Dynamic Passage Selector (DPS)

• Dynamic Passage Selector (DPS) is a supervised framework that dynamically selects a query-tailored subset of passages for retrieval-augmented generation.
• It leverages an autoregressive selection process to model inter-passage dependencies, ensuring comprehensive evidence collection while minimizing redundancy.
• Empirical evaluations show DPS boosts F1 scores by up to 30% over conventional reranking methods on benchmarks like MuSiQue and HotpotQA.

A Dynamic Passage Selector (DPS) refers to a supervised learning framework for adaptive passage selection that directly models inter-passage dependencies and dynamically selects a query-tailored subset of passages for retrieval-augmented generation (RAG) pipelines. Unlike conventional reranking systems that independently score and truncate with a rigid Top-K selection, DPS considers collective evidence across passages, allowing for more robust multi-document reasoning, especially in scenarios involving complex, multi-hop queries. DPS is architected to be a plug-and-play module and does not require modifications to downstream generators, facilitating seamless integration and empirical improvements across diverse tasks (Meng et al., 13 Aug 2025).

1. Design Principles and Framework

The central design of DPS reframes passage reranking as a selection (structured prediction) problem, focusing on directly predicting a minimal, sufficient subset of passages whose union best supports the query’s information need. For a query $q$ and a set of candidate passages $P = \{p_1, ..., p_n\}$, DPS models the conditional probability of selecting sequence $S = (i_1, ..., i_K)$ (where each $i_j$ is the index of a selected passage):

$$P_\theta(S | q, P) = \prod_{j=1}^K P_\theta(i_j | q, P, i_{<j})$$

Each selection step is autoregressive, conditioning on both the original query and all previously chosen passages. This mechanism allows the model to reason about complementarity and potential redundancy, distinguishing DPS from point-wise and list-wise rerankers that ignore inter-passage structure.

Selection decisions are trained using a sequence-level cross-entropy loss:

$$L(\theta) = -\frac{1}{M} \sum_m \sum_{j=1}^K \log P_\theta(i_j^* | q^{(m)}, P^{(m)}, i_{<j}^*)$$

where the target $S^*$ is a ground-truth sequence that jointly covers the evidential requirements of each training query.

2. Inter-Passage Dependency Modeling

A distinctive feature of DPS is the explicit conditioning of each passage selection on the context of previously chosen passages ($i_{<j}$). This enables the model to optimize for sets of passages that collectively provide comprehensive and complementary evidence while minimizing redundancy. Such dependency modeling is crucial for multi-hop QA and composite retrieval tasks, where answers require concatenated facts scattered across several documents—passages must be selected not simply for individual relevance but as a set that synthesizes necessary evidence.

This paradigm is in contrast to fixed Top-K selection, which may omit essential passages for low $K$ or admit distracting noise for high $K$, since the passage dependencies are not considered.

3. Performance Evaluation and Benchmarks

DPS demonstrates consistent improvements over competitive reranking and fine-tuning methods, as empirically validated across five standard benchmarks: HotpotQA, MuSiQue, 2WikiMQA, Legal, and CS. On the challenging MuSiQue dataset, DPS yielded a 30.06% improvement in F1-score compared to Qwen3-reranker and a 15.4% improvement over RankingGPT (Meng et al., 13 Aug 2025). This magnitude of gain suggests that the dynamic, query-specific adaptation and inter-passage reasoning substantially bolster answer generation quality for composite information needs.

DPS’s dynamic selection strategy results in robust evidence gathering for complex queries, while maintaining or surpassing competitive metrics (e.g., F1, Exact Match) on both multi-hop and single-hop tasks.

Dataset	DPS F1 Gain over Qwen3-reranker	DPS F1 Gain over RankingGPT
MuSiQue	30.06%	15.4%
Other Bench.	Consistent outperformance	Consistent outperformance

These gains reflect increased recall of relevant evidence while controlling for passage noise.

4. Integration in Retrieval-Augmented Generation Systems

DPS is developed as a seamless plug-and-play module. It operates post-retrieval and pre-generation, taking as input the output of a fast retriever and producing a structured, query-specific selection of passages. Its workflow requires no modification to existing downstream generators or to the rest of the RAG pipeline, and no generator fine-tuning is necessary.

This integration strategy allows practitioners to retrofit DPS into legacy systems or experimental architectures, immediately leveraging its adaptive selection features without costly reengineering. A plausible implication is the rapid deployment of dynamic passage selection improvements across existing RAG applications.

5. Impact on Reasoning Capabilities and Multihop QA

By dynamically selecting a minimal but sufficient set of passages, DPS enables RAG systems to reason over evidence that may be distributed across multiple documents. This is particularly impactful in multi-hop settings, where the answer synthesis depends on the interplay of facts and relations scattered throughout the corpus.

For single-hop or simple queries, DPS acts as a relevance-aware selector. For complex queries, the dynamic, structured approach allows the system to provide comprehensive coverage without unnecessary redundancy, optimizing the input context for generative models.

The adaptive nature of DPS means that it can select differing numbers of passages per query, balancing between information sufficiency and noise suppression. This suggests improved efficiency and answer quality over rigid fixed-size selection approaches.

6. Technical Distinction from Previous Methods

DPS differs from point-wise and list-wise rerankers in its supervision and selection formulation:

• Point-wise rerankers: Score passages independently; omit inter-passage dependency.
• List-wise rerankers: Optimize ranking distributions globally but still select Top-K with no inter-passage conditioning.
• DPS: Autoregressively selects passages, modeling dependencies at each selection step and dynamically determining the subset size.

This paradigm shift toward structured selection enables better reasoning with composite evidence and more adaptive passage set sizes tailored to query needs.

7. Directions for Further Research

Empirical evidence supports the adoption of DPS as a foundation for dynamic evidence selection in advanced RAG pipelines. Further research may explore optimizing the autoregressive selection process, extensions to domain-specific retrieval, and efficient scaling to very large candidate sets. A plausible implication is that more sophisticated dependency modeling—potentially integrating interaction graphs or attention mechanisms over passage sets—could further enhance evidence synthesis in generative QA and other composite reasoning tasks.

Continued benchmarking on diverse multi-hop, domain-specific, or fact-checking tasks will characterize the boundaries of DPS’s improvements and inform subsequent architectural refinements.