Pith. sign in

REVIEW 4 major objections 6 minor 60 references

Beyond Independent Passages: Adaptive Passage Combination Retrieval for Retrieval Augmented Open-Domain Question Answering

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Retrieving passages as pairs, not one at a time, improves open-domain question answering.

desk verdict A genuinely new passage-pair retrieval idea with plausible potential, but the empirical case is statistically unproven and needs significance tests, multiple seeds, and code before the central claim can be accepted. read the letter →

arxiv 2507.04069 v1 pith:44E4CDBC submitted 2025-07-05 cs.CL cs.AIcs.LG

Ting-Wen Ko , Jyun-Yu Jiang , Pu-Jen Cheng This is my paper
classification cs.CLcs.AIcs.LG
keywords retrieval-augmentedgenerationopen-domainquestionansweringpassagecombinationretrievalmulti-hopreasoningdensererankingblack-boxlanguagemodelsadaptive
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Conventional retrieval-augmented question answering treats each passage as an independent candidate, so contexts can become redundant, noisy, or missing the complementary evidence that multi-hop questions require. This paper proposes Adaptive Passage Combination Retrieval (AdaPCR), which treats a passage combination—either a single passage or a pair built by using one passage to reformulate the query—as the retrieval unit, and reranks combinations so that retriever scores track a frozen language model's answer likelihood. The paper claims that on NQ, TriviaQA, and HotpotQA this combination-based retriever beats independent-passage baselines after reranker training, with the largest gains on multi-hop HotpotQA, and that it adaptively chooses how many passages to keep without a separate stopping module. If right, fixed-size independent retrieval is a bottleneck, and modelling inter-passage dependencies can improve black-box LLM question answering without fine-tuning the generator.

What carries the argument

The load-bearing object is the passage-combination scoring function built on a shared bi-encoder. For a singleton $\langle d_i\rangle$ it is $\cos(E_q(x), E_d(d_i))$; for a pair $\langle d_i, d_{ij}\rangle$ it is $\cos(E_q(d_i \oplus x), E_d(d_{ij}))$, which uses the first passage as part of the query and thereby captures dependency between the two passages. The training objective is the predictive RAG loss from Equations (6)--(7), which averages retriever weights with the frozen generator's answer likelihood over candidate combinations, avoiding softmax normalization of generator outputs. The adaptive selection step, taking the maximum over combinations of one or two passages, replaces a fixed retrieval count and removes the need for a separate stopping classifier. The paper's theoretical appendix shows that maximizing answer probability is achieved by a one-hot retriever, connecting the reranking objective to downstream generation.

What would settle it

Compute the rank correlation, on held-out HotpotQA questions, between the Eq. (3) scores of candidate passage pairs and the frozen generator's actual answer likelihood when each pair is in the context; a near-zero correlation would indicate that the pairwise score is not carrying the retrieval gains.

Watch

Extended reading notes

Core claim

The central discovery is that scoring whole passage combinations, rather than passages one at a time, improves open-domain QA when the generator is frozen. AdaPCR first retrieves a small pool $D_1=\{p_1,\dots,p_k\}$ using the standard bi-encoder cosine score $\cos(E_q(x), E_d(d_i))$, then forms a second candidate set by concatenating each first-stage passage with the question and scoring a second passage via $\cos(E_q(d_i \oplus x), E_d(d_{ij}))$. The final selection is the best-scoring singleton or pair, so the number of passages adapts to the question. The reranker is trained with a RAG-style predictive objective, a weighted sum of retriever confidence and the frozen LLM's answer probability, and the paper argues that the optimal retriever under this objective is one-hot, selecting the single highest-likelihood combination. Empirically, the trained reranker outperforms the independent-passage IC-RALM reranker on all three benchmarks, with the largest EM and F1 advantages on HotpotQA.

Load-bearing premise

The pairwise score rests on an untested assumption: that the cosine similarity between the embedding of first-passage-plus-question and the embedding of a second passage tells you how much that second passage improves the answer.

