Generative Information Retrieval

• Generative Information Retrieval (GenIR) is an emerging paradigm that uses generative models to produce document identifiers or grounded answers instead of traditional ranking.
• It encompasses two main directions—generative document retrieval and grounded answer generation—merging retrieval with end-to-end answer synthesis using encoder–decoder architectures.
• Challenges in evaluation, training misalignments, and dynamic corpus adaptation drive novel methods and interactive strategies to improve robustness and effectiveness.

Generative Information Retrieval (GenIR) is an information retrieval paradigm in which a generative model satisfies an information need by generating either document identifiers or user-facing responses, rather than only ranking items from an explicit index. The literature usually divides GenIR into two major directions: Generative Document Retrieval (GDR/GR), where the model generates document identifiers, and Grounded Answer Generation (GAG/GAR) or Reliable Response Generation, where the model generates an answer tied to evidence (Bénédict et al., 2023, Li et al., 2024, Kuo et al., 2024). Relative to classical index-retrieve-then-rank systems, GenIR shifts IR from explicit matching toward conditional generation, and this shift propagates into document identifier design, training objectives, evaluation, robustness, and interaction.

1. Definitions, taxonomy, and conceptual scope

A concise formalization of GDR treats retrieval as conditional sequence generation. In the survey formulation, for a query $q$, a generative retriever models

$$\mathcal{R}(q,d) = P(d' \mid q; \theta) = \prod_{i=1}^{T} P(d'_i \mid d'_{<i}, q; \theta),$$

where $d'$ is a generated document identifier sequence (Li et al., 2024). The workshop formulation expresses the same idea as a conditional distribution over document identifiers, $p(\text{doc-id}\mid q)$, and contrasts it with answer generation, $p(\text{answer}\mid q,\text{evidence})$ (Bénédict et al., 2023).

Direction	Output	Typical role
GDR / GR	Generated document identifier	Retrieval proper
GAG / GAR	Generated answer grounded in evidence	Answer synthesis
Reliable Response Generation	Direct user-facing response	End-to-end information access

This taxonomy matters because GenIR is not reducible to DocID generation alone. The SIGIR Gen-IR workshop explicitly framed the field around both GDR and GAG, and the surveys similarly distinguish GR from reliable response generation (Bénédict et al., 2023, Li et al., 2024). A recurrent misconception is therefore that GenIR is synonymous with “generate a DocID”; the literature instead treats it as a broader family of generation-based information access systems.

The same sources also identify why GenIR has been discussed as a possible paradigm change. End-to-end generative models can compress multiple pipeline stages, potentially avoid a traditional document index, and unify retrieval with answer synthesis (Bénédict et al., 2023). At the same time, this reformulation introduces the “infinite index paradigm”: a generator can produce arbitrarily many outputs, so standard offline assumptions about a finite candidate space and fixed relevance sets become unstable (Bénédict et al., 2023, Arabzadeh et al., 2024). This is one reason GenIR evaluation rapidly became a central research topic.

A bibliographic caveat is also documented in the supplied literature. The record titled “From Relevance to Authority: Authority-aware Generative Retrieval in Web Search Engines” (Lee et al., 15 Apr 2026) does not describe a GenIR method; it is identified in the supplied analysis as an ACL formatting guide. It therefore does not provide evidence about authority-aware generative retrieval.

2. Generative document retrieval architectures and document identifiers

Most GDR systems use encoder-decoder transformers that generate DocIDs autoregressively. The surveys describe the dominant recipe as fine-tuning T5- or BART-like models so that, given a query, the model decodes a valid identifier under constrained decoding or constrained beam search (Kuo et al., 2024, Li et al., 2024). Inference is typically restricted by a prefix trie or analogous structure so that only valid identifier continuations can be produced. This turns ranking into a sequence-generation problem, but it also makes identifier construction a first-order design decision.

The identifier taxonomy is broad. The survey distinguishes numerical identifiers and string identifiers, with numerical DocIDs further divided into single-token atomic IDs and sequential IDs, and string identifiers including titles, URLs, substrings, pseudo-queries, and multi-view identifiers (Kuo et al., 2024). Another survey similarly contrasts numeric DocIDs, textual DocIDs, semantically structured IDs, and learnable identifiers, emphasizing that identifier choice shapes scalability, interpretability, and how much the model can exploit linguistic pretraining (Li et al., 2024). Across both surveys, semantically meaningful identifiers often perform better than arbitrary ones, but the empirical picture is not uniform: atomic IDs can remain competitive under some parameter budgets and datasets.

