Pseudo-Relevance Feedback in IR

• Pseudo-Relevance Feedback (PRF) is a technique that refines user queries by assuming the top retrieved documents are relevant, effectively addressing vocabulary mismatch.
• It leverages diverse methods ranging from classical Rocchio and RM3 approaches to advanced neural and transformer-based models for robust query expansion.
• Recent research focuses on mitigating query drift and enhancing robustness through selective feedback, LLM integration, and efficient model architectures.

Pseudo-Relevance Feedback (PRF) is a foundational technique in information retrieval (IR) for improving recall and precision by automatically reformulating user queries using information extracted from initial retrieval results. The core assumption is that the top-$k$ documents retrieved by the first-pass ranking are, on average, relevant to the query. PRF methods leverage these “pseudo-relevant” documents to enrich or reweight the query, mitigating vocabulary mismatch and other retrieval challenges across both sparse and dense representational modalities. Modern research has extended PRF beyond classic bag-of-words formulations to advanced neural architectures, LLMs, and efficient transformer-based models, resulting in robust, highly configurable feedback-driven retrieval pipelines.

1. Foundational Principles of Pseudo-Relevance Feedback

Classical PRF operates under two main assumptions: (1) the relevance assumption, that the highest-ranked documents are relevant; (2) the model assumption, that query reformulation occurs in a feature space (term-weight, embedding, or otherwise) compatible with the retrieval architecture (Tu et al., 29 Oct 2025). The canonical workflow comprises three stages: (i) initial retrieval with the original query $q$, (ii) extraction or construction of feedback signals from top-$k$ documents $F = \{d_1,\ldots,d_k\}$, (iii) reformulation of $q$ into an expanded or shifted query $q'$, which is used in a second retrieval step (Li et al., 2024, Li et al., 2021). Classical approaches include:

• Rocchio (vector-space) update: $q' = \alpha q + \frac{\beta}{k} \sum_{i=1}^k d_i$, with tunable hyperparameters $\alpha, \beta$ (Jedidi et al., 11 Mar 2026).
• RM3 (relevance model): Interpolates the unigram distributions of $q$ and $F$: $q$0 (Mackie et al., 2023).
• Dense retrieval (vector PRF): Aggregates feedback in the embedding space, e.g., via Rocchio-style vector averaging.

Extensions to these formalisms (see Section 3) address limitations related to signal drift, representation transferability, and robustness to noisy feedback.

2. Methodological Extensions and Architectural Innovations

Recent PRF research encompasses a rich design space, spanning both classical and neural retrieval. Notable architectural innovations include:

• Transformer-based PRF models: TPRF leverages a lightweight transformer, interposed between two rounds of dense retrieval, which aggregates the initial query and feedback vectors by stacking them and processing the resulting matrix through multiple self-attention layers. The final refined query embedding achieves near state-of-the-art effectiveness at a fraction of the memory and latency cost of full-scale LLM encoders. TPRF’s inference time remains invariant with the number of feedback passages (i.e., agnostic to $q$1), and its memory footprint can be tuned by adjusting the number of layers and attention heads (Li et al., 2024).
• Text classification–based PRF: Simple classifier-based PRF methods treat the top-$q$2 documents as pseudo-positives and bottom-$q$3 as pseudo-negatives to train a query-specific linear model, which is then used to rerank the candidate set. This approach delivers additive improvements over both BM25 and RM3 baselines (Lin, 2019).
• Deep learned feedback fusion: NPRF (Neural PRF) reframes feedback as a document–document matching problem, using learned interaction models (e.g., DRMM, K-NRM) to score target documents relative to each feedback document. Neural gating or learned aggregation produces the final ranking signal (Li et al., 2018).
• Attention-based and graph-structured models: PRF can be cast as a sparse graph of sequence nodes (PGT), with efficient [CLS]-token attention linking the query, candidate, and feedback nodes, reducing the computational complexity over full-transformer architectures while retaining the ability to condition on multiple feedback documents. PGT enables tractable PRF with more feedback passages and richer context (Yu et al., 2021). ColBERT-PRF extends late-interaction dense retrieval by clustering token-level feedback embeddings, weighting them by IDF, and incorporating them into the expanded query representation (Wang et al., 2021).
• PRF with LLMs: Corpus-only, LLM-only, and hybrid PRF methods have been systematically studied, revealing that LLM-only (e.g., HyDE) query expansion is highly cost-effective, and hybrid strategies (combining corpus and LLM-generated feedback) often outperform single-source approaches. The impact of the feedback modeling mechanism (e.g., Rocchio vs. RM3) is often more significant than the feedback source itself, particularly for LLM-generated expansions (Jedidi et al., 11 Mar 2026).