Editorial extensions

If this is right

  • Multi-hop QA improves even before reranker training: on HotpotQA, AdaPCR beats IC-RALM by 4.30 EM and 6.45 F1 points with an off-the-shelf retriever, showing that combination scoring itself, not only training, recovers complementary evidence.
  • After reranker training, AdaPCR Reranker surpasses IC-RALM Reranker on NQ (+1.10 EM), TriviaQA (+0.60 EM), and HotpotQA (+0.40 to +0.60 EM), so the gains are not limited to the multi-hop setting.
  • Adaptive selection means no fixed retrieval count or stopping module is needed; the model chooses one or two passages per question, which should reduce the distraction of irrelevant fixed-size contexts.
  • Using a balanced positive-sampling strategy during training materially improves EM and F1 on all datasets, particularly HotpotQA, indicating that stable pair-level supervision matters for learning passage dependencies.
  • RAG loss beats CE and KL divergence losses across both DPR and BERT initializations, suggesting that optimizing an unnormalized answer-likelihood objective is more stable than softmax-normalized alternatives.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Extension: the scoring rule for pairs assumes that cosine similarity between the embedding of the first passage plus question and the embedding of a second passage measures the second passage's marginal value; this geometric correspondence is not validated in the paper and could be tested directly.
  • Extension: the two-round scheme could be iterated to more rounds, but retrieval error would likely compound; a natural test is whether three or more rounds help on datasets requiring longer reasoning chains.
  • Extension: since untrained AdaPCR already produces large TriviaQA gains, part of the reported improvement may come from the concatenated-query reformulation itself rather than from reranker training; an ablation holding the reranker fixed and varying only query construction would separate these contributions.
  • Extension: if the result transfers beyond QA, adaptive pair retrieval could be applied to other black-box generation tasks where selecting complementary documents matters, although the paper does not test this.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes AdaPCR, a two-stage retrieval framework for open-domain QA with a frozen black-box generator. In the first stage it retrieves k passages with a bi-encoder retriever; in the second stage it concatenates each first-stage passage with the question, scores a second passage via cosine similarity, and selects either a single passage or a pair as the final context. The reranker is trained with a RAG-style loss that weights passage combinations by their downstream answer likelihood, and the number of retrieved passages is chosen adaptively as one or two. Experiments on NQ, TriviaQA, and HotpotQA (1,000-sample test sets) compare AdaPCR against IC-RALM and an IC-RALM reranker. The paper reports consistent gains after reranker training, especially on multi-hop HotpotQA, and attributes them to explicit modeling of inter-passage dependencies.

Significance. If the central claim is reliable, modeling passage combinations rather than independent passages is a useful design principle for RAG with black-box LMs, particularly for multi-hop questions. The paper has several strengths: the comparison against IC-RALM Reranker is a reasonable way to isolate the effect of combination modeling; the training objective is clearly specified; the frozen-generator setup avoids confounding from generator fine-tuning; and the case studies concretely illustrate how the two-stage scoring can retrieve complementary evidence. The untrained AdaPCR gains on TriviaQA and HotpotQA are large, suggesting the two-stage heuristic has practical potential. However, the post-training margins over IC-RALM Reranker are small (0.4–0.6 EM on all but one dataset, 0.5 on NQ), no statistical significance or variance information is reported, and at least one reported number is inconsistent with the table. The significance of the contribution therefore depends on additional statistical evidence that the current manuscript does not provide.