Large-scale effectiveness depends heavily on these design choices. RIPOR argued that earlier generative retrievers underperformed on realistic collections because DocIDs were not constructed from relevance structure and training ignored beam-search prefix survival (Zeng et al., 2023). It therefore introduced relevance-based document ID construction with residual quantization and prefix-oriented ranking optimization. On MS MARCO Dev, RIPOR reached MRR@10 = 0.333 and Recall@10 = 0.562, reported as a 30.5% improvement in MRR@10 over the strongest prior generative baseline; it also improved NDCG@10 on TREC DL 2019 and 2020 (Zeng et al., 2023). This result is important because it directly addresses the longstanding criticism that GenIR only works on artificially small collections.

A later zero-shot line extends these ideas to heterogeneous and instruction-driven settings. ZeroGR treats retrieval as generation of semantically meaningful DocIDs conditioned on natural-language instructions, adds an instruction-tuned pseudo-query generator, and uses reverse annealing decoding to balance precision and recall (Sun et al., 12 Oct 2025). In the reported results, ZeroGR-3B achieved 41.1 Acc@1 on MAIR overall and 48.1 nDCG@10 average on BEIR (Sun et al., 12 Oct 2025). This suggests that GenIR can be pushed beyond supervised single-corpus settings into instruction-conditioned zero-shot retrieval, though the method still assumes access to task instructions and large-scale synthetic data generation.

3. Grounded answer generation, RAG, and unified models

The second major branch of GenIR replaces document output with generated responses. The central distinction, emphasized in both the workshop and the “Search Still Matters” perspective, is that a generative system may provide an answer while a traditional IR system returns documents that can be inspected, compared, and verified (Bénédict et al., 2023, Hersh, 2023). This distinction is operationally important in academic, biomedical, and other high-stakes settings, where users care about authoritativeness, timeliness, contextualization, citations, and traceability (Hersh, 2023).

The literature therefore treats retrieval-augmented generation (RAG) as a key intermediary architecture. In the perspective paper, RAG is described as using search engine output to prompt existing LLMs so that they produce more focused and up-to-date information (Hersh, 2023). The same paper is explicit that generative AI may augment the IR process but does not replace search for serious information needs; document-level access, source verification, and current literature coverage remain indispensable (Hersh, 2023). This is a recurring corrective to stronger replacement narratives around GenIR.

A more integrated approach is exemplified by UniGen, which argues that GDR and grounded QA should be learned jointly rather than as separate retriever-reader pipelines (Li et al., 2023). UniGen uses a shared encoder, a retrieval decoder, and a QA decoder, with a joint objective

$$\mathcal{L} = \lambda \mathcal{L}_{\text{retr}'} + (1-\lambda)\mathcal{L}_{\text{qa}'}.$$

It further introduces LLM-generated Q-Connectors and D-Connectors to bridge the query-to-docid and docid-to-answer gaps, and an iterative enhancement strategy in which retrieved documents and generated answers are used to improve later rounds (Li et al., 2023). Empirically, UniGen-Base improved MRR@10 by 1.81% on MS MARCO and 0.83% on NQ over the best retrieval baselines, while UniGen-Iter improved R@1 by 11.76% on MS MARCO and 2.17% on NQ; on QA, it beat DPR+FID on MS MARCO by 28.94% in BLEU-1 and surpassed Ultron+FID on NQ by 3.73% in EM (Li et al., 2023).

These results support a broader interpretation of GenIR as a unified learning problem spanning retrieval and grounded answer generation. They also clarify why answer-generation systems cannot be evaluated solely by classical ranking metrics: the system output is not necessarily an item from a fixed collection, and end-to-end utility depends on grounding, attribution, and factual control rather than on document ranking alone (Arabzadeh et al., 2024, Alaofi et al., 2024).

