acceptodds
Under review as a conference paper at ICLR 2027

Pay to Reveal: Adaptive Information Acquisition under Communication Budgets

Abstract

Retrieval-augmentedgeneration(RAG)groundslanguage-modelanswersinexter nalcorpora,andmulti-hopquestionsoftenmakelaterretrievaldependonevidence foundearlier.Existingiterativesystemstypicallyquerybody-derivedindexesor rerankretrievedcandidatepools.Westudyacomplementarydecision:acontroller canenumeratedocumenttitles,buteachbodymustbeacquiredunderalimited readbudget.WeformalizethissettingasPaytoReveal.Apolicymayacquireat mostKdocuments;eachacquisitionreturnsaquestion-conditionedexcerptcapped atLunderacharacter-basedtokenestimate;andaccumulatedexcerptsguidelater choices.WeproposeBM25-RF,aniterativepseudo-relevance-feedbackpolicy thatexpandsthequerywithacquiredexcerptsandreranksunopenedtitleswith BM25.WealsoproposeResidual-CMG,whichaddstothestatictitle-BM25score thedifferencebetweenasupport-documentclassifier’sacquired-statescoreandits empty-statescoreatthesamestep.Theclassifieristrainedonlyonthetrainingsplit. Underanexplicitexpected-progresscondition,afinite-Kanalysisrelatesper-step selectiongapstofinalsupport-documentrecall.Weevaluatesixmethodsunderone protocolon22,246retrievalquestionsand3,000Qwen3-4B-Instruct-2507answer questionsfrom2WikiMultiHopQA,MuSiQue,andHotpotQAFullWiki.Relative toStatic-BM25,BM25-RFchangesanswerF1by+.067,+.049,and−.028across thethreedatasets.Addingthestatictitle-BM25scoretothelearnedscoreimproves supportAPoverDirect-CMGonallthreedatasets.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
retrieval-augmented generation (RAG), adaptive information acquisition, sequential evidence acquisition, multi-hop question answering, communication-efficient retrieval.

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_kqMrOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies Pay to Reveal, a sequential retrieval setting in which a controller can inspect a fixed catalog of document titles but must spend a limited acquisition budget to obtain question-conditioned, length-capped excerpts from document bodies. It proposes two policies: BM25-RF, which expands the retrieval query with previously acquired excerpts, and Residual-CMG, which combines static title-BM25 relevance with the residual signal from a support-document classifier, defined as the difference between the classifier's scores in the acquired state and those in the empty state. The paper also derives a finite-horizon support-recall bound under an explicit expected-progress assumption. Experiments cover 22,246 retrieval questions and 3,000 downstream answer questions across 2WikiMultiHopQA, MuSiQue, and HotpotQA FullWiki. BM25-RF improves both acquisition and downstream answer quality on the first two datasets but degrades on HotpotQA. Retaining the static BM25 prior improves the learned policies' support-document average precision relative to a classifier-only control.

Critical strengths
  1. The paper specifies the information-access protocol precisely enough to support replication and evaluates it at substantial scale: 22,246 acquisition questions and 3,000 downstream answer questions, with paired bootstrap confidence intervals and Holm-corrected tests where reported. This is a substantial empirical investment for a sequential-retrieval study.

  2. It provides credible evidence that simple lexical feedback (BM25-RF) improves both acquisition and downstream F1 on 2Wiki and MuSiQue—including raising all-support coverage from 0.3432 to 0.6104 on 2Wiki—and it reports the HotpotQA degradation transparently rather than suppressing it.

Critical weaknesses
  1. The residual‑state subtraction component in Residual‑CMG lacks independent empirical validation. As shown in Table 11, Prior+Full‑CMG accounts for 89 %‑99 % of the AP improvement from Direct‑CMG to Residual‑CMG. Residual‑CMG only brings marginal AP increments (0.0171, 0.0002, 0.0049). Critically, there are no paired statistical significance tests for these small gaps, nor corresponding downstream answer‑quality comparisons. Therefore it remains unclear whether the residual correction provides tangible practical benefits beyond simply retaining static title‑BM25 prior.