major comments (4)
  1. [Section 5.1, Table 1] The central claim that AdaPCR-based models 'consistently outperform IC-RALM variants across all datasets after reranker training' is not statistically supported. For the trained comparison, the EM margins over IC-RALM Reranker are NQ +0.50, TriviaQA +0.60, HotpotQA +0.60 (normal) and +0.40 (hard), all on test sets of 1,000 questions. A 0.4–0.6 point EM difference is 4–6 additional correct answers, which is within the range of sampling noise even under a rough unpaired proportion calculation, and no confidence intervals, McNemar tests, or repeated-seed variance are reported. The paper should report paired significance tests (e.g., McNemar) and variance across training seeds, or soften the consistency claim accordingly.
  2. [Section 5.1, Table 1] The text states that 'AdaPCR Reranker achieves a 1.10-point improvement in EM and a 0.49-point improvement in F1 over IC-RALM Reranker' for NQ, but Table 1 shows 23.32 vs. 22.82 EM (+0.50) and 32.34 vs. 32.07 F1 (+0.27). The reported 1.10 and 0.49 values correspond to the comparison against the non-reranked IC-RALM, not against IC-RALM Reranker. This is a factual inconsistency in the headline result and must be corrected.
  3. [Section 5.3, Table 3] The claim that 'RAG loss consistently outperforms CE loss' is contradicted by Table 3. For BERT on HotpotQA-hard, the one-round RAG loss EM is 14.82 versus 15.52 for CE loss, a 0.70-point deficit. Additionally, for BERT on NQ and HotpotQA, the two-round RAG loss is lower than the one-round RAG loss (11.71 vs. 12.31 and 15.12 vs. 16.02). The aggregate averages reported in the text hide these row-level inconsistencies; the conclusion about loss-function choice needs to be qualified.
  4. [Section 3.1, Eq. (3)] The pair scoring function in Eq. (3), s(⟨di,dij⟩,x)=cos(Eq(di⊕x), Ed(dij)), assumes that the cosine similarity between the bi-encoder embedding of the concatenated query and the candidate passage embedding measures the marginal value of dij given di. This is an ad hoc extension of the original bi-encoder scoring function, and the paper provides no targeted validation, such as a correlation between this score and answer likelihood for pairs, or an ablation that replaces Eq. (3) with an alternative composition operator. Since the untrained AdaPCR already produces large gains on TriviaQA and HotpotQA, the reader cannot tell whether the gains come from the combination-scoring mechanism or from the query-reformulation step alone; this should be disentangled.
minor comments (6)
  1. [Abstract] The abstract contains the duplicated phrase 'particularly problematic - particularly problematic'; this should be fixed.
  2. [Section 5.2 and Section 5.5] Figure 2 is referenced for two different purposes: Section 5.2 says the positive-sampling results are shown in Figure 2, and Section 5.5 uses Figure 2 for the second case study. The positive-sampling numbers actually appear in Table 2, so the figure/table references need to be reconciled.
  3. [Section 3.1] The phrase 'we concatenated it with the input x' should be 'we concatenate it with the input x'; the tense is inconsistent with the surrounding description.
  4. [Appendix B] Appendix B proves that a one-hot retriever maximizes the RAG objective, which is immediate from the linearity of Eq. (7) over the simplex. The additional claim that this implies the same optimal behavior as CE loss is not derived, and the paper itself notes that empirical validation is left to future work; the claim should be stated as a conjecture.
  5. [Section 4.4] The sentence 'Since the questions in our dataset are simple questions with short lengths' is inaccurate for HotpotQA, which is a multi-hop dataset; the justification for omitting the retrieval stride should be rephrased.
  6. [References] The Chung et al. (2022) reference is incomplete; it should include the full title and venue, and several other 'arXiv preprint' entries lack version or access information.

Circularity Check

1 steps flagged · score 2.0 of 10

No load-bearing circularity: the empirical comparison is self-contained, though Appendix B's one-hot optimality is a definitional tautology and the RAG/CE equivalence is unproven.

  1. self definitional [Section 5.3 (Comparison of Loss Functions) and Appendix B (Equations 7-10, Proposition 1)]
    "In this setup, dn is the best candidate passage that leads to the highest answer probability, which we aim to find in our optimization process of the retriever: dn = arg max d PLM (y | [d; x]) (10) ... Proposition 1 The one-hot retriever is an optimal retriever. ... Interestingly, however, we theoretically show that the optimal behavior of the retriever under RAG loss should have the same result as CE loss (Appendix B). However, we don’t have the budget to run the training until full convergence, so we leave the empirical validation of this finding to future work."

