Pith. sign in

REVIEW 3 major objections 5 minor 71 references

UniConv: Unifying Retrieval and Response Generation for Large Language Models in Conversations

T0 review · 3 major / 5 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 The paper claims that a single decoder-only LLM jointly fine-tuned for dense retrieval, response generation, and context identification can serve as both the conversational retriever and the RAG responder, and that this joint training…

desk verdict A novel unification idea undermined by train/test contamination on its headline dataset; the core claim of mutual improvement is not supported by the paper's own ablations. read the letter →

arxiv 2507.07030 v1 pith:FFHWH64E submitted 2025-07-09 cs.CL cs.IR

Fengran Mo , Yifan Gao , Chuan Meng , Xin Liu , Zhuofeng Wu , Kelong Mao , Zhengyang Wang , Pei Chen
show 4 more authors
Zheng Li Xian Li Bing Yin Meng Jiang
This is my paper
classification cs.CLcs.IR
keywords conversationalsearchdenseretrievalretrieval-augmentedgenerationunifiedlanguagemodelmulti-turndialogueresponsecontextidentificationinstructionjointfine-tuning
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

The paper argues that conversational search systems do not need separate retriever and generator models. It proposes UniConv, one decoder-only LLM fine-tuned jointly for dense retrieval, response generation, and evidence-context identification, so the same model finds passages and then writes an answer grounded in them. The central claim is mutual improvement: joint training prevents retrieval fine-tuning from destroying generation ability and lets retrieved evidence feed generation more consistently than separate pipelines. If the claim holds, a single 7B model can do the work of two systems at lower serving cost and with more reliable, evidence-grounded answers.

What carries the argument

The load-bearing machinery is joint fine-tuning of a decoder-only LLM with three objectives: the retrieval loss $\mathcal{L}_R$ (an InfoNCE contrastive loss over session-query and passage embeddings), the generation loss $\mathcal{L}_G$ (next-token likelihood with session masking, so generation is conditioned on the session representation $V_{q'_n}$ rather than full attention over all history), and the context identification instruction loss $\mathcal{L}_{CII}$ (a contrastive loss on the input $q'_n \circ p^+_n$ that scores the gold response against negative responses). A fourth component, data discrepancy mitigation, adds conversational search data whose samples carry both a relevant passage and a ground-truth response, so the two objectives see aligned supervision. The CII mechanism is what carries evidence from retrieved passages into generation; ablations show it helps generation and faithfulness while slightly reducing retrieval performance.

What would settle it

Train UniConv with the same mixture minus TopiOCQA, then evaluate on a held-out portion of TopiOCQA and on OR-QuAC and INSCIT; if retrieval NDCG@3 and RAG F1 gains over ChatRetriever and Conv-GRIT shrink or vanish, the claimed mutual improvement depends on seeing the test data during training.

Watch

Extended reading notes

Core claim

UniConv is a single decoder-only LLM initialized from Mistral-7B-chat and fine-tuned with a weighted sum of three losses: a contrastive retrieval loss, a session-masked response generation loss, and a context identification instruction loss. The paper's central claim is that this joint optimization makes retrieval and generation mutually improve: the same model retrieves better than the prior conversational dense retriever on most datasets and, in the RAG setting, generates more accurate and more faithful responses than separate retriever-plus-generator systems, including stronger zero-shot LLMs. The reported numbers on four main datasets (TopiOCQA, QReCC, OR-QuAC, and INSCIT) plus the reliability evaluation on TopiOCQA and FaithDial are the evidence offered for the claim.

Load-bearing premise

The results stand or fall on the assumption that the TopiOCQA test numbers are not inflated by the fact that the whole TopiOCQA dataset was used for fine-tuning; if they are, the mutual-improvement claim rests only on the other datasets.

Editorial extensions

If this is right

  • The same 7B model can act as both first-stage retriever and RAG generator, so deployment and maintenance costs drop to one model instead of two.
  • Joint fine-tuning is a workable answer to the collapse problem: retrieval-only fine-tuning of an LLM destroys generation, but joint objectives preserve and even improve RAG generation.
  • Unified retrieval gives larger RAG gains than using a separate retriever's top passages with a separate generator, which is evidence that consistency between retrieval and generation matters.
  • Context identification instruction is responsible for much of the generation benefit and faithfulness gain, while data discrepancy mitigation lifts both tasks; future two-stage training could recover the small retrieval loss from CII.

