acceptodds
Under review as a conference paper at ICLR 2027

Beyond the Query: Corpus-conditioned Predictive Retrieval for Underspecified Queries

Abstract

Users search because they may lack the knowledge needed to describe the information they seek. Query expansion can help, but may emphasize only some concepts or relations relevant to an underspecified query. We introduce CPR (Corpus-conditioned Predictive Retrieval), which uses corpus documents as partial representations of possible intents. Motivated by the posterior-predictive principle, CPR averages corpus-based relevance predictions under a support prior constructed from semantic and lexical evidence to refine candidate support, without additional training. Experiments on reasoning-intensive retrieval show that better prior rankings need not yield better document predictions. Composition with query expansion recovers shared failures, while weaker direct retrievers can still provide useful document relations. These findings support evaluating evidence through the predictions it enables and using corpus evidence to complement generated queries and stronger embeddings.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Retrieval

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_qdmFOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies reasoning-intensive retrieval for underspecified queries whose text alone does not determine relevance. It proposes CPR, which requires no additional training and treats corpus documents as partial candidates for the latent intent, averaging corpus-based relevance predictions under a prior built from semantic and semantic-lexical evidence. Scoring combines semantic neighbor relations with lexical coverage relative to a corpus background and weights predictions by candidate support. Experiments on BRIGHT and R2MED with three frozen retrievers show improvements over direct retrieval, leading results among the compared LLM-free methods, and complementarity with query expansion methods such as HyDE and ThinkQE.

Critical strengths

1. Improvements across the full evaluation suite CPR improves the task-average nDCG@10 of direct retrieval on all three frozen retrievers (pooled mean 20.30 to 27.04) and leads every compared LLM-free method across 11 BRIGHT and 8 R2MED tasks, with corresponding recall gains. This matters because these results, spanning 19 tasks and three retrievers, form the paper's central empirical case.

2. Demonstrated complementarity with query expansion Composing CPR with HyDE or ThinkQE improves over both standalone methods and recovers 7.20% and 6.89% of Base misses that neither recovers, and the HyDE composition improves on both methods on every backbone. The paper also separates recovery of missed queries from ranking gains on retrieved queries. This matters because it identifies a practical complementarity between corpus evidence and generated query evidence, which is the paper's most transferable finding.

Critical weaknesses

1. Limited novelty relative to closely related retrieval methods CPR combines components that are individually close to established retrieval techniques, including document-side evidence estimation, rank-based neighbor weighting, and multiplicative semantic and lexical coverage. The paper compares with recent corpus-guided and probabilistic methods such as LexBoost, ECLIPSE-PRF, RDIME-PRF, SemRank, Parry et al., PQR, and QuDAR. All of these comparisons remain at the full-pipeline level, and no ablation isolates what corpus-conditioned prediction adds beyond the neighbor and lexical signals already used by these methods. The paper also lacks a matched-budget control that fuses the same signals without the intent mixture, so the source of the reported gains remains difficult to identify. For these reasons, the boundary between CPR and prior work is unclear.

2. The core empirical claim remains unsupported The headline comparison against dense retrieval and the baselines reports no significance tests or confidence intervals, and the intervals that do appear cover only component and diagnostic analyses. The gains also decline as the backbone strengthens (E5 +8.76, Qwen3 +6.64, ReasonIR +4.80), HyDE surpasses CPR on ReasonIR nDCG (31.33 vs 28.96), and task-level outcomes include losses (TheoremQA-Theorems -2.71, MedXpertQA-Exam -1.25, TheoremQA-Questions -0.40) alongside large gains. The abstract presents a negative observation about prior rankings as a central finding, and the paper reports no latency, memory, or aggregation cost even though it acknowledges that corpus-wide relations are expensive. The main empirical claim of a robust and practical advantage over direct retrieval is therefore not established.

3. Unverified design choices and inconclusive diagnostics The method depends on many design choices that receive no ablation or sensitivity analysis, including the setting k = ⌈√N⌉, the CSLS-style density correction, reciprocal ranks, the rank product, equal weighting of the two priors, alternating folds, top-h truncation, IDF weighting, and the forced retention of the top-1 document. No development set or tuning protocol is reported, so benchmark-specific tuning cannot be ruled out. The diagnostic analyses are circular because ΔL and the coverage-gain correlations are computed post hoc with relevance labels from the same tasks and with weights derived from the quantity under evaluation. Formal issues also remain. The normalization of u(d) over C is combined with sums over Aq in Eq. 2 and Eq. 10, ρs denotes two different rank functions, Eq. 10 is justified only by intuition, and an uncalibrated ratio is treated as a valid binary likelihood on the strength of a range argument. These problems prevent the reported gains from being attributed to the proposed design and raise serious concerns about reproducibility.