    Appendix B's 'optimal retriever' is derived from the paper's own definitions rather than from independent content: Eq. (7) defines P(y|x) as a weighted sum of PLM(y|[d;x]) with retriever weights, and Eq. (10) defines dn as the passage maximizing that same PLM term. A one-hot retriever that places all mass on dn therefore maximizes the weighted sum by construction, making Proposition 1 a tautology. Section 5.3 then imports this tautology as evidence that RAG loss and CE loss have the same optimum, a claim Appendix B never actually proves because it does not mention either loss. This is definitional rather than empirical, but it is not load-bearing for the headline EM/F1 results, which are produced by training with a frozen external LM and evaluating on public benchmarks.

full rationale

The paper's central claim is an empirical comparison: AdaPCR-based models are trained with a reranking objective that uses answer likelihood from a frozen Flan-T5-large generator, and they are evaluated with EM/F1 on public QA test sets (NQ, TriviaQA, HotpotQA). No fitted parameter is renamed as a prediction, no external result is imported from the authors' prior work, and the main scoring functions in Eqs. (2)-(3) are explicit modeling choices rather than circular definitions. The only self-referential element is Appendix B, which proves a one-hot retriever is optimal; that result is a tautology following immediately from the definitions of dn and the weighted-sum objective. Section 5.3 additionally claims this proves RAG loss and CE loss have the same optimum, but Appendix B does not analyze either loss, and the paper explicitly leaves empirical validation to future work. That unsupported equivalence is non-load-bearing: even if it were false, the measured retrieval gains would stand. The more serious concern is statistical, not circular: the post-training EM gains over IC-RALM Reranker are 0.4-1.1 points on 1,000-question sets with no confidence intervals or significance tests, so the 'consistently outperform' claim is not strongly established. Such fragility is a correctness/robustness issue, not circularity, and does not change the verdict that the derivation itself is self-contained.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method depends on standard ML components plus two domain assumptions: the frozen LM's answer likelihood as a reward signal and the completeness of BM25 top-100. The concatenated-query scoring is an ad hoc modeling assumption with no targeted validation.

free parameters (4)
  • k (first-stage top-k) = 5
    Number of first-stage candidate passages; chosen to balance efficiency and performance, no sensitivity analysis reported.
  • Softmax temperature gamma (Eq. 5) = not reported
    Temperature for retriever confidence in the predictive reranking loss; never reported, making training non-reproducible.
  • KL temperature beta (Section 5.3) = not reported
    Temperature for the KL divergence baseline; not reported, so the baseline comparison cannot be reproduced.
  • Balanced positive sampling ratio = not reported
    Proportion of single-passage vs pair positives in training sampling is not specified, yet Figure 2 shows it affects results.
assumptions (4)
  • domain assumption Frozen LM answer likelihood is a reliable reward for passage utility.
    Training objective (Eq. 6-7) assumes PLM(y|[d;x]) correlates with passage usefulness; for short answers many relevant passages give near-zero likelihood, motivating the filtering in Section 3.3.
  • domain assumption BM25 top-100 preresults contain all necessary evidence.
    Section 4.3 restricts retrieval to pre-retrieved top-100; training examples lacking the answer in this set are discarded (Section 3.3), so the system cannot recover evidence BM25 missed.
  • ad hoc to paper Bi-encoder cosine similarity transfers to concatenated queries.
    Eq. (3) extends the scoring function to concatenated queries with no re-training or validation of the embedding space for concatenated inputs.
  • standard math Autoregressive factorization of answer likelihood.
    Eq. (1) factorizes PLM(y|[d;x]) as a product of token probabilities; standard for LM scoring.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Independent Passages: Adaptive Passage Combination Retrieval for Retrieval Augmented Open-Domain Question Answering." pith.science (2026). https://pith.science/paper/44E4CDBC

@misc{pith2026250704069,
  author       = {Pith},
  title        = {Pith review of: Beyond Independent Passages: Adaptive Passage Combination Retrieval for Retrieval Augmented Open-Domain Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/44E4CDBC}},
  note         = {Machine review of arXiv:2507.04069}
}
read the original abstract