  2. Given the linear logistic classifier setup, subtracting empty‑state logit cancels out all classifier features invariant to acquired evidence. The residual term is purely an internal state‑dependent contrast computed inside the pre‑trained classifier, rather than a genuinely separate, independent ranking signal. This blurs the conceptual novelty of Residual‑CMG as a new scoring formulation.

Questions for authors
  1. Classifier training uses local candidate pools and local BM25 normalization, but inference evaluation uses full‑catalog normalization and retrieval‑generated candidate sets. What empirical evidence separates genuine algorithmic gains from artifacts compensating this train‑evaluation mismatch?

  2. Is there any evaluated experimental setting where both progress coefficient γ and per‑step selection‑regret produce non‑vacuous, numerically informative bounds for Theorem 1? If no such setting exists, should Theorem 1 be re‑framed in the manuscript as a conditional accounting framework rather than an approximate performance guarantee for BM25‑RF / Residual‑CMG?

  3. Subtracting the empty‑state logit cancels static features in this linear classifier, making the residual merely an internal contrast rather than an independent ranking signal. Can targeted ablations separate gains from the state‑dependent contrast operation versus combining classifier outputs with the static title‑BM25 prior?

1 comment
AuthorsOpenReview ↗

Response to Reviewer kqMr

Q1. Does the residual gain survive training with retrieval candidates and catalog normalization?

The original answer-F1 advantage on 2Wiki and HotpotQA reversed under this training condition. At the same λ = 0.1, simple fusion (Prior+Full-CMG) led Residual-CMG by 0.0856 F1 on 2Wiki and 0.0697 on HotpotQA. On MuSiQue the residual led by 0.0084, with a paired 95% interval of [−0.0003, +0.0173] (Table 17, p. 23).

We varied candidate source and BM25 normalization using the same training questions, histories, state-weighting rule, feature definitions and procedure for selecting regularization. Each fit selected its own C. The table gives support AP differences (Residual minus Prior+Full); paired 95% intervals appear in Table 16 (p. 22).

Training candidates / normalization2WikiMuSiQueHotpotQA
Context / context−0.0013+0.0007+0.0018
Context / catalog−0.0003+0.0012+0.0053
Retrieval / context−0.1227+0.0066−0.0214
Retrieval / catalog−0.1236+0.0103−0.1113

On 2Wiki, switching to retrieval candidates favored simple fusion under either normalization. On HotpotQA, its lead was largest with catalog normalization and retrieval candidates. On MuSiQue, retrieval candidates favored the residual on AP. The original fit used different histories, so only comparisons among the four retrained models isolate the two training choices.

Q2. Does Theorem 1 give a numerically informative guarantee for either policy?

No. None of the evaluated settings supplies both an informative progress coefficient γ and a bound on per-step selection regret. We adopted the reviewer’s suggested framing: Section 4.3 presents Theorem 1 as a conditional lower bound on expected support-document recall. Equations (7) and (8) show how a given rate of progress and the accumulated losses from action selection determine that bound (p. 5, proof in Appendix A.1, p. 13).

The branching experiment measures final recall after continuing from eight alternative next reads (Table 25, p. 25). Estimating the theorem’s selection regret would require conditional one-step gains over the full catalog. The branch outcomes also supply no estimate of γ.

Q3. What does empty-state subtraction add beyond static score fusion?

You are right about the algebra. In this linear classifier, subtracting the same candidate’s empty-evidence logit at the same step cancels the intercept and every feature contribution that stays fixed. The residual measures how the evidence changes the score inside one classifier. Static title relevance enters separately through the BM25 prior.

With b(e) the normalized static title-BM25 score and f the classifier logit, Equation 5 (p. 5) defines the residual score. The three scoring rules are

sdirect(e)=f(e,Ht),sfull(e)=b(e)+λf(e,Ht),sres(e)=b(e)+λ[f(e,Ht)−f(e,∅t)].\begin{aligned} s_{\mathrm{direct}}(e)&=f(e,H_t),\\ s_{\mathrm{full}}(e)&=b(e)+\lambda f(e,H_t),\\ s_{\mathrm{res}}(e)&=b(e)+\lambda\big[f(e,H_t)-f(e,\varnothing_t)\big]. \end{aligned}

Direct versus Prior+Full tests the difference between using the classifier alone and adding its λ-scaled logit to BM25. Prior+Full versus Residual holds the classifier, prior and λ fixed and changes only the subtraction in the ranking rule. At the submitted λ = 0.1, subtraction improved answer F1 on 2Wiki and HotpotQA. The table reports question-paired results on 1,000 questions per dataset with six reads and the same Qwen3-4B-Instruct-2507 reader (Table 15, p. 22).