Questions for authors

What is the minimal essential difference between CPR and the corpus-guided methods it already compares with, and how much does CPR add over them at matched budget? Do the headline gains survive significance testing and remain robust beyond E5, particularly on ReasonIR where HyDE leads? What latency, memory, and aggregation cost does CPR add relative to Base and HyDE, and were the hyperparameters selected on a held-out development set? Please align the posterior-predictive framework with the actual estimators and provide a derivation. Report query-weighted results, additional backbones, and measured efficiency. Replace the label-dependent diagnostics with label-free validation or state their interpretation limits explicitly.

3 comments

We sincerely thank the reviewer for their thoughtful and constructive feedback. We greatly appreciate the time and effort devoted to reviewing our work. We have revised the manuscript to clarify the formulation and provide additional experimental evidence addressing the concerns. We address each point below.

W1

Contribution. Underspecified queries may not express the information needed to judge document relevance. Rather than simply augmenting the query or ranking corpus evidence, CPR uses eligible corpus documents as partial evidence sources and combines their source-conditioned relevance predictions. A document that matches the query closely need not be the most informative source for predicting other documents' relevance. This distinction between direct ranking quality and predictive contribution is the central contribution of CPR.

Controlled comparisons. Across 19 tasks and three backbones, applying corpus-conditioned prediction improves prior-only ranking by +1.273 nDCG@10 (95% CI [0.921, 1.638]). Holding the candidate prior, semantic and lexical evidence, source scope, and candidate set fixed, CPR outperforms same-signal RRF60 by +1.437 [1.208, 1.660]. Under the same resource configuration (one RTX 6000 Ada and four CPU cores) and a retrieval-time cap of 2T, the gain over RRF60 remains +1.436 [1.207, 1.663] across all 19 tasks; CPR also exceeds ECLIPSE-PRF by +11.522 [10.688, 12.360] within that cap. Here, T is the measured full-CPR retrieval time for the corresponding task and backbone, and configuration selection does not use relevance labels.

The RRF control isolates the final combination of fixed signals; it does not, by itself, remove source aggregation. Replacing summed source contributions with the strongest contribution per candidate yields a smaller, statistically inconclusive difference (+0.124 [−0.018, 0.266]). We therefore attribute the controlled gains to CPR's predictive scoring and signal combination, without claiming that every aspect of the mixture aggregation is individually indispensable.

The contribution of predictive scoring is discussed in Section 5.3 and Appendix E.3, with same-signal comparisons reported in Appendix E.4. We will further include the matched computational budget comparisons in Appendix H.

W2

Significance and backbone robustness. We reanalyzed 19 tasks and 2,157 queries with 10,000 paired within-task bootstrap resamples. CPR improves over Base, and CPR+HyDE improves over HyDE, with 95% confidence intervals excluding zero for all three evaluated backbones.

BackboneCPR − Base [95% CI]CPR+HyDE − HyDE [95% CI]
E5+8.762 [7.821, 9.715]+8.546 [7.585, 9.557]
Qwen3+6.645 [5.722, 7.567]+7.172 [6.289, 8.058]
ReasonIR+4.802 [3.845, 5.758]+3.068 [2.135, 4.014]
Overall+6.736 [6.058, 7.426]+6.262 [5.591, 6.954]

CPR also exceeds ECLIPSE-PRF by +5.947 [5.255, 6.645] and RDIME-PRF by +6.246 [5.559, 6.938]. The gains persist with query-weighted aggregation (Base 20.370 → CPR 26.909; HyDE 26.581 → CPR+HyDE 32.423). On an additional BGE-M3 backbone evaluated over all 19 tasks, CPR − Base is +6.050 [5.279, 6.820] and CPR+HyDE − HyDE is +6.432 [5.622, 7.265].