Retrieval-augmented generation (RAG) enhances large language models (LLMs) by incorporating external documents at inference time, enabling up-to-date knowledge access without costly retraining. However, conventional RAG methods retrieve passages independently, often leading to redundant, noisy, or insufficiently diverse context-particularly problematic - particularly problematic in noisy corpora and for multi-hop questions. To address this, we propose Adaptive Passage Combination Retrieval (AdaPCR), a novel framework for open-domain question answering with black-box LMs. AdaPCR explicitly models dependencies between passages by considering passage combinations as units for retrieval and reranking. It consists of a context-aware query reformulation using concatenated passages, and a reranking step trained with a predictive objective aligned with downstream answer likelihood. Crucially, AdaPCR adaptively selects the number of retrieved passages without additional stopping modules. Experiments across several QA benchmarks show that AdaPCR outperforms baselines, particularly in multi-hop reasoning, demonstrating the effectiveness of modeling inter-passage dependencies for improved retrieval.

Figures

Figures reproduced from arXiv: 2507.04069 by the authors.

Figure 1
Figure 1. Case study 1: baseline vs. our method when asked the same question "Which band is from England, [PITH_FULL_IMAGE:figures/full_fig_p012_1.png] view at source ↗
Figure 2
Figure 2. Case Study 2 Proof of Lemma 1 PLM(y | x) = Xn i=1 Pret (di | x) PLM (y | [di ; x]) ≤ Xn i=1 Pret (di | x) PLM (y | [dn; x]) = PLM (y | [dn; x])Xn i=1 Pret (di | x) = PLM (y | [dn; x]) · 1 = PLM (y | [dn; x]) Proposition 1 The one-hot retriever is an optimal retriever. Consider Pˆ ret (di | x) = ( 1, if i = n 0, otherwise then Pˆ ret = arg max Pret PLM(y | x) = arg max Pret Xn i=1 Pret (di | x) PLM (y | [di ; x]) Tha… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 17 canonical work pages

  1. [1]

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2024. https://openreview.net/forum?id=hSyW5go0v8 Self- RAG : Learning to retrieve, generate, and critique through self-reflection . In The Twelfth International Conference on Learning Representations

    work page 2024

  2. [2]

    Payal Bajaj, Daniel Campos, Nick Craswell, Li Deng, Jianfeng Gao, Xiaodong Liu, Rangan Majumder, Andrew McNamara, Bhaskar Mitra, Tri Nguyen, Mir Rosenberg, Xia Song, Alina Stoica, Saurabh Tiwary, and Tong Wang. 2018. https://arxiv.org/abs/1611.09268 Ms marco: A human generated machine reading comprehension dataset . Preprint, arXiv:1611.09268

    arXiv 2018

  3. [3]

    Jonathan Berant, Andrew Chou, Roy Frostig, and Percy Liang. 2013. https://www.aclweb.org/anthology/D13-1160 Semantic parsing on F reebase from question-answer pairs . In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1533--1544, Seattle, Washington, USA. Association for Computational Linguistics

    work page 2013

  4. [4]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and 1 others. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877--1901

    2020

  5. [5]

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, and 12 others. 2022. https://doi.org/10.48550/ARXIV.2210.11416 Scaling instruction-fine...

    arXiv doi:10.48550/arxiv.2210.11416 2022

  6. [6]

    Florin Cuconasu, Giovanni Trappolini, Federico Siciliano, Simone Filice, Cesare Campagnano, Yoelle Maarek, Nicola Tonellotto, and Fabrizio Silvestri. 2024. The power of noise: Redefining retrieval for rag systems. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 719--729

    work page 2024

  7. [7]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805

    arXiv 2018

  8. [8]

    Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. 2024. A survey on rag meeting llms: Towards retrieval-augmented large language models. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 6491--6501

    work page 2024

Show all 60 references
  1. [9]

    Michael Fromm, Max Berrendorf, Sandra Obermeier, Thomas Seidl, and Evgeniy Faerman. 2021. Diversity aware relevance learning for argument search. In Advances in Information Retrieval, pages 264--271, Cham. Springer International Publishing

    2021

  2. [10]