DatasetPrior+Full F1Residual F1Residual − Prior+Full F1 [95% CI]Holm p
2Wiki0.41000.4522+0.0422 [0.0217, 0.0627]0.0032
MuSiQue0.06810.0682approximately 0 [−0.0057, 0.0056]0.9927
HotpotQA0.28090.2994+0.0186 [0.0057, 0.0319]0.0252

For the support-AP increments you identified, the Holm-adjusted p values were 0.0030 for +0.0171 on 2Wiki, 0.8661 for +0.0002 on MuSiQue and 0.0030 for +0.0049 on HotpotQA. Table 15 gives the paired intervals for AP, F1 and exact match. AP and F1 each use a separate 15-test Holm correction.

The answer advantage depended on the weight. When each rule selected λ by AP@6 on 500 held-out training questions per dataset, the paired F1 differences (Residual minus Prior+Full) were −0.0054 on 2Wiki, approximately zero on MuSiQue and +0.0015 on HotpotQA. All three intervals included zero (Tables 17 and 18, p. 23).

Reviewer_nCb3OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a new problem setting in RAG: Pay to Reveal. Different from traditional RAG, which retrieves based on semantic matching on the document contents, Pay to Reveal takes into the consideration of the costs of reading the document body. It studies the title of the document before its body content can be observed. The paper proposed two policies, BM25-RF and Residual-CMG, for this task. Experiments conducted on three QA datasets showing the proposed retrievel method brings improvements compared to traditional BM25 and other baselines.

In general, this paper needs significant improvement before publishing. The problem framing, experiment settings, and the methods need to be clearly defined.

Critical strengths

The paper targets a special angle in RAG: instead of retrieving relevant documents based on the content, it defines policies and protocols with certain budgets to retrieve only document titles before revealing their contents.

Critical weaknesses
  1. The presentation and writing of the paper need significant improvement. Many concepts introduced in the paper lack proper description and explanation, making it hard to follow. For example, what is “state” in the problem setting? What are “acquired-state” and “empty-state” at L061 in RAG? The task and problem setting are also hard to understand.
  2. The proposed methods, BM25-RF and Residual-CMG, are essentially policies and protocols, which are more like engineering methods instead of ML/DL methods. Furthermore, the improvements of the methods are marginal in all three QA datasets compared to established baselines, making the effectiveness of the methods questionable.
Questions for authors
  1. What does the task of Pay to Reveal try to do? Why is it useful in the real world or to the community?
  2. What is “state” in the problem setting? What are “acquired-state” and “empty-state” at L061 in RAG?
3 comments
AuthorsOpenReview ↗

Response to Reviewer nCb3

Q1. What does Pay to Reveal study, and where does this decision arise?

Pay to Reveal asks which document to open next when only a limited number can be read. In Figure 1 (p. 4), the first excerpt from Everyday Is Christmas (song) names Jacky Cheung. His title had a BM25 score of 0 under the original question. With the excerpt added, it scores 21.56 and becomes the next document selected. The static ranking would open Everyday instead.

In the main experiment, the selector starts with the question and a title catalog. Each read spends one of six opportunities. A shared extractor scores sentences in the opened document by IDF-weighted overlap with the question and returns selected sentences in document order, using a sentence-selection budget of 256 estimated tokens. The selector sees this excerpt before choosing again. The answer model receives the accumulated excerpts after selection (Section 3, p. 3).

A search result card presents a similar choice: a reader sees a description before deciding whether to fetch the document. On the same 830 BrowseComp-Plus queries, BM25-RF’s document AP difference from static ranking was −0.0007 [−0.0052, 0.0039] with titles alone and +0.0111 [0.0047, 0.0177] with native result cards. Thus feedback improved the ranking of relevant documents with native cards. The answer-accuracy difference intervals included zero for both readers (Appendix G.3, p. 24).

Q2. What are the state, acquired state and empty state?