Limits of the improvements. The gain over Base decreases with stronger backbones. Standalone HyDE exceeds CPR on ReasonIR (31.335 vs. 28.960) and BGE-M3 (22.518 vs. 19.340), and CPR loses to Base on TheoremQA-Theorems (−2.71), MedXpertQA-Exam (−1.25), and TheoremQA-Questions (−0.40) in task-macro nDCG@10. Nonetheless, CPR+HyDE improves over HyDE on both ReasonIR and BGE-M3. We will state these task- and backbone-dependent limitations explicitly instead of claiming uniform superiority.

Meaning of the prior-ranking observation. The abstract's observation distinguishes a prior's direct ranking performance from its predictive usefulness: Semantic–Lexical ranks better initially, whereas Semantic achieves better final ranking after prediction. We will clarify this interpretation in the abstract.

Online cost. We measured generation, query-side encoding, and retrieval as applicable across the 19 tasks. The following values are batch-amortized seconds per query, not isolated single-query latencies.

MethodE5Qwen3ReasonIR
Base0.0150.0410.046
ECLIPSE-PRF0.0160.0430.048
CPR0.3160.3800.409
HyDE7.9858.1788.433
ThinkQE50.33050.56450.839
CPR+HyDE8.0688.2608.542

CPR is more costly than LLM-free retrieval baselines but avoids query-time LLM generation. In the measured setup, adding CPR to HyDE costs approximately 0.08–0.11 additional seconds/query, compared with a +6.262 nDCG@10 gain. Recorded failed ThinkQE generation attempts are excluded. CPR retrieval was timed using four AMD EPYC 7313 CPU cores (FP64); generation, query encoding, and other methods' final retrieval used RTX 6000 Ada GPUs. Model loading and corpus preprocessing are excluded.

Million-document scalability. On one million unique documents with E5 and 2,157 cached queries, search time is 4.826 ms/query for Base, 841.716 ms/query for CPR, and 1,069.006 ms/query for CPR+HyDE. CPR requires 25.096 GiB peak GPU allocation, 28.688 GiB peak CPU RSS, and 24.052 GiB of relation-index storage; offline index construction takes 281.897 seconds. These cached-input measurements exclude generation and encoding, and retrieval accuracy on the combined million-document corpus was not evaluated. They quantify the added storage and aggregation costs rather than establishing a cost advantage over direct retrieval.

The statistical significance results are reported in Section 4 and Appendix D, while the computational analysis is presented in Section 8 and Appendix H. We will additionally report the BGE-M3 and query-weighted results in Appendix D.

W3

We first clarify the definition of Aq\mathcal{A}_q in Section 2. Our intention was to allow Aq\mathcal{A}_q to be a configurable subset of the corpus, ranging from a restricted set of retrieved documents to the full eligible corpus. The original description did not clearly distinguish these cases, and we have revised it accordingly. Importantly, CC and Aq\mathcal{A}_q play different roles: CC provides the corpus-wide background u(d)u(d), whereas Aq\mathcal{A}_q determines which sources contribute to the relevance prediction. We make this distinction explicit in the derivation of Equations (2) and (10).

Design sensitivity. We varied one design choice at a time across 19 tasks, 2,157 queries, and E5/Qwen3/ReasonIR while retaining the original evaluation inputs. Below, differences are alternative configuration minus the manuscript CPR baseline in task-macro nDCG@10, with paired 95% bootstrap intervals.

Changed design choiceΔnDCG@10 [95% CI]
Prior mixture weight α=0.25 instead of 0.5−1.378 [−1.788, −0.968]
Prior mixture weight α=0.75 instead of 0.5−0.247 [−0.593, 0.109]
Density neighborhood k=0.5√N+0.147 [−0.028, 0.322]
Density neighborhood k=2√N−0.302 [−0.463, −0.145]
Remove CSLS-style density correction−1.533 [−1.891, −1.181]
Prior reciprocal-rank exponent β=0.5−1.316 [−1.566, −1.067]
Replace semantic–lexical rank product with RRF60−0.409 [−0.669, −0.147]
Top-h at 25%, with head-only renormalization+0.208 [0.110, 0.308]
Replace query-term IDF with uniform weighting+0.004 [−0.069, 0.078]
Remove forced top-1 retention−1.002 [−1.373, −0.639]

Random 2- and 4-fold alternatives yield changes between −0.090 and +0.077 nDCG@10 across the tested seeds, with intervals including zero. The default equal-weight prior mixture (α=0.5) attains 27.035 nDCG@10, compared with 25.658 at α=0.25 and 26.789 at α=0.75. Removing density correction and top-1 retention decreases performance, whereas the top-h experiment identifies a modest improvement over the default. These tests support several design choices without implying that every default is optimal.

