Multi-Task Retrieval-Augmented Generation

• Multi-task RAG is a framework that integrates evidence retrieval and answer synthesis across varied tasks to enhance accuracy and scalability.
• Architectural patterns range from decomposing documents into finer units to unifying retrieval and generation in a single model for efficient operation.
• Evaluation frameworks use metrics like nDCG@5, Recall@5, and faithfulness scores to assess both retrieval quality and generative performance.

Multi-task retrieval-augmented generation (RAG) denotes RAG systems that jointly optimize retrieval and generation across multiple tasks, domains, modalities, or interaction settings, rather than treating retrieval as a single fixed preprocessing step. In the literature, the term covers standalone retrieval and end-to-end generation in multi-turn conversations, long narrative search tasks with attribution, domain-specific enterprise retrieval served by one encoder across many use cases, multimodal open-domain tasks, and adaptive selection among retrieval strategies of different cost and depth (Elchafei et al., 1 May 2026, Upadhyay et al., 10 Mar 2026, Béchard et al., 8 Jan 2025, Liu et al., 24 Feb 2025, Tang et al., 2024). The field is therefore defined less by one canonical architecture than by a recurring design problem: how to couple evidence acquisition, evidence selection, and answer synthesis so that systems remain accurate, grounded, scalable, and efficient under heterogeneous task requirements.

1. Task scope and problem settings

Multi-task RAG has been instantiated in several distinct benchmark regimes. In SemEval-2026 Task 8 (MTRAGEval), Task A evaluates standalone retrieval quality and Task C evaluates end-to-end retrieval-augmented generation in multi-turn conversational settings, requiring both accurate answer generation and faithful grounding in retrieved evidence (Elchafei et al., 1 May 2026). In the TREC 2025 RAG Track, the task structure explicitly separates Retrieval (R), Augmented Generation (AG), Retrieval Augmented Generation (RAG), and Relevance Judgment (RJ), with systems required to return top-100 ranked document segments, generate well-formed answers using a fixed set of top-$k$ retrieved segments with sentence-level citations, or perform end-to-end pipelines that map all evidence to segment IDs (Upadhyay et al., 10 Mar 2026).

Open-domain multimodal work broadens this scope further. The M$^2$RAG benchmark comprises four tasks—image captioning, multi-modal question answering, multi-modal fact verification, and image reranking—and requires models to retrieve query-relevant information from a multi-modal document collection and use it as input context for RAG modeling (Liu et al., 24 Feb 2025). Other systems frame multi-task RAG through application structure rather than benchmark design. MARAGS was built for the CRAG competition, whose three subtasks combine standard web RAG, web plus knowledge-graph API use, and a larger-scale setting with 50 HTML documents (DeHaven, 2024). Enterprise-oriented retriever tuning defines the problem as supporting multiple retrieval tasks such as steps, tables, table fields, and catalog items with one small encoder (Béchard et al., 8 Jan 2025).

These settings share a common feature: retrieval and generation are evaluated both independently and in combination. This suggests that “multi-task” in RAG refers not only to downstream task variety, but also to the coexistence of retrieval-centric, generation-centric, and attribution-centric objectives within one system.

2. Architectural patterns

A first major pattern is decomposition of retrieval and generation into different units of representation. H-RAG uses a hierarchical parent-child pipeline in which each document is split into overlapping sentence-based child chunks for fine-grained retrieval, while the entire original document is stored as a parent unit for generation (Elchafei et al., 1 May 2026). Retrieval operates on child chunks, generation consumes aggregated parent documents, and parent relevance is computed by the maximum score among retrieved child chunks: $$\text{Score(parent)}=\max_{\text{child in parent}} \text{Score(child)}.$$ This decoupling is intended to balance retrieval precision and contextual grounding (Elchafei et al., 1 May 2026).

A second pattern is unification of retrieval and generation within a single model. ImpRAG is a query-free RAG system that divides a decoder-only LLM into three specialized layer groups: a bottom group optimized for retrieval, a middle group that reads retrieved passages through cross-attention, and a top group that decodes the response (Zhang et al., 2 Jun 2025). It eliminates explicit textual queries by deriving a latent query embedding from the model’s hidden states and scores passages with

$s(q,p)=E_q \cdot E_p.$

RoleRAG pursues another unified design: six modules are all driven by the same underlying LLM, distinguished only by optimized role-specific tokens, with modules for query graph building, retrieval judgment, sub-answer generation, summarization, new-query generation, and final answer reasoning (Zhu et al., 21 May 2025).

A third pattern is modular specialization without full model duplication. MARAGS uses a single Llama 3 8B backbone with different LoRa adapters for each CRAG subtask and for API-call generation, allowing task-specific specialization while keeping one base model in memory (DeHaven, 2024). DuetRAG also formalizes specialization, but at the level of answer pathways: an internal knowledge model, an external knowledge model, and a referee model collaborate so that internalized domain knowledge and retrieved evidence can be compared or synthesized (Jiao et al., 2024).

