Parametric-verified Adaptive Information Retrieval (PAIRS)

Updated 10 August 2025

The paper introduces PAIRS, an architecture that integrates LLM parametric knowledge with external retrieval using a dual-generation verification mechanism.
It employs dual-path retrieval by generating both query-based and pseudo-context-based embeddings to enhance the precision and recall of selected documents.
Experimental results demonstrate that PAIRS boosts QA performance (e.g., +1.1% EM, +1.0% F1) while reducing retrieval operations by 25%, optimizing computational resources.

Parametric-verified Adaptive Information Retrieval and Selection (PAIRS) is an architecture for information retrieval and question answering systems that integrates parametric knowledge from LLMs with external, retrieved knowledge sources in an adaptive and efficiency-oriented manner. PAIRS provides a mechanism for dynamic decision-making—verifying when the internal memory of the model suffices for a query and adaptively controlling if and how retrieval should be performed. This dual approach addresses inefficiencies inherent in prior systems that blindly retrieve for every query and improves accuracy on complex or underspecified queries via context-guided, dynamic selection of external information.

1. Framework Overview: Integration of Parametric and Retrieved Knowledge

PAIRS operates on the principle that an LLM’s parametric knowledge is strong for well-represented facts and entities, but external retrieval remains crucial for long-tail, complex, or context-dependent queries. The system therefore employs a two-stage, dual-path generation mechanism:

Dual-generation for answer verification: The LLM generates a direct answer from the query (parametric path) and a context-augmented answer using a self-generated pseudo-context (contextual path).
If these answers converge (are identical or semantically similar), the system infers that parametric knowledge suffices and bypasses external retrieval.
If the answers diverge (indicating uncertainty or possible insufficiency in the parametric memory), PAIRS activates a dual-path retrieval mechanism and selection procedure to incorporate relevant external context (Chen et al., 6 Aug 2025).

This architecture is training-free and relies primarily on generative and retrieval components, minimizing unnecessary retrieval operations and improving system efficiency.

2. Dual-Path Generation and Verification Mechanism

The dual-path generation is central to parametric verification:

For a given query $q$ , the system first generates answer $\hat{a}_1 = \mathcal{G}(\mathcal{P}(q))$ (parametric) and self-generates a pseudo-context $p$ to produce $\hat{a}_2 = \mathcal{G}(\mathcal{P}(q, p))$ (context-augmented).
Verification decision: If $\hat{a}_1$ and $\hat{a}_2$ agree, retrieval is suppressed; if not, retrieval is triggered (Chen et al., 6 Aug 2025).
This dual answer path exploits the LLM’s ability to internally simulate context, paralleling recent adaptive decision strategies where LLMs emit special tokens when uncertain (e.g., $<$ RET $>$ , cf. (Labruna et al., 30 Apr 2024)), but here the agreement is evaluated by direct answer comparison.
Statistically, PAIRS reduces retrieval invocations by approximately $25\%$ , retrieving for only $75\%$ of queries on benchmark datasets (Chen et al., 6 Aug 2025).

This dual-generation setup contrasts with prior approaches that use static popularity thresholds (Mallen et al., 2022) or explicit uncertainty tagging (Labruna et al., 30 Apr 2024), and provides a more flexible, context-sensitive parametric verification routine.

3. Dual-Path Retrieval (DPR): Query and Pseudo-Context Guidance

When external knowledge is required, PAIRS activates a dual-path retrieval protocol:

Dual queries: Two embedding-based retrievals are launched—one using the query ( $q \rightarrow \mathfrak{f}(q)$ ) and the other using the LLM-generated pseudo-context ( $p \rightarrow \mathfrak{f}(p)$ ).
Top- $n$ documents from each retrieval path are merged, forming a candidate pool for selection.
Dual-path retrieval is particularly effective on sparse or underspecified queries, as pseudo-context responses typically introduce complementary signals not captured by the raw query alone (Chen et al., 6 Aug 2025).
This advances previous work where dynamic retrieval agents adaptively select numbers of retrieved documents per query (Kratzwald et al., 2018), but PAIRS leverages internal generative signals rather than relying solely on input features or external metrics.

By combining query and context-driven retrieval, PAIRS increases coverage and recall, especially in multihop and compositional reasoning settings (Su et al., 7 Jun 2025).