Diagnostic interpretation. The ΔL measure uses benchmark relevance labels to compare query-term coverage in relevant documents and semantic competitors, but does not use CPR's predicted gains to define ΔL. In the source-supported terminology analysis, documents used to construct the relevant/competitor term sets are excluded from the sources being examined. Both analyses are post-hoc, label-dependent associations, not independent performance validation or causal evidence. The main comparisons and controlled ablations, reported above, provide the performance evidence.

The prior-mixture sensitivity analysis is presented in Section 5.1 and Appendix E.1, while the scoring controls and additional sensitivity analyses are provided in Appendix E.4. We clarify the role of Aq\mathcal{A}_q in Section 2 and discuss the interpretation limits of the label-dependent diagnostics in Section 6.1 and Appendix F.4.

Dear Reviewer qdmF,

Thank you again for your thoughtful feedback. We would greatly appreciate it if you could review our responses and let us know whether they have addressed your concerns or if any questions remain. Thank you for your time and consideration.

Best regards, The Authors

Reviewer_qHuYOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes CPR, which addresses the challenges of retrieving relevant documents for underspecified queries. CPR computes a support prior and relevance likelihood between the corpus and query, combining them to produce a final score for each document and perform retrieval. Through experiments on BRIGHT and R2MED, this paper demonstrate that CPR consistently outperforms existing methods across various dense retrievers and corpus types.

Critical strengths
  • Addressing an important problem with a conceptually clear approach: The problem of underspecified queries is an important challenge in modern IR, and the proposed method addresses this problem by leveraging concepts that have played important roles in the development of IR, making the overall approach intuitive and easy to understand.
  • Extensive experiments and baselines: This work demonstrates the effectiveness of the proposed method across diverse corpus types using BRIGHT and R2MED, covering various domains and formats (e.g., natural language, mathematics, etc.). Furthermore, this paper provides rigorous comparisons against PRF-based methods, which share similar characteristics with the proposed approach, as well as LLM-based methods.
Critical weaknesses
  • Insufficient evidence supporting the claim that LLMs tend to emphasize only certain concepts: One of the main limitations highlighted in the Introduction is that LLM-generated expansions may emphasize certain concepts. While this claim is plausible, this paper does not provide experimental evidence or analysis to support it. As a result, it is difficult to fully understand why CPR should be preferred over LLM-based expansion methods.

    • More specifically, unless this paper demonstrates that LLMs inherently tend to overemphasize certain concepts, I may question whether improving LLM-based expansion methods would be a more practical and efficient direction, especially when considering the computational cost and latency concerns discussed below.
  • Per-query computational cost and latency that scale with corpus size: When performing retrieval for a given query, CPR requires additional embedding-based and term-based (TF, IDF) computations over individual documents, beyond simply computing relevance scores between the query and documents. This raises concerns about increasing computational costs and latency as the corpus grows, particularly regarding whether CPR can be efficiently applied to realistic retrieval settings where corpus sizes often exceed one million documents.

    • If the additional computational cost and latency in such settings turn out to be greater than those of LLM-based expansion methods, this raises the question of whether further improving LLM-based expansion methods might be a more practical alternative.
Questions for authors

Could you please address the following questions? I would be willing to increase my rating if the responses sufficiently address the weaknesses discussed above.

  1. Could you provide additional results for BM25 with existing query expansion methods? In reasoning-intensive retrieval tasks, query expansion methods have often shown larger performance improvements with sparse retrievers than with dense retrievers. Therefore, I would like to see whether the proposed method still achieves superior performance compared to these settings.
  2. Could you provide a comparison of the per-query computational cost and latency of CPR against other methods as the corpus size increases? Such an analysis would help clarify the practical scalability of CPR, particularly in large-scale retrieval settings.
  3. Could you provide a simple analysis or experiment demonstrating that LLM-generated expansions tend to overemphasize certain concepts? I believe that adding such evidence would more effectively motivate the necessity of CPR and strengthen the paper's argument for its advantages over LLM-based expansion methods!
8 comments

We sincerely thank the reviewer for their thoughtful and constructive feedback. We greatly appreciate the time and effort devoted to reviewing our work. We have revised the manuscript to clarify the formulation and provide additional experimental evidence addressing the concerns. We address each point below.