The state records the documents already opened, in order, and the excerpt returned by each. The question and the catalog stay fixed. In the example above, after one read the acquired state contains the Everyday Is Christmas (song) document ID and its excerpt naming Jacky Cheung.

To compute the empty-state score, we score the same candidate at the same step with the evidence input empty. For example, the classifier scores the Jacky Cheung title once with the first excerpt and once with that excerpt withheld. Residual-CMG adds the λ-scaled difference between those two logits to the static title-BM25 score. BM25-RF uses the excerpt directly to expand the question and rerank the remaining titles (Section 3, p. 3 and Equations 4–5, pp. 4–5).

What do the two policies contribute, and how large are the gains?

BM25-RF uses iterative pseudo-relevance feedback. Its contribution here is the test of when to update: changing only the update schedule changes answer quality even when all policies read six documents and use the same reader. Compared with no updates, updating after the first two reads raised answer F1 by 0.0976 on 2Wiki and 0.0596 on MuSiQue (Holm-adjusted p = 0.0030 for each). On HotpotQA the difference was +0.0119 [−0.0108, 0.0347]. Against a body-indexed reranker, the original every-read BM25-RF policy led on 2Wiki and the reranker led on MuSiQue and HotpotQA (Tables 6 and 14, pp. 8 and 22).

Residual-CMG uses a logistic classifier trained to predict whether an unopened candidate is a supporting document. The experiment asks whether adding the difference between the acquired-state and empty-state logits improves on adding the full acquired-state logit to the same static score. At the original λ = 0.1, the score change gave higher answer F1 on 2Wiki and HotpotQA. With each rule’s weight selected on training data, all three paired intervals for the F1 difference included zero. When we retrained on retrieval-generated candidates with catalog normalization, simple fusion gave higher F1 on 2Wiki and HotpotQA (Tables 15–17, pp. 22–23).

Reviewer_nCb3OpenReview ↗

Response to Authors

Thank the authors for the detailed response.

The clarifications regarding the Pay to Reveal setting and the definition of the state help address my concerns about the presentation and problem formulation. The new results provide stronger evidence that adaptive feedback can improve retrieval and downstream QA performance, particularly on 2Wiki and MuSiQue. However, I still have several concerns:

  1. Practical motivation and generalisability. While the provided example helps motivate the setting, the practical relevance of the title-only acquisition protocol remains limited. The additional BrowseComp-Plus results show a small improvement in document AP, but no statistically significant improvement in downstream answer accuracy
  2. Methodological novelty. BM25-RF is largely based on existing pseudo-relevance feedback techniques, while Residual-CMG applies a relatively straightforward score adjustment. The contribution therefore lie primarily in the problem formulation and empirical evaluation
  3. Robustness of the empirical improvements. The new experiments demonstrate meaningful improvements on two datasets, which addresses part of my original concern. However, the results remain sensitive to the dataset and feedback schedule. Moreover, the advantage of Residual-CMG over simple score fusion is not consistently supported after separate hyperparameter tuning or changes to the classifier training configuration

Overall, the rebuttal has clarified several important aspects of the paper and strengthened its empirical evaluation. However, I remain concerned about the limited methodological novelty and the generalisability of the proposed approaches.

Thank you for clarifying your concerns. The practical decision we study arises when a search returns a list of results and the reader can open only a few. Suppose the first document names someone absent from the question. Should the reader now open a different result? IRCoT can use that clue to issue another query against a passage index. Yoo and Jeong (2026) study how investigators can identify relevant files from their names. We study what happens after a file is opened: can its contents help the reader choose the next document from the same list? Figure 1 shows the difference in a real question. Once the first excerpt names Jacky Cheung, BM25-RF opens his entry next. Static BM25 opens Everyday.

BM25-RF uses classical pseudo-relevance feedback. We compare two ways of choosing six documents from the same title catalog. Static BM25 follows the initial ranking. The original BM25-RF policy uses each new excerpt to rerank the remaining titles. Across 12,536 retrieval questions on 2Wiki, BM25-RF found all labeled supporting documents on 61.04% of questions, compared with 34.32% for static BM25 (Table 5). Answer F1 also improved on 2Wiki and MuSiQue and fell on HotpotQA (Table 4). With the richer result cards in BrowseComp-Plus, feedback improved document AP, while the difference in answer accuracy was statistically inconclusive for both readers (Appendix G.3).