Across these systems, a plausible implication is that multi-task RAG architectures are converging on a small set of solutions to the same structural problem: separate what benefits from specialization, but minimize mismatch between retrieval-time and generation-time representations.

3. Retrieval adaptation and evidence selection

Retrieval adaptation is central because many of the cited systems identify fixed-$k$, single-pass retrieval as a bottleneck. H-RAG combines dense and sparse retrieval at the child-chunk level, with an interpolation parameter $\alpha$ controlling the weighting and a cosine-similarity rescoring stage over chunk embeddings (Elchafei et al., 1 May 2026). SAM-RAG generalizes this idea to dynamic retrieval in multimodal settings: instead of retrieving a fixed number of documents, it continues retrieving in batches until relevant information is found, then iteratively verifies answer usability and support through isRel, isUse, and isSup modules (Zhai, 2024).

Adaptive selection can also be posed as a policy problem. MBA-RAG frames retrieval strategy selection as a multi-armed bandit problem, where each retrieval method is an arm and the reward balances answer accuracy and retrieval cost: $r_a=\mathcal{A}(y,\hat{y}_a)-\lambda C(a).$ This formulation is explicitly intended to avoid indiscriminate retrieval and rigid single-class routing across queries of varying complexity (Tang et al., 2024).

Another line of work adapts retrieval through better shared encoders. Multi-task retriever fine-tuning for enterprise RAG instruction-tunes a 305M-parameter mGTE-base encoder on a variety of domain-specific retrieval tasks so that one encoder can serve many use cases with low cost, scalability, and speed (Béchard et al., 8 Jan 2025). The same paper emphasizes hard negative mining, class-imbalance mitigation, and instruction-style supervision as critical for structured enterprise data (Béchard et al., 8 Jan 2025).

Graph-based approaches replace flat passage ranking with structured expansion. LinearRAG constructs a relation-free hierarchical “Tri-Graph” over entities, sentences, and passages, then performs local semantic bridging followed by global importance aggregation through Personalized PageRank (Zhuang et al., 11 Oct 2025). TIGRAG instead builds a token co-occurrence knowledge graph, uses Personalized PageRank for semantic expansion, and then performs graph-weighted candidate scoring and neural reranking; it also introduces iterative entity-driven retrieval for multi-hop reasoning (Bonifazi et al., 29 Jun 2026). In the multimodal LVLM setting, “mRAG: Elucidating the Design Space of Multi-modal Retrieval-Augmented Generation” shows that listwise re-ranking with a prompted LVLM mitigates positional bias better than pointwise or pairwise strategies, and that supplying only the most relevant document can outperform feeding more context to the generator (Hu et al., 29 May 2025).

One recurring misconception is that better RAG always follows from retrieving more context. Several of these works argue otherwise: fixed-$k$ retrieval often introduces noise, dynamic filtering can improve both retrieval and generation, and in multimodal generation accuracy may decline when additional retrieved documents are appended (Zhai, 2024, Hu et al., 29 May 2025).

4. Multi-agent and cooperative formulations

A substantial branch of multi-task RAG treats the pipeline as collaboration among specialized agents. MA-RAG defines Planner, Step Definer, Extractor, and QA Agents, invoked on demand to handle query disambiguation, subquery construction, evidence extraction, and answer synthesis in complex or ambiguous information-seeking tasks (Nguyen et al., 26 May 2025). MASS-RAG uses three role-specialized evidence-processing agents—summarizer, extractor, and reasoner—whose outputs are integrated by an overview stage, optionally with per-view answer generation before final synthesis (Xiao et al., 20 Apr 2026). SEMA-RAG adapts this pattern to medical reasoning with an Interpreter Agent for clinical schema interpretation, an Explorer Agent for sufficiency-driven self-evolving retrieval, and an Arbiter Agent for evidence adjudication and answer selection (Huang et al., 16 May 2026).

Other papers define cooperation more formally. CoRAG argues that conventional RAG is ranking-centric and asymmetric, with the generator overly dependent on reranking results, and reformulates RAG as a cooperative multi-agent decision-making problem in which reranker and generator are peer decision-makers jointly optimized toward downstream response quality (Song et al., 21 Feb 2026). The objective is written as an expected task reward over reranker-selected documents and generated answers: $$\max_{\theta,\phi}\; \mathbb{E}_{D,\hat{a}}[R(a^*,\hat{a})].$$ DuetRAG likewise treats internal knowledge use and external retrieval as complementary pathways resolved by an arbiter or referee model (Jiao et al., 2024).

The CIIR LiveRAG system, titled mRAG, decomposes the pipeline into coordinator, planner, searcher, summarizer, reasoner, validator, and generator/reviser agents, and then optimizes inter-agent collaboration through self-training with reward-guided trajectory sampling (Salemi et al., 12 Jun 2025). This work is explicit that the coordinator manages trajectories adaptively rather than enforcing a fixed agent order (Salemi et al., 12 Jun 2025).

Taken together, these systems suggest that multi-task RAG increasingly treats “retrieval,” “reasoning,” and “generation” as heterogeneous cognitive operations whose coupling must be learned or orchestrated rather than hard-coded.