W1 & Q3

Thank you for valuable comment. We recognize that our original motivation was not sufficiently clear and that additional empirical evidence was needed. Our concern is not that LLMs inherently favor particular concepts. Rather, an underspecified query may correspond to numerous latent intents. Although query expansion could explore these possibilities through sufficiently many generations, doing so would incur substantial generation and encoding costs. Under practical computational budgets, only a limited number of expansions can be generated, potentially leaving many plausible intents unexplored.

We also appreciate the reviewer's observation that generated expansions may concentrate on similar concepts. To investigate this, we analyzed HyDE and ThinkQE expansions and CPR's eight highest-prior corpus supports across all queries, using the same sentence encoder to measure cosine similarity.

EvidenceOriginal query vs. evidenceWithin-evidence similarityAfter deduplication
HyDE0.73760.95640.9413
ThinkQE0.69400.92150.9168
CPR (E5)0.44380.45750.4557
CPR (Qwen3)0.46050.51170.5097
CPR (ReasonIR)0.45670.50120.4995

Although generated expansions differ from the original query, they exhibit high similarity to one another, even after removing exact duplicates. In contrast, CPR's highest-prior corpus supports are substantially less similar to one another, despite being selected without any diversity objective.

These findings demonstrate that generating multiple expansions does not necessarily provide diverse semantic evidence. CPR instead leverages existing corpus documents as complementary sources of evidence without additional LLM generation. Moreover, applying CPR to the same HyDE-generated evidence improves nDCG@10 by 6.262 points, further demonstrating its complementary value. We have clarified this motivation in Introduction and added the semantic similarity analysis in Section 6.3, with experimental details and results provided in Appendix F.5.

W2 & Q2

We measured the online components of Base, ECLIPSE-PRF, CPR, HyDE, ThinkQE. Costs include query-side encoding, LLM generation and feedback where applicable, and final retrieval.

We also clarify that Aq\mathcal{A}_q is not required to contain the entire corpus. CPR allows the support set to be restricted to a subset of corpus documents relevant to the query, while using all eligible documents is one possible configuration. We have revised Section 2 to make this flexibility explicit.

MethodE5Qwen3ReasonIR
Base0.0150.0410.046
ECLIPSE-PRF0.0160.0430.048
CPR0.3160.3800.409
HyDE7.9858.1788.433
ThinkQE50.33050.56450.839
CPR+HyDE8.0688.2608.542

Times are seconds per query, amortized over batches and summed across separately measured stages. Recorded failed ThinkQE generation attempts are excluded. CPR costs more than LLM-free Base and ECLIPSE-PRF, but substantially less than generation-based HyDE and ThinkQE. When added to the same HyDE evidence, CPR incurs only 0.081–0.108 additional seconds per query and improves mean nDCG@10 by 6.26.

Comparison with further query expansion The expansion analysis shows substantial semantic redundancy among multiple generated texts. Generating still more alternatives increases generation and encoding costs; CPR instead reuses corpus evidence without additional LLM calls. In the reported setup, CPR alone takes 0.316–0.409 seconds per query versus 7.985–8.433 seconds for HyDE and 50.330–50.839 seconds for ThinkQE. CPR+HyDE shares HyDE's existing generations, so its additional online cost is limited to the CPR retrieval step.

CorpusDocumentsCPR search (ms/query)GPU peak (GiB)CPU peak RSS (GiB)
Pony7,8944.930.091.07
Earth Science121,24927.432.456.14
LeetCode413,932316.9014.8232.34

The GPU cached-input scaling results above cover 7,894 to 413,932 documents (E5) and show increased search cost with corpus size. They are not the same execution setup as the CPU-based CPR processing reported in the six-method table, and do not establish online latency beyond 413,932 documents.

CPR retrieval was measured on four AMD EPYC 7313 CPU cores (FP64), while generation, query encoding, and the other methods' final retrieval were run on RTX 6000 Ada GPUs. Logical FLOPs were counted from executed matrix operations (one multiply–add = two FLOPs), not converted from device runtimes; the CPU/GPU timings therefore reflect the stated heterogeneous setup. Model loading, corpus encoding, graph construction, and lexical-statistics preprocessing are excluded from online totals.