4. Adaptive Information Selection (AIS): Weighted Filtering and Joint Relevance

After assembling a candidate pool, PAIRS applies an Adaptive Information Selection (AIS) module to maximize joint relevance:

Geometric scoring: For each document $d$ , compute similarities to both query and pseudo-context:

$s_1(d,q) = \langle \mathfrak{f}(d), \mathfrak{f}(q) \rangle = \cos(\theta_1),\quad s_2(d,p) = \langle \mathfrak{f}(d), \mathfrak{f}(p) \rangle = \cos(\theta_2)$

Joint score:

$s(d) = \cos(\theta_1 + \theta_2) = s_1 s_2 - \sqrt{1-s_1^2}\sqrt{1-s_2^2}$

Documents with high joint alignment are favored; dynamic weighting (via $\alpha$ as a regression factor over the angular distance between $q$ and $p$ ) can further refine tradeoffs between query and context relevance (Chen et al., 6 Aug 2025).
This selection avoids simply summing scores or relying on conventional reranking alone and provides robustness on queries with low information signals, improving both precision and recall.

AIS complements existing pairwise reranking approaches (cf. (Gienapp et al., 2022)) by exploiting both parametric and context-informed relevance vectors.

5. Efficiency and Accuracy: Experimental Results and Evaluation

Empirical evaluation of PAIRS demonstrates:

On six QA benchmarks (HotpotQA, NaturalQA, WebQuestions, SQuAD, TriviaQA, 2WikiMultiHopQA), PAIRS achieves improved accuracy ( $+1.1\%$ EM, $+1.0\%$ F1 versus prior baselines) while reducing retrieval costs by $25\%$ (Chen et al., 6 Aug 2025).
The dual-path retrieval plus AIS selection outperform both standard RAG and recent dual-retrieval hybrids (e.g., HyDE, Q2D, CoT) on average EM and F1.
Additional reranking further boosts performance, with the DPR-AIS-rerank variant yielding $+2.3\%$ EM and $+2.5\%$ F1 gains.
This suggests that parametric verification and adaptive selection provide measurable improvements for QA and document retrieval in terms of both computational efficiency (reduced retrieval frequency) and answer quality.
PAIRS also shows robustness to sparse or underspecified queries, mitigating risks of irrelevant retrievals common in static systems (Chen et al., 6 Aug 2025).

6. Context: Adaptive RAG, Decision Strategies, and Verification

PAIRS is part of a broader movement in Retrieval-Augmented Generation towards dynamic, adaptive information access and parametric knowledge verification (Su et al., 7 Jun 2025):

Dynamic RAG: Retrieval is triggered iteratively based on real-time needs, as opposed to static, one-shot pipelines.
Parametric vs. non-parametric interplay: Decision strategies, such as threshold heuristics based on popularity (Mallen et al., 2022), special tokens for retrieval needs (Labruna et al., 30 Apr 2024), and comparative mediation analysis (Ghosh et al., 1 Oct 2024), all address how to verify the sufficiency of parametric memory.
Verification via dual-path generation: PAIRS innovates by directly comparing generative responses rather than relying solely on external scoring or metadata, providing a generalizable approach for selective retrieval activation.
The dual-path embedding selection process in AIS is complementary to techniques that align attention, causal mediation, or preference optimization to improve robustness in conflicting or context-rich information landscapes (Farahani et al., 7 Oct 2024, Zhang et al., 6 Aug 2024).

7. Future Directions: Modular Integration and Extensibility

Possible logical extensions and open research avenues based on the presented data include:

Integration with dynamic parameter injection (cf. Parametric RAG (Su et al., 7 Jun 2025)): Modules representing retrieved knowledge could be merged with base LLM parameters for deeper, more persistent grounding.
Active sampling and risk-sensitive adaptation: Techniques from active pairwise sampling (Mikhailiuk et al., 2020) and risk-sensitive configuration selection (Mothe et al., 2023) can be applied to further optimize retrieval triggers, candidate selection, and pipeline efficiency.
Generalization to new domains: As shown in out-of-distribution experiments (Zhang et al., 6 Aug 2024), the framework generalizes to settings with varying knowledge conflicts, sparse queries, or complex reasoning tasks.

A plausible implication is that further improvements in adaptive interleaving of parametric verification, dual-context retrieval, and multi-stage selection will push the field towards more efficient and accurate information retrieval and question answering systems.

PAIRS advances retrieval-augmented architectures by implementing dual-path parametric verification, context-guided dynamic retrieval, and adaptive information selection, yielding demonstrable gains in QA efficiency and accuracy while mitigating the risk and cost of unnecessary retrievals (Chen et al., 6 Aug 2025). This framework provides a foundation for scalable, self-improving retrieval in LLM-powered systems.