3. Managing Drift and Robustness to Noisy Feedback

Query drift—where expansion vectors or terms diverge from the original intent due to noisy feedback—is a central challenge in PRF design. Recent works address these effects:

• Selective PRF via learned decision models: Rather than applying expansion indiscriminately, models such as Deep-SRF-BERT predict, using a transformer-based bi-encoder plus sequence aggregator, whether PRF will yield an effectiveness gain for a query. The model produces a confidence score $q$4 that enables either hard selection or reciprocal-rank fusion between original and expanded retrieval results, significantly reducing unwanted drift; accuracy in selecting when to expand approaches the oracle ceiling (Datta et al., 2024).
• Comparative regularization (Loss-over-Loss): The LoL framework enforces the “comparison principle”: reformulations based on more feedback should not worsen retrieval loss. It achieves this by directly penalizing cases where increasing the feedback depth $q$5 results in larger retrieval loss, jointly optimizing over multiple reformulation depths and yielding more robust performance across variable feedback quality (Zhu et al., 2022).
• Feedback signal quality studies: Empirical investigations show that classic bag-of-words Rocchio is highly sensitive to feedback signal quality and frequently degrades under moderate or weak feedback. Dense vector–based PRF is more robust, but only learned PRF encoders can reliably tolerate fully noisy feedback sets, retaining effectiveness even when the assumed relevant set contains mostly non-relevant documents (Li et al., 2022).
• LLM-assisted filtering: A hybrid of RM3 and LLM-based pseudo-labeling filters out noisy top-$q$6 documents prior to feedback modeling, thereby grounding expansions in corpus evidence and eliminating hallucination or topic drift. This method outperforms blind PRF baselines across multiple datasets (Otero et al., 16 Jan 2026).

4. Generative and Advanced Neural Feedback Mechanisms

Recent PRF paradigms employ strong generative architectures and utility-oriented learning:

• Generative PRF (GRF, GPRF): LLMs are prompted to generate synthetic feedback documents or direct query rewrites, which are then used as expansion evidence. Hybrid approaches—fusing PRF and GRF ranking signals—deliver complementary gains in both recall and precision. The “Generalized PRF” (GPRF) framework leverages utility-optimized, reinforcement learning–driven LLMs to generate rewrites that maximize retrieval metrics, thus relaxing both the relevance and model assumptions of classic PRF, and consistently outperforming RM3, Vector-PRF, and zero-shot GRF baselines (Mackie et al., 2023, Tu et al., 29 Oct 2025).
• QA-formulated feedback: The QA4PRF approach recasts PRF as a question-answering task, using attention-based pointer networks to extract semantically salient expansion terms directly from the feedback context, rather than relying on term frequency or statistical association alone. Integration with learned ranking models (LambdaRank module) further refines term selection (Ma et al., 2021).
• Adversarial and generative expansion models: Generative Query Expansion with PRF (GQE-PRF) uses conditional GANs over sequence models (e.g. BART) to generate expansion tokens conditioned on the query and feedback documents, outperforming extractive RM3/PRF baselines on retrieval and reranking metrics (Huang et al., 2021).