Offline document-relation statistics are built once and reused. A separate E5 test at one million unique documents measured graph construction (137.32 seconds) and density computation (27.48 seconds), but not one-million-document online search latency.

We have clarified the flexibility of Aq\mathcal{A}_q in Section 2 and included the computational analysis in Section 8 and Appendix H, covering online processing time, corpus-size-dependent retrieval costs, memory usage, and offline preprocessing.

Q1

We additionally evaluated BM25 with eight HyDE-generated passages per query, using two expansion strategies: concatenation and RRF over the original query and eight expansions (RRF9). The table reports the overall mean nDCG@10.

MethodnDCG@10
BM2513.381
BM25 + HyDE (Concat)24.609
BM25 + HyDE (RRF9)23.753
CPR27.035
CPR + HyDE32.836

Query expansion substantially improves BM25 performance, confirming its effectiveness for sparse retrieval. However, using the same HyDE-generated passages, CPR+HyDE outperforms BM25+HyDE (Concat) and BM25+HyDE (RRF9) by 8.227 and 9.083 nDCG@10 points, respectively. CPR alone also outperforms both BM25+HyDE variants. These results suggest that CPR remains competitive against sparse retrieval enhanced with query expansion, while benefiting further from combining corpus-conditioned evidence with generated expansions.

We have included the BM25 query expansion comparisons in Section 6.3 and Appendix F.3, using the same HyDE-generated passages for a controlled comparison.

Thank you for your response! First, my concerns regarding the computational cost have been addressed.

However, I still have concerns regarding the motivation and necessity of the proposed method.

  • About experiment provided to support the motivation: I believe that when multiple information are relevant to a query, it is important to evaluate whether the generated evidence can diversely cover these “relevant information”. However, the provided experiment measures similarity without distinguishing between relevant and irrelevant information. Therefore, I believe that the conclusion drawn from this experiment needs to be reconsidered.
    • More specifically, you argue that LLMs fail to generate semantically diverse evidence based on within-evidence similarity alone. However, couldn't this be because LLMs consistently generate evidence that comprehensively covers all the diverse information relevant to the query?
      • For example, suppose that information "A", "B", and "C" are all relevant to query 1. If an LLM consistently generates outputs that fully cover "A", "B", and "C", these outputs could have high similarity to the query (as dense retrievers are trained to make query-document similarity reflect relevance) and high similarity to one another, even though they comprehensively cover all the relevant information.
    • Conversely, CPR shows not only lower within-evidence similarity but also lower query-evidence similarity. Could this be because the incorporated evidence includes irrelevant information such as "D" and "E", rather than the relevant information "A", "B", and "C"?
  • About Q1 experiment: What I hoped to see from this experiment was a comparison with reasoning query expansion methods such as ThinkQE. When I mentioned in my review that sparse retrievers perform well on reasoning-intensive retrieval tasks, I was referring to the strong performance of methods that incorporate reasoning into query expansion, as explored in BRIGHT [1].

Furthermore, while reading the rebuttal, I noticed a statement that I found somewhat unfamiliar. I would like to clarify whether I may have communicated my point unclearly in the original review or whether this was simply included by mistake.

  • I do not recall explicitly stating in my review that "generated expansions may concentrate on similar concepts." Was there any part of my review that might have led to this interpretation?

[1] BRIGHT: A Realistic and Challenging Benchmark for Reasoning-Intensive Retrieval; Su et al

Reviewer_KoxbOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes Corpus-conditioned Predictive Retrieval (CPR), aiming to use corpus documents as support points for latent intents in underspecified queries. Based on posterior-predictive decomposition, the method combines semantic evidence, lexical evidence, and inter-document relations, achieving performance improvements across multiple tasks and frozen retrievers.

Critical strengths

1, The experiments are relatively extensive, covering multiple tasks and frozen retrievers and reporting multiple metrics.

Critical weaknesses

1, Equation (1) defines zz as an aspect of the information need that is not explicitly expressed in the query. Equation (2) then directly restricts the support points of zz to corpus documents. This can be understood as a corpus-supported discrete approximation, but it is not a strict equivalent transformation naturally derived from Equation (1). Documents may contain evidence related to latent intents or serve as partial representations of those intents, but documents themselves are generally not equivalent to the user’s unexpressed information needs.

