid}$, the model interpolates between time vectors of adjacent periods ($t_{start}$ and $t_{end}$) using a coefficient $\alpha \in [0, 1]$:
  $$\theta_{tmid} = \theta_{base} + (1 - \alpha)\tau_{tstart} + \alpha\tau_{tend}$$
  This allows the retriever to generalize to intermediate time periods (e.g., 2019 or 2020) by simply interpolating the weights of models trained on 2018 and 2021 data.

2.3. Timestamp Inference and Extrapolation

• Inference: TPOUR can infer document timestamps using a "mixture-of-TPOUR" approach, where embeddings from multiple time-specific retrievers are concatenated and passed to a linear classifier.
• Extrapolation: The authors demonstrate that time vectors can be extrapolated to future periods by combining vectors from distinct past periods to approximate a "future-oriented" time vector, improving retrieval on recent data without training on that specific data.

3. Key Contributions

1. Identification of Temporal Misalignment: The paper demonstrates that existing time-unaware retrievers fail to distinguish between semantically similar but temporally misaligned documents in mixed-timestamp collections.
2. TPOUR Framework: The introduction of a preference-based training method (TRPO) that enables unsupervised retrievers to learn temporal relevance from unlabeled corpora collected at different times, without explicit timestamp supervision.
3. Continuous Generalization: Validation that time vector interpolation allows TPOUR-trained retrievers to generalize to intermediate and future time periods without additional training, effectively creating a continuous temporal alignment capability.
4. Temporal Sensitivity in General Retrieval: Evidence that temporal awareness improves performance not only on temporal QA tasks but also on general retrieval benchmarks (BEIR), where optimal performance correlates with the dataset's publication year.

4. Experimental Results

The authors evaluated TPOUR on SituatedQA (yearly temporal QA), RealTimeQA (monthly temporal QA), and the BEIR benchmark.

• Performance on Temporal QA: TPOUR significantly outperformed both unsupervised baselines (e.g., Contriever) and supervised baselines (e.g., DPR).
  • On SituatedQA (2018, implicit queries), TPOUR achieved an nDCG@5 of 44.11, surpassing Contriever (29.89) by a large margin.
  • TPOUR maintained strong performance across all years (2018–2021) and months, even for queries with implicit temporal information.
• Comparison with Large Models: Despite being approximately 72.7 times smaller than the Qwen-Embedding-8B model, TPOUR (based on Contriever) outperformed Qwen-Embedding-8B variants (including those with query rewriting and time-aware instructions) on both explicit and implicit queries.
  • TPOUR improved average nDCG@5 by +4.04 (+12.15%) on explicit queries and +4.98 (+15.21%) on implicit queries compared to Qwen-Embedding-8B.
• Interpolation Effectiveness: Interpolated models achieved performance comparable to or better than models directly fine-tuned on intermediate years. For example, interpolating between 2018 and 2021 models yielded an average improvement of +13.4 nDCG@5 over the start-year retriever on SituatedQA.
• Timestamp Prediction: A mixture-of-TPOUR predictor achieved 76.56% accuracy in predicting document years, outperforming a baseline predictor built on Contriever (50.18%) and a larger Nomic-Embed v2 MoE model (53.03%).
• BEIR Benchmark: Analysis showed that the optimal interpolation weight ($\alpha$) for BEIR datasets correlated with their creation years (e.g., older datasets like MS MARCO performed best with $\alpha=0.0$, while newer datasets like TREC-COVID performed best with $\alpha=1.0$), indicating that temporal modeling benefits general retrieval tasks.

5. Significance and Claims

The paper claims that TPOUR offers a scalable solution to temporal misalignment in information retrieval by leveraging unsupervised preference optimization. Its significance lies in:

• Eliminating the need for labeled data: It achieves temporal awareness using only unlabeled document corpora from different time periods, making it feasible for large-scale deployment where labeled temporal data is unavailable.
• Continuous Adaptability: The ability to interpolate time vectors allows the system to adapt to any point in time (past, present, or future) without retraining, addressing the static nature of traditional retrievers.
• Efficiency: TPOUR demonstrates that a relatively small model (110M parameters) can outperform massive embedding models (8B parameters) in temporal retrieval tasks when equipped with the correct temporal preference training.

The authors conclude that temporal modeling is not just a niche requirement for temporal QA but a fundamental component that enhances retrieval quality across general domains, suggesting that future retrieval systems should inherently account for the temporal context of both queries and documents.