    Hang Gao and Yongfeng Zhang. 2024. Vrsd: Rethinking similarity and diversity for retrieval in large language models. arXiv preprint arXiv:2407.04573

    2024 arXiv

  3. [11]

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yixin Dai, Jiawei Sun, Haofen Wang, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2:1

    2023 arXiv

  4. [12]

    Michael Glass, Gaetano Rossiello, Md Faisal Mahbub Chowdhury, Ankita Rajaram Naik, Pengshan Cai, and Alfio Gliozzo. 2022. https://arxiv.org/abs/2207.06300 Re2g: Retrieve, rerank, generate . Preprint, arXiv:2207.06300

    2022 arXiv

  5. [13]

    Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. 2020. https://arxiv.org/abs/2002.08909 Realm: Retrieval-augmented language model pre-training . Preprint, arXiv:2002.08909

    2020 arXiv

  6. [14]

    Eduard Hovy, Laurie Gerber, Ulf Hermjakob, Chin-Yew Lin, and Deepak Ravichandran. 2001. https://www.aclweb.org/anthology/H01-1069 Toward semantics-based answer pinpointing . In Proceedings of the First International Conference on Human Language Technology Research

    2001

  7. [15]

    Jonas Hübotter, Sascha Bongni, Ido Hakimi, and Andreas Krause. 2025. https://arxiv.org/abs/2410.08020 Efficiently learning at test-time: Active fine-tuning of llms . Preprint, arXiv:2410.08020

    2025 arXiv

  8. [16]

    Srinivasan Iyer, Sewon Min, Yashar Mehdad, and Wen-tau Yih. 2021. Reconsider: improved re-ranking using span-focused cross-attention for open domain question answering. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Lin...

    2021

  9. [17]

    Gautier Izacard and Edouard Grave. 2021. https://arxiv.org/abs/2007.01282 Leveraging passage retrieval with generative models for open domain question answering . Preprint, arXiv:2007.01282

    2021 arXiv

  10. [18]

    Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. 2022. https://arxiv.org/abs/2208.03299 Atlas: Few-shot learning with retrieval augmented language models . Preprint, a...

    2022 arXiv

  11. [19]

    Pengcheng Jiang, Jiacheng Lin, Lang Cao, Runchu Tian, SeongKu Kang, Zifeng Wang, Jimeng Sun, and Jiawei Han. 2025. https://arxiv.org/abs/2503.00223 Deepretrieval: Hacking real search engines and retrievers with large language models via reinforcement learning . Preprint, arXiv...

    2025 arXiv

  12. [20]

    Li Jiapeng, Liu Runze, Li Yabo, Zhou Tong, Li Mingling, and Chen Xiang. 2024. https://arxiv.org/abs/2404.14464 Tree of reviews: A tree-based dynamic iterative retrieval framework for multi-hop question answering . Preprint, arXiv:2404.14464

    2024 arXiv

  13. [21]

    Bowen Jin, Jinsung Yoon, Jiawei Han, and Sercan O Arik. 2024. Long-context llms meet rag: Overcoming challenges for long inputs in rag. arXiv preprint arXiv:2410.05983

    2024 arXiv

  14. [22]

    Vladimir Karpukhin, Barlas O g uz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. arXiv preprint arXiv:2004.04906

    2020 arXiv

  15. [23]

    Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. 2019. Generalization through memorization: Nearest neighbor language models. arXiv preprint arXiv:1911.00172

    2019 arXiv

  16. [24]

    Minsang Kim, Cheoneum Park, and Seungjun Baek. 2024. Qpaug: Question and passage augmentation for open-domain question answering of llms. arXiv preprint arXiv:2406.14277

    2024 arXiv

  17. [25]

    Bernhard Kratzwald and Stefan Feuerriegel. 2018. Adaptive document retrieval for deep question answering. arXiv preprint arXiv:1808.06528

    2018 arXiv

  18. [26]

    Benno Kruit, Yiming Xu, and Jan-Christoph Kalo. 2024. Retrieval-based question answering with passage expansion using a knowledge graph. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)...