2, The paper motivates latent intents by query underspecification, but it does not justify why a single corpus document is an appropriate representation of an intent. Many intents are compositional or relational, and may not be well represented by individual documents.

Questions for authors

1, What is the computational and memory cost of using all eligible corpus documents as support points, and how does this cost scale with corpus size?

2, Why are the semantic prior and semantic–lexical prior assigned equal weights?

3 comments

We sincerely thank the reviewer for their thoughtful and constructive feedback. We greatly appreciate the time and effort devoted to reviewing our work. We have revised the manuscript to clarify the formulation and provide additional experimental evidence addressing the concerns. We address each point below.

W1

We agree that Equation (2) is not an exact transformation of Equation (1). Our intention is to construct a corpus-supported discrete approximation to the posterior-predictive formulation, rather than to identify corpus documents with the user's true latent intents.

Specifically, our motivation is that corpus documents may contain concepts and relations that are missing from an underspecified query but are necessary to determine relevance. We therefore operationalize latent intents using corpus documents, treating each document as a partial representation of a possible information need rather than an exact equivalent of the user's intent.

Under this modeling choice, we define Aq\mathcal{A}_q as the set of eligible document-based intent candidates and assign each z∈Aqz \in \mathcal{A}_q a support prior p(z∣q)p(z\mid q). Equation (2) then approximates the marginalization in Equation (1) by averaging source-conditioned relevance predictions across these candidates. This is a corpus-supported discrete approximation, not an exact transformation of the original latent-intent distribution.

During the reviewing process, we noticed that our statement defining Aq\mathcal{A}_q (in Section 2) might mislead the readers. What we intended to write was that (1) Aq\mathcal{A}_q is usually a relatively small subset of CC, (2) retriever can restrict the search space using documents Aq\mathcal{A}_q, and (3) C−qC-{q} can be a possible case of Aq\mathcal{A}_q. As the current manuscript misleads readers to (3) only, we modified the description about Aq\mathcal{A}_q in Section 2.

The support prior combines reciprocal-rank distributions constructed from semantic and semantic–lexical evidence. This allows the model to retain semantic support when the query omits relevant terminology while incorporating lexical evidence when available.

We have revised Section 2 to clarify the corpus-supported approximation, the definition of Aq\mathcal{A}_q, and its role in Equation (2). The construction of the support prior is described in Section 2.1.

W2.

CPR does not assume that a single corpus document fully represents a user's latent intent. The support space is the set of eligible documents Aq\mathcal{A}_q, whose cardinality is not restricted to one. Each document provides partial evidence about potentially relevant concepts or relations, and p(z∣q)p(z\mid q) determines its contribution to the relevance prediction.

Importantly, Equation (2) does not select one document as the user's intent. It forms a weighted mixture of source-conditioned relevance predictions across the eligible support set. Thus, multiple documents can contribute different aspects of the information needed to assess a candidate.

We nevertheless agree that averaging individual source contributions does not explicitly model interactions between complementary documents. To examine this limitation, we additionally compared individual-document supports with a two-document centroid representation on 30 prespecified queries across three backbones. Relative to matched individual-document controls, the changes in nDCG@10 were +0.683 [0.063, 1.388] for E5, +0.285 [−0.665, 1.759] for Qwen3, and −0.250 [−0.794, 0.117] for ReasonIR.

These limited results do not establish a consistent advantage for group representations. We therefore distinguish aggregating evidence from multiple sources, which CPR already performs, from explicitly modeling their joint interactions. The latter remains an important direction for extending CPR to compositional and relational information needs. We have clarified the role of multiple document supports in Section 2 and discuss the limitations of individual-document representations in Conclusion and Limitations.

Q1

We measured the online components of Base, ECLIPSE-PRF, CPR, HyDE, ThinkQE, and CPR+HyDE. Costs include query-side encoding, LLM generation and feedback where applicable, and final retrieval.