5. Resource-Efficiency and Practical Deployment

Hardware efficiency, latency, and scalability are critical for deploying PRF at scale:

• Transformer-based feedback fusion: TPRF demonstrates that a small transformer can provide almost all the recall gains of a full cross-encoder for dense feedback, reducing model size from >500 MB to as little as 62 MB and query batch latency from 100 s to <0.2 s/100 queries on CPU. Effectiveness is robust across configurations, with performance-efficiency trade-offs tunable via the number of layers and attention heads (Li et al., 2024).
• Prompt-based offline PRF: PromptPRF integrates LLM-generated features (entities, keywords, summaries) offline into dense query representations, permitting small dense retrievers (3B parameters) to achieve effectiveness on par with or exceeding larger (8B+) baselines at 63% reduced latency (Li et al., 19 Mar 2025).
• Offline PRF with pseudo-queries: By shifting heavy computation to offline pseudo-query generation and caching, fast single-pass dense retrieval can achieve the effectiveness gains of standard two-pass dense PRF at an order-of-magnitude lower online latency (Wen et al., 2023).
• Per-query online distillation: Performed at query time, student models (e.g., sparse lexical models) can be distilled from cross-encoder outputs and used for efficient second-stage retrieval, matching or exceeding established PRF baselines in both effectiveness and recall (MacAvaney et al., 2023).

6. Application Areas and Task-specific Adaptations

Beyond ad hoc document and passage retrieval, PRF has been successfully adapted to other IR tasks:

• Query categorization (APRF-Net): PRF engines, layered with hierarchical attention over multiple document and field levels, enrich query representations for fine-grained product categorization, especially benefitting rare or tail queries (Ahmadvand et al., 2021).
• Multi-vector and late-interaction dense retrieval: Expansion embeddings derived from feedback clustering (ColBERT-PRF) enable expansion beyond the term level, integrating semantic themes and improving retrieval for multiple-representation dense models (Wang et al., 2021).
• Domain and cross-modal expansion: GPRF's model-agnostic rewriting and ColBERT-PRF’s centroids have been shown to generalize across domains (in-domain and BEIR) and retrieval backends, affirming PRF’s applicability to broad information access tasks (Tu et al., 29 Oct 2025, Wang et al., 2021).

7. Open Challenges and Future Research Directions

Despite strong recent progress, several challenges remain:

• Adaptive feedback depth and weighting: Optimal PRF depth ($q$7) is query-dependent; learning to predict or adapt $q$8 per query remains an open problem (Li et al., 2024).
• Drift mitigation and interpretability: Continued development of comparative, confidence-calibrated, and hybrid LLM-based PRF is needed to robustly suppress drift while maintaining both interpretability and effectiveness (Zhu et al., 2022, Datta et al., 2024, Otero et al., 16 Jan 2026).
• Cross-modal, end-to-end, and cost-aware pipelines: Extending PRF frameworks to support multi-modal evidence, joint retriever–expansion training, and cost-effective, hardware-aware deployment (including quantization/distillation) are active areas of research (Li et al., 2024, Li et al., 19 Mar 2025).
• Integration with reranking and QA pipelines: Advances in integrating PRF directly into multi-stage retrieval and QA pipelines, possibly using reinforcement or utility-based training objectives, are promising future directions (Tu et al., 29 Oct 2025, Ma et al., 2021).
• Extensive evaluation over new benchmarks: Broader and more rigorous assessment of PRF efficacy and efficiency across diverse retrieval tasks, collections, and operational settings is necessary to inform practical adoption (Jedidi et al., 11 Mar 2026, Li et al., 19 Mar 2025).

In summary, pseudo-relevance feedback remains a cornerstone of information retrieval, with ongoing advances in neural and generative architectures, robustness mechanisms, hybrid corpus/LLM strategies, and practical deployment methodologies. These innovations collectively position PRF as both a classic and continually evolving methodology for retrieval effectiveness and efficiency across the contemporary IR landscape.