4. Training objectives, ranking alignment, and internal retrieval mechanisms

A major technical issue in GDR is the mismatch between next-token DocID prediction and document-level relevance. The DDRO paper states this explicitly as token-level misalignment: maximizing next-token likelihood does not ensure correct document ordering by relevance (Mekonnen et al., 7 Apr 2025). DDRO addresses this with direct document relevance optimization, a pairwise learning-to-rank objective layered on top of supervised fine-tuning, avoiding reward modeling and RL fine-tuning. On benchmark datasets, DDRO reported a 7.4% improvement in MRR@10 for MS MARCO and a 19.9% improvement for Natural Questions relative to reinforcement-learning-based approaches (Mekonnen et al., 7 Apr 2025). The significance of this result is methodological as much as empirical: it shows that GenIR ranking can be optimized directly with preference-style objectives rather than by treating ranking alignment as a separate RL problem.

A second training extension concerns multi-graded relevance. GR$^2$ argues that much prior generative retrieval work is effectively binary and ignores the ordered structure of relevance labels (Tang et al., 2024). It therefore combines docid generation and autoencoding to obtain identifiers that are both relevant and distinct, then trains the query-to-docid retriever with multi-graded constrained contrastive (MGCC) learning. On Gov 500K, GR$^ {2S}$ and GR$^ {2P}$ improved over RIPOR by about 14% and 11% relative on P@20, respectively, while also remaining strong on binary datasets (Tang et al., 2024). This places graded relevance modeling inside the generative paradigm rather than treating it as a post hoc evaluation issue.

Mechanistic work has also begun to analyze how GenIR models compute relevance internally. Reverse-engineering experiments on encoder-decoder GenIR models found that the decoder is the primary component responsible for successful retrieval, not the encoder, and that decoder computation can be divided into three stages: priming, bridging, and interaction (Reusch et al., 25 Mar 2025). In that account, cross-attention is most active in the bridging stage, while query-document interaction occurs only in the final interaction stage, where MLPs dominate the final ranking decision (Reusch et al., 25 Mar 2025). This finding is conceptually important because it replaces a black-box view of GenIR with a more structured account of where retrieval knowledge is stored and where ranking is actually computed.

Together, these papers indicate that GenIR training is not well described by “seq2seq likelihood” alone. The literature has moved toward pairwise ranking, graded constraints, curriculum over prefixes, and mechanistic analyses that isolate which parts of the network carry retrieval-specific computation (Zeng et al., 2023, Mekonnen et al., 7 Apr 2025, Tang et al., 2024, Reusch et al., 25 Mar 2025).

5. Evaluation, trustworthiness, and robustness

Evaluation is arguably the most distinctive methodological problem in GenIR. When a system returns novel text rather than a ranked list from a fixed corpus, classical Cranfield assumptions break down (Arabzadeh et al., 2024, Alaofi et al., 2024). A comparison study therefore examined five offline methods for evaluating generative IR responses: binary relevance, graded relevance, subtopic relevance, pairwise preferences, and embedding-based evaluation (Arabzadeh et al., 2024). On TREC DL 2019 and 2020 validation against human judgments, pairwise preferences showed the best alignment with human ordering, while subtopic relevance was presented as a practical compromise because it is autonomous, interpretable, and auditable (Arabzadeh et al., 2024).

The broader evaluation chapter treats this transition as a two-part development: LLMs are increasingly used as evaluation infrastructure, and GenIR systems themselves require end-to-end evaluation beyond document ranking (Alaofi et al., 2024). It proposes two conceptual safeguards against circularity. First, LLM-based assessment can be understood as “slow search”, where a slower system is used to evaluate or train a faster production system. Second, evaluation still needs human grounding, because relevance, usefulness, and trust remain human constructs (Alaofi et al., 2024).

A stricter answer to circularity is proposed in work on Normative Generative Ranking, which formalizes candidate selection as a within-query ranking problem and introduces strict estimator separation: Judge B provides supervision and Judge A provides evaluation, with no overlap in training, tuning, or model selection (Nahhas et al., 24 Feb 2026). On NGR-33k, the distilled BGE-M3 student achieved nDCG@5 = 0.771 \pm 0.02 and outperformed the Judge-B-supervised Cross-Encoder, which achieved nDCG@5 = 0.577, under leakage-free Judge-A evaluation (Nahhas et al., 24 Feb 2026). Although the task is normative ranking rather than standard web retrieval, the methodological claim generalizes: evaluator separation is necessary if GenIR performance claims are to avoid preference leakage.