MethodE5Qwen3ReasonIR
Base0.0150.0410.046
ECLIPSE-PRF0.0160.0430.048
CPR0.3160.3800.409
HyDE7.9858.1788.433
ThinkQE50.33050.56450.839
CPR+HyDE8.0688.2608.542

Times are seconds per query, amortized over batches and summed across separately measured stages. Recorded failed ThinkQE generation attempts are excluded. CPR costs more than LLM-free Base and ECLIPSE-PRF, but substantially less than generation-based HyDE and ThinkQE. When added to the same HyDE evidence, CPR incurs only 0.081–0.108 additional seconds per query and improves mean nDCG@10 by 6.26.

CPR retrieval was measured on four AMD EPYC 7313 CPU cores (FP64), while generation, query encoding, and the other methods' final retrieval were run on RTX 6000 Ada GPUs. Logical FLOPs were counted from executed matrix operations (one multiply–add = two FLOPs), not converted from device runtimes; the CPU/GPU timings therefore reflect the stated heterogeneous setup. Model loading, corpus encoding, graph construction, and lexical-statistics preprocessing are excluded from online totals.

Cost as corpus size increases

To assess corpus-size scaling, we separately measured CPR on three existing collections, retaining all eligible corpus documents as supports. The following search times use cached query inputs on the same GPU setup; peak GPU allocation and CPU RSS were measured in isolated method processes.

CorpusDocumentsCPR search (ms/query)GPU peak (GiB)CPU peak RSS (GiB)
Pony7,8944.930.091.07
Earth Science121,24927.432.456.14
LeetCode413,932316.9014.8232.34

The measured search latency and memory footprint increase with collection size; these figures describe the tested collections rather than a fitted asymptotic scaling law. The document-relation graph and lexical statistics are computed offline and reused across queries. The cached GPU search timings above should not be equated with the CPU-based CPR online costs in the six-method table.

Offline construction costs

At 413,932 LeetCode documents (E5), graph construction took 22.51 seconds, density computation 7.29 seconds, and lexical indexing 33.17 seconds. These one-time preprocessing costs are separated from online search.

We have added the computational analysis to Appendix H (Computational Cost and Scalability), including online processing time, corpus-size-dependent retrieval costs, memory usage, and offline preprocessing. The main findings are summarized in Section 8.

Q2

We assigned equal weights to the semantic and semantic–lexical priors to avoid favoring either source of evidence. Following the reviewer's suggestion, we conducted a sensitivity analysis with α = 0.25, 0.5, and 0.75. The overall mean nDCG@10 scores were 25.658, 27.035, and 26.789, respectively. Equal weighting achieved the highest score among the tested settings, supporting its use as a reasonable default. This sensitivity analysis is reported in Section 5.1 (Support Prior) and Appendix E.1.

Dear Reviewer Koxb,

Thank you again for your thoughtful feedback. We would greatly appreciate it if you could review our responses and let us know whether they have addressed your concerns or if any questions remain. Thank you for your time and consideration.

Best regards, The Authors

Comments on the paper

1 comment

Dear Area Chair,

We thank you and the reviewers for your careful consideration. To facilitate your assessment, we summarize how we addressed the principal concerns through manuscript revisions and additional experiments.

  1. Formulation and intent representation (Koxb, qdmF). We clarified Eq. (2) as a corpus-supported discrete approximation, distinguished the roles of the support set and corpus background, and explained how multiple documents contribute partial evidence. We also evaluated paired-document supports and explicitly discussed the limitation concerning joint interactions between documents.

  2. Contribution beyond existing retrieval signals (qdmF). With the underlying signals and candidates held fixed, CPR improves over same-signal RRF by 1.44 nDCG@10 (95% CI: [1.21, 1.66]). This gain persists under a matched computational budget, providing controlled evidence for CPR’s predictive scoring.

  3. Empirical reliability and design choices (qdmF, Koxb). Paired-bootstrap analyses across all task support improvements of CPR over Base and CPR+HyDE over HyDE on all four evaluated backbones, including ReasonIR and the additional BGE-M3 evaluation. We also provided query-weighted results and design-sensitivity analyses, and clarified the interpretation limits of the label-dependent diagnostics.

  4. Efficiency and complementarity with query expansion (qHuY, Koxb, qdmF). We reported runtime, memory, preprocessing, and million-document search costs using cached inputs. We also compared BM25, dense ThinkQE, and ThinkQE+CPR using identical generated inputs; adding CPR improves nDCG@10 and Recall@100 over dense ThinkQE across all three tested retrievers.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.