We agree that the gains depend on how feedback is used. At the same six-read budget, updating after reads one and two and keeping that ranking for the remaining reads gave higher mean answer F1 than updating throughout on all three datasets. The difference was significant on MuSiQue and HotpotQA after correction (Tables 6 and 14). For Residual-CMG, separate weight tuning and changes to classifier training likewise yielded no consistent advantage over simple score fusion (Appendix G).

Reviewer_gCTAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  1. Multi-hop RAG where the policy sees only titles, may open K = 6 documents, and gets a 256-token excerpt per opening.
  2. Two greedy policies: BM25-RF appends acquired excerpts to the query and reranks unopened titles (Eq. 4); Residual-CMG adds a classifier-score change to static title-BM25 (Eq. 5). A recall bound (Thm. 1) holds under an expected-progress premise.
  3. BM25-RF raises answer F1 over Static-BM25 by .067 and .049 on 2Wiki and MuSiQue, and lowers it by .028 on HotpotQA, where a static dense retriever wins.
Critical strengths
  1. Controlled protocol: all six methods share one extractor, budget, answer model, and title index (Sec. 3, 5).
  2. Paired bootstrap with Holm correction, and the negative HotpotQA result is reported.
  3. Honest about the theorem: Sec. 4.3 says the progress premise is not established for either policy.
Critical weaknesses
  1. No comparison with the conventional pipeline. Every method searches titles only (Sec. 5); body-indexed sequential RAG appears in Table 1 as an adjacent setting and is never run. The paper shows what works inside its restriction, not that it beats retrieve-then-rerank over bodies [1].
  2. The motivation is asserted. The title-only, pay-per-open setting is imposed, with no deployment shown. The 2Wiki and MuSiQue catalogs are dev gold plus distractors (App. C.3), and the budget counts openings, not cost (Sec. 6.3).
  3. Adaptation is never isolated. The prefix analysis (Sec. 6.3, Table 7) varies the number of acquisitions, i.e. the budget. The needed ablation fixes K = 6 and varies when the ranking updates: never, first two openings only, every step.
  4. Greedy is not tested. The policy opens the top-scored title each step (Sec. 4.1); App. A itself notes the bound weakens when an action matters for what it reveals later. No lookahead, beam, or same-state comparison.
  5. The gain is narrow. Feedback helps on two constructed catalogs and hurts on the only open corpus (F1 −.028, AP −.054). For the learned selector, Sec. 6.2 concludes the gain over Direct-CMG comes from retaining the static prior; against Static-BM25 it is indistinguishable on HotpotQA, and the empty-state subtraction adds at most .017 AP and is missing from the answer matrix (Sec. 6.1). λ = 0.1 is a preset, and K and L are never varied.
Questions for authors
  1. A body-indexed retrieve-and-rerank baseline with the same answer model and six documents. How far is the title-only pipeline from it?
  2. The fixed-K ablation: no update / update for two openings then freeze / update throughout.
  3. Rerun 2Wiki and MuSiQue on a full-Wikipedia title catalog. Does BM25-RF still help?
  4. From the same state, how often does the top-scored title give the best final recall?
  5. Was λ = 0.1 tuned on 2Wiki / MuSiQue dev data? Give a λ and K sweep and add Prior+Full-CMG to the answer matrix.
1 comment
AuthorsOpenReview ↗

Response to Reviewer gCTA

Q1. Body retrieval and reranking

BM25-RF led the body reranker by 0.0510 answer F1 on 2Wiki. The reranker led by 0.0668 on MuSiQue and 0.2088 on HotpotQA (all Holm-adjusted p < 0.01). The baseline retrieved 100 candidates by title-and-body BM25 and reranked their text with cross-encoder/ms-marco-MiniLM-L6-v2. Both methods supplied six excerpts to Qwen3-4B-Instruct-2507 on the same 1,000 questions per dataset, using the same extractor (Tables 6 and 14, pp. 8 and 22).