Robustness remains a major weakness. A dedicated study on out-of-distribution generalization concluded that the OOD robustness of generative IR models is in need of improvement, examining query variations, unseen query types, unseen tasks, and corpus expansion (Liu et al., 2024). In its KILT experiments, generative retrievers showed strong IID performance but substantial degradation under misspellings, word-order changes, cross-dataset query shifts, and cross-task transfer, although corpus expansion was somewhat more favorable (Liu et al., 2024). This complicates claims that GenIR is a simple replacement for dense or sparse retrieval: strong in-distribution results do not imply reliable behavior under realistic shifts.

6. Dynamic corpora, interaction, and research directions

Dynamic corpora expose another structural tension in GenIR. Because many GDR systems store corpus knowledge parametrically, updating the corpus can induce catastrophic forgetting. Early work on dynamic corpora found that generative retrieval on StreamingQA was more adaptable to evolving knowledge, more robust to temporal information, and more efficient than dual encoders in terms of indexing time and storage footprint, although online decoding latency remained a bottleneck (Kim et al., 2023). The same paper also showed that dense retrievers could exploit timestamp shortcuts, whereas generative retrieval was more robust to such temporal artifacts (Kim et al., 2023).

More recent continual-learning work sharpened this diagnosis. The continual retrieval paper argues that GenIR exhibits a pronounced stability-plasticity trade-off under sequential corpus increments and proposes post-adaptation memory tuning (PAMT) with a Parametric Memory Head (PMH) to stabilize retrieval without retraining the backbone (Mekonnen et al., 25 Apr 2026). Under expanded-trie evaluation, PAMT improved backward transfer for Full FT–SPQ on MS MARCO from $-55.72$ to $$\mathcal{R}(q,d) = P(d' \mid q; \theta) = \prod_{i=1}^{T} P(d'_i \mid d'_{<i}, q; \theta),$$0 and for Full FT–TU from $$\mathcal{R}(q,d) = P(d' \mid q; \theta) = \prod_{i=1}^{T} P(d'_i \mid d'_{<i}, q; \theta),$$1 to $$\mathcal{R}(q,d) = P(d' \mid q; \theta) = \prod_{i=1}^{T} P(d'_i \mid d'_{<i}, q; \theta),$$2, while updating only a sparse subset of memory values per session (Mekonnen et al., 25 Apr 2026). This suggests that modular memory mechanisms may be necessary if model-as-index retrieval is to remain viable on living collections.

Interaction research extends GenIR from retrieval algorithms to user-facing information access systems. The interaction chapter argues that a major benefit of generative IR is the move from predefined actions such as clicks and requeries toward richer natural-language and multimodal interaction (Aliannejadi et al., 2024). It organizes this space around expressing vague or formalized information needs, explicit and implicit feedback, clarification and preference elicitation, proactive system behavior, explanation, and multimodal interfaces (Aliannejadi et al., 2024). A plausible implication is that GenIR cannot be evaluated solely as a ranking model once it becomes conversational, mixed-initiative, and explanation-bearing.

Query reformulation is one concrete example of this broader generative turn. GenCRF treats reformulation itself as a GenIR-style task: an LLM generates multiple reformulations, clusters them into 1 to 3 intent groups, aggregates them with Similarity Dynamic Weights or Score Dynamic Weights, and uses a Query Evaluation Rewarding Model (QERM) to trigger regeneration when needed (Seo et al., 2024). On six BEIR datasets, the reported best averages were 0.4176 nDCG@10 for GenCRF/ScoreDW-FT-QREM with Mistral and 0.4270 with Llama (Seo et al., 2024). This line connects GenIR to intent diversification, feedback loops, and retrieval-time query synthesis rather than to DocID generation alone.

The forward-looking workshop report on the future of IR in the age of generative AI generalizes these developments into a broader IR-GenAI agenda: end-to-end evaluation of retrieval plus generation, instruction-following retrieval, learning when to retrieve versus generate, user modeling and personalization, agentic multi-step information access, and efficiency-aware system design (Allan et al., 2024). The cumulative picture is therefore not that GenIR simply supersedes classical IR. Rather, the literature presents a layered future in which generative retrieval, grounded answer generation, human-auditable evaluation, dynamic adaptation, and richer interaction coexist with a continuing need for authoritative sources, source inspection, and robust search infrastructure (Hersh, 2023, Allan et al., 2024).