    2024

  19. [27]

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, and 1 others. 2019. Natural questions: a benchmark for question answering research. Transactions of the Associatio...

    2019

  20. [28]

    Tamera Lanham, Anna Chen, Ansh Radhakrishnan, Benoit Steiner, Carson Denison, Danny Hernandez, Dustin Li, Esin Durmus, Evan Hubinger, Jackson Kernion, and 1 others. 2023. Measuring faithfulness in chain-of-thought reasoning. arXiv preprint arXiv:2307.13702

    2023 arXiv

  21. [29]

    Angeliki Lazaridou, Elena Gribovskaya, Wojciech Stokowiec, and Nikolai Grigorev. 2022. Internet-augmented language models through few-shot prompting for open-domain question answering. arXiv preprint arXiv:2203.05115

    2022 arXiv

  22. [30]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, and 1 others. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Informati...

    2020

  23. [31]

    Xin Li and Dan Roth. 2002. https://www.aclweb.org/anthology/C02-1150 Learning question classifiers . In COLING 2002: The 19th International Conference on Computational Linguistics

    2002

  24. [32]

    Zijian Li, Qingyan Guo, Jiawei Shao, Lei Song, Jiang Bian, Jun Zhang, and Rui Wang. 2024. https://arxiv.org/abs/2406.06572 Graph neural network enhanced retrieval for question answering of llms . Preprint, arXiv:2406.06572

    2024 arXiv

  25. [33]

    Jimmy Lin, Xueguang Ma, Sheng-Chieh Lin, Jheng-Hong Yang, Ronak Pradeep, and Rodrigo Nogueira. 2021. Pyserini: A python toolkit for reproducible information retrieval research with sparse and dense representations. In Proceedings of the 44th International ACM SIGIR Conference ...

    2021

  26. [34]

    Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 12:157--173

    2024

  27. [35]

    Xiaozhong Lyu, Stefan Grafberger, Samantha Biegel, Shaopeng Wei, Meng Cao, Sebastian Schelter, and Ce Zhang. 2023. Improving retrieval-augmented large language models via data importance learning. arXiv preprint arXiv:2307.03027

    2023 arXiv

  28. [36]

    Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Dmytro Okhonko, Samuel Broscheit, Gautier Izacard, Patrick Lewis, Barlas O g uz, Edouard Grave, Wen-tau Yih, and 1 others. 2021. The web is your oyster-knowledge-intensive nlp against a very large web corpus. arXiv preprint...

    2021 arXiv

  29. [37]

    Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. 2023. In-context retrieval-augmented language models. Transactions of the Association for Computational Linguistics, 11:1316--1331

    2023

  30. [38]

    Ruiyang Ren, Yingqi Qu, Jing Liu, Wayne Xin Zhao, Qiaoqiao She, Hua Wu, Haifeng Wang, and Ji-Rong Wen. 2023. https://arxiv.org/abs/2110.07367 Rocketqav2: A joint training method for dense passage retrieval and passage re-ranking . Preprint, arXiv:2110.07367

    2023 arXiv

  31. [39]

    Stephen Robertson and Hugo Zaragoza. 2009. https://doi.org/10.1561/1500000019 The probabilistic relevance framework: Bm25 and beyond . Found. Trends Inf. Retr., 3(4):333–389

    2009 doi

  32. [40]

    Parth Sarthi, Salman Abdullah, Aditi Tuli, Shubh Khanna, Anna Goldie, and Christopher D Manning. 2024. Raptor: Recursive abstractive processing for tree-organized retrieval. In The Twelfth International Conference on Learning Representations

    2024

  33. [41]

    Timo Schick, Jane Dwivedi-Yu, Roberto Dess \` , Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36:68539--68551

    2023

  34. [42]

    Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Rich James, Mike Lewis, Luke Zettlemoyer, and Wen tau Yih. 2023. https://arxiv.org/abs/2301.12652 Replug: Retrieval-augmented black-box language models . Preprint, arXiv:2301.12652

    2023 arXiv

  35. [43]

    Dan Su, Mostofa Patwary, Shrimai Prabhumoye, Peng Xu, Ryan Prenger, Mohammad Shoeybi, Pascale Fung, Anima Anandkumar, and Bryan Catanzaro. 2022. Context generation improves open domain question answering. arXiv preprint arXiv:2210.06349

    2022 arXiv

  36. [44]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, and 1 others. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288

    2023 arXiv

  37. [45]

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2023. https://doi.org/10.18653/v1/2023.acl-long.557 Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions . In Proceedings of the 61st Annual Meeting of th...