Q2. Feedback timing at six reads

Restricting updates to the first two reads gave the highest mean F1 of the three schedules on every dataset:

Update schedule2Wiki F1MuSiQue F1HotpotQA F1
No updates0.39180.06420.3003
Updates after reads one and two0.48940.12380.3122
Updates throughout0.46820.11320.2741

Compared with no updates, two updates improved F1 on 2Wiki and MuSiQue (Holm p = 0.0030 each). On HotpotQA the gain was 0.0119 (95% CI [−0.0108, 0.0347]). Two updates also beat updates throughout on MuSiQue and HotpotQA after correction (Tables 6 and 14).

Q3. Full Wikipedia catalog

The gain did not reproduce on either full answer cohort. Across 5.23 million Wikipedia abstracts, BM25-RF / Static-BM25 scored 0.3277 / 0.3328 F1 on 2Wiki and 0.0374 / 0.0346 on MuSiQue (1,000 questions each; both paired intervals included zero). Among questions with one-to-one support mappings, the differences were +0.1432 on 77 2Wiki questions and +0.0289 on 55 MuSiQue questions; only the 2Wiki interval excluded zero (Table 23, p. 24).

Q4. Does the highest-scoring next read give the best final recall?

Usually, in a tie. We continued each of the eight highest-scoring next reads to six reads with the same policy (900 states per policy and dataset). In 2Wiki / MuSiQue / HotpotQA order, the top action attained the best final recall in 80.11% / 85.78% / 87.11% of states for BM25-RF and 84.11% / 91.33% / 95.33% for Residual-CMG. It was uniquely best in at most 1.56% of states (Table 25, p. 25).

Q5. Weight selection, read budget and Prior+Full-CMG

The original λ = 0.1 was selected by Residual-CMG MAP@4 on held-out official training data for 2Wiki and MuSiQue, then carried over to HotpotQA. The sweep below gives validation AP@6 for Prior+Full-CMG / Residual-CMG (Table 18, p. 23). When each rule maximized support AP@6 on 500 held-out training questions per dataset, both chose 0.03 on 2Wiki and HotpotQA and 0.1 on MuSiQue. Table 6 now includes Prior+Full-CMG. At the original weight, Residual-CMG had higher answer F1 on 2Wiki and HotpotQA; after separate tuning, all three paired F1 difference intervals included zero (Tables 15 and 17, pp. 22–23).

λ2WikiMuSiQueHotpotQA
00.6443 / 0.64430.4766 / 0.47660.4326 / 0.4326
0.030.6555 / 0.64690.4815 / 0.48320.4450 / 0.4343
0.10.6471 / 0.63960.4930 / 0.49270.4319 / 0.4309
0.30.5824 / 0.48100.4633 / 0.42550.3608 / 0.3204
10.4667 / 0.48050.4444 / 0.42110.3253 / 0.3194

At K = 2, 4, 6 and 10, BM25-RF exceeded static F1 on 2Wiki and MuSiQue and fell below it on HotpotQA (300 questions per dataset, L = 256). The same ordering held at six reads with excerpt budgets of 128, 256 and 512 estimated tokens (Tables 19–20, p. 23).

Dataset and methodK = 2K = 4K = 6K = 10
2Wiki, Static-BM250.38490.41970.41240.4211
2Wiki, BM25-RF0.44770.48300.48720.4961
MuSiQue, Static-BM250.05520.05670.06810.0790
MuSiQue, BM25-RF0.10700.11360.11850.1155
HotpotQA, Static-BM250.28490.29770.29860.3202
HotpotQA, BM25-RF0.27180.27730.28360.2771

Motivation and measured cost

A search result card lets a reader decide what to open next. In Figure 1 (p. 4), the first excerpt names Jacky Cheung and brings his title to the next read. On 830 BrowseComp-Plus queries, feedback changed document AP by −0.0007 with titles alone and +0.0111 with native result cards (95% CI [0.0047, 0.0177]); both readers’ answer-accuracy difference intervals included zero with native cards (Appendix G.3, p. 24).

With byte limits matched across policies within each dataset (three limits, 300 questions per dataset), BM25-RF gained 0.0742–0.0935 F1 over static selection on 2Wiki and 0.0480–0.0508 on MuSiQue. All three HotpotQA intervals included zero (Table 21, p. 24). Median selection time was lower for BM25-RF than for body retrieval and reranking on 2Wiki and MuSiQue and higher on HotpotQA (Table 24, p. 25).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.