5. Evaluation frameworks and metrics

Evaluation in multi-task RAG is notably multi-layered. In SemEval MTRAGEval, retrieval is measured by nDCG@5 and Recall@5, while generation is scored with RB$_{agg}$ for document-level grounding, RL$^2$0 for faithfulness, RB$^2$1 for reference-less answer quality, and a harmonic mean that integrates evidence quality and answer generation (Elchafei et al., 1 May 2026). The TREC 2025 RAG Track extends evaluation to narrative queries and uses a layered framework including relevance assessment, response completeness, attribution verification, and agreement analysis (Upadhyay et al., 10 Mar 2026).

TREC’s response evaluation is particularly explicit about decomposition. It uses nugget extraction and sub-narrative assignment, measures strict vital recall and sub-narrative coverage, and evaluates support at sentence level with weighted precision and weighted recall over citations (Upadhyay et al., 10 Mar 2026). The same overview reports high run-level correlation between automated and manual assessments, but also more variance at the per-narrative level and only moderate agreement for automated relevance judgment (Upadhyay et al., 10 Mar 2026). This makes evaluation itself part of the multi-task problem.

Multimodal evaluation varies by application. M$^2$2RAG evaluates image captioning and multi-modal QA with BLEU-4, ROUGE-L, and CIDEr, fact verification with accuracy and F1, and image reranking with FID (Liu et al., 24 Feb 2025). Multi-RAG, a system for adaptive video understanding, is benchmarked on MMBench-Video with GPT-4-based semantic scoring from 0 to 3 and reports separate reasoning and perception means (Mao et al., 29 May 2025). The multimodal design-space study mRAG evaluates retrieval with Recall@$^2$3 and MRR, and generation with ROUGE-L plus semantic correctness judged by InternVL3-14B and GPT-4.1 (Hu et al., 29 May 2025).

A common controversy concerns how much trust to place in automatic evaluators. The TREC overview states that automation is reliable at the run level but remains challenging at fine granularity, and that LLM-based relevance judgment is not yet robust enough for high-stakes use (Upadhyay et al., 10 Mar 2026). This suggests that multi-task RAG evaluation is itself still an open systems problem rather than a settled measurement protocol.

6. Empirical results, recurring findings, and open directions

The empirical literature reports gains from multiple design families rather than one dominant paradigm. H-RAG achieved an nDCG@5 score of 0.4271 on Task A and a harmonic mean score of 0.3241 on Task C, with RB$^2$4, RL$^2$5, and RB$^2$6, and the paper emphasizes that fluent answer scores outpace grounding and faithfulness in multi-turn RAG (Elchafei et al., 1 May 2026). DuetRAG reports 36.3 accuracy on HotPotQA under GPT-4 evaluation, compared with 32.7 for its external model and 23.1 for its internal model, supporting the claim that arbitration across internal and external knowledge improves multi-hop answering (Jiao et al., 2024). ImpRAG reports an average exact match of 54.4 across eight knowledge-intensive tasks, versus 49.2 for RA-IT, 49.9 for RA-DIT, and 50.9 for RA-DIT-Llama, along with an average retrieval recall of 67.9 (Zhang et al., 2 Jun 2025). SEMA-RAG reports improvement of the strongest baseline by +6.46 accuracy points on average, measured per backbone, across five medical benchmarks and five LLM backbones (Huang et al., 16 May 2026).

The same literature also documents strong negative results that shape current best practice. In multimodal retrieval-augmented generation for LVLMs, answer accuracy does not necessarily increase with more context, and the best practice reported in that study is to provide only the most relevant re-ranked document to the generator (Hu et al., 29 May 2025). In TREC, attribution completeness and high-quality automated relevance judgment remain difficult (Upadhyay et al., 10 Mar 2026). In MARAGS, training on original targets increased hallucination while relabeling unanswerable instances to “i don’t know” drastically reduced hallucination, improving the CRAG score despite lower raw accuracy (DeHaven, 2024).

Several forward directions are repeatedly identified. The TREC overview calls for improved sub-narrative decomposition, meta-evaluation of LLM evaluators, better attribution and factuality scoring, more agentic systems, and further work on multi-turn, interactive, and dynamic narratives (Upadhyay et al., 10 Mar 2026). Multimodal papers point toward stronger modality alignment, better use of retrieved images rather than primarily text, and unified frameworks that can dynamically select or suppress evidence through self-reflection (Liu et al., 24 Feb 2025, Hu et al., 29 May 2025). Graph-based systems point toward multi-modal extensions, structured knowledge integration, and more precise multi-hop expansion without expensive LLM-based graph construction (Zhuang et al., 11 Oct 2025, Bonifazi et al., 29 Jun 2026).

A plausible synthesis is that multi-task RAG is becoming less a single technique than a family of coordinated design decisions about decomposition, routing, retrieval depth, evidence representation, and verification. The surveyed systems differ sharply in implementation, but they converge on one proposition: robust retrieval-augmented generation requires explicit management of task heterogeneity rather than a one-pass retrieve-then-read pipeline.