    2023 doi

  38. [46]

    Lakshmi Vikraman, Ali Montazeralghaem, Helia Hashemi, W Bruce Croft, and James Allan. 2021. Passage similarity and diversification in non-factoid question answering. In Proceedings of the 2021 acm sigir international conference on theory of information retrieval, pages 271--280

    2021

  39. [47]

    Haoyu Wang, Ruirui Li, Haoming Jiang, Jinjin Tian, Zhengyang Wang, Chen Luo, Xianfeng Tang, Monica Cheng, Tuo Zhao, and Jing Gao. 2024. https://arxiv.org/abs/2402.11129 Blendfilter: Advancing retrieval-augmented large language models via query generation blending and knowledge...

    2024 arXiv

  40. [48]

    Yile Wang, Peng Li, Maosong Sun, and Yang Liu. 2023. Self-knowledge guided retrieval augmentation for large language models. arXiv preprint arXiv:2310.05002

    2023 arXiv

  41. [49]

    Zhiguo Wang, Patrick Ng, Ramesh Nallapati, and Bing Xiang. 2021. Retrieval, re-ranking and multi-task learning for knowledge-base question answering. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pa...

    2021

  42. [50]

    Nuojia Xu, Jun Hu, Quan Fang, Dizhan Xue, Yongxi Li, and Shengsheng Qian. 2024 a . Tri-relational multi-faceted graph neural networks for automatic question tagging. Neurocomputing, 576:127250

    2024

  43. [51]

    Shicheng Xu, Liang Pang, Huawei Shen, Xueqi Cheng, and Tat-Seng Chua. 2024 b . Search-in-the-chain: Interactively enhancing large language models with search for knowledge-intensive tasks. In Proceedings of the ACM Web Conference 2024, pages 1362--1373

    2024

  44. [52]

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan Salakhutdinov, and Christopher D Manning. 2018. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. arXiv preprint arXiv:1809.09600

    2018 arXiv

  45. [53]

    Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Rich James, Jure Leskovec, Percy Liang, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. 2022. Retrieval-augmented multimodal language modeling. arXiv preprint arXiv:2211.12561

    2022 arXiv

  46. [54]

    Ori Yoran, Tomer Wolfson, Ben Bogin, Uri Katz, Daniel Deutch, and Jonathan Berant. 2023 a . Answering questions by meta-reasoning over multiple chains of thought. arXiv preprint arXiv:2304.13007

    2023 arXiv

  47. [55]

    Ori Yoran, Tomer Wolfson, Ori Ram, and Jonathan Berant. 2023 b . Making retrieval-augmented language models robust to irrelevant context. arXiv preprint arXiv:2310.01558

    2023 arXiv

  48. [56]

    Wenhao Yu, Zhihan Zhang, Zhenwen Liang, Meng Jiang, and Ashish Sabharwal. 2023. https://arxiv.org/abs/2305.14002 Improving language models via plug-and-play retrieval feedback . Preprint, arXiv:2305.14002

    2023 arXiv

  49. [57]

    Peitian Zhang, Shitao Xiao, Zheng Liu, Zhicheng Dou, and Jian-Yun Nie. 2023. https://arxiv.org/abs/2310.07554 Retrieve anything to augment large language models . Preprint, arXiv:2310.07554

    2023

  50. [58]

    Zihan Zhang, Meng Fang, and Ling Chen. 2024. Retrievalqa: Assessing adaptive retrieval-augmented generation for short-form open-domain question answering. arXiv preprint arXiv:2402.16457

    2024 arXiv

  51. [59]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  52. [60]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.