Reading between the lines

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

  • A direct test the paper does not run is to retrain without TopiOCQA in the mixture and evaluate on held-out conversational data; because TopiOCQA is in both training and test, the OR-QuAC and INSCIT results are the cleaner evidence for mutual improvement.
  • The CII loss is essentially a contrastive grounding objective, so it could plausibly transfer to single-turn RAG and other knowledge-grounded generation tasks where retrieval and generation are trained separately.
  • The per-turn advantage of the unified model suggests the shared session representation, not just shared parameters, may be doing the work; ablating parameter sharing on one base model would separate these causes.
  • Synthetic well-formatted training data, which the paper mentions as an alternative, would test the data-discrepancy mechanism without touching any test set.
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

3 major / 5 minor

Summary. The paper proposes UniConv, a single Mistral-7B decoder-only model jointly fine-tuned with three objectives: conversational dense retrieval (LR), response generation (LG), and a context-identification instruction loss (LCII), together with a data-discrepancy-mitigation (DDM) step that adds TopiOCQA training data. The model is evaluated for retrieval (NDCG@3, R@10) and generation (F1, zero-shot and RAG) on TopiOCQA, QReCC, OR-QuAC, and INSCIT, and for reliability on FaithDial and TopiOCQA. The central claim is that a unified model can mutually improve both retrieval and generation and outperform separate-model baselines, supported by additional per-turn, history-aware, and reliability analyses.

Significance. If the central claim held, this would be a practically useful contribution: a single 7B LLM acting as both first-stage conversational retriever and RAG generator would reduce deployment cost and allow end-to-end optimization. The paper compares against strong recent baselines such as ChatRetriever, QRACDR, and Conv-GRIT, and it is candid about the zero-shot generation degradation and about the retrieval-side cost of the CII mechanism. The held-out gains on OR-QuAC and INSCIT appear real but modest, and the overall significance is conditional on resolving the TopiOCQA train/test overlap and on reconciling the CII ablation with the mutual-improvement claim. The paper is transparent in its Limitations section, which is a strength, but the headline evidence is currently not reliable enough to support the abstract's claims.

major comments (3)
  1. [Sec. 3.4, Sec. 4.1, Tables 2-6] TopiOCQA is used both as fine-tuning data and as a test set. Section 3.4 states 'we include the conversational search data (Adlakha et al., 2022)' and Section 4.1 lists 'the whole conversational search dataset TopiOCQA for fine-tuning', while Tables 2, 3, 4, 5, 6, and 7 report TopiOCQA as an evaluation set. This is a train/test overlap, so the TopiOCQA columns do not measure generalization. The authors must either train on only the official TopiOCQA train split and re-report all TopiOCQA numbers, or exclude TopiOCQA from evaluation entirely. The DDM ablation in Table 4 is especially confounded, since removing DDM removes a training split rather than isolating a mechanism.
  2. [Sec. 4.5, Tables 4 and 5] The retrieval ablation undermines the mutual-improvement claim: removing CII improves NDCG@3 on all four datasets (45.5 vs 42.6 on TopiOCQA, 49.7 vs 46.6 on QReCC, 47.6 vs 43.5 on OR-QuAC, and 40.0 vs 36.2 on INSCIT), while Table 5 shows CII helps generation. The paper acknowledges this 'interesting phenomenon' but still frames CII as a mechanism that improves consistency and the abstract claims mutual improvement. The claims need to be reformulated to separate generation-side benefits from retrieval-side costs, and the proposed two-stage training alternative should either be tested or CII should be repositioned as a generation-only component.
  3. [Sec. 4.3, Sec. 4.4, Tables 2 and 3] Even setting TopiOCQA aside, the claim that UniConv 'outperforms the existing baselines' is not supported on QReCC: in Table 2, UniConv's retrieval NDCG@3 is 47.6 versus ChatRetriever's 52.5, and in Table 3 the RAG F1 is 26.2 versus 26.3. The paper should qualify the outperformance claim to the datasets and settings where it actually holds. In addition, Table 2 reports t-test significance without describing the number of runs, seeds, or whether significance is computed over instances or model runs; no variance or confidence intervals are reported anywhere, which is load-bearing for the claimed improvements.
minor comments (5)
  1. [Table 4 vs Table 2] The QReCC NDCG@3 for the full UniConv model is reported as 46.6 in Table 4 but 47.6 in Table 2; these numbers need to be reconciled.
  2. [Figure 2] The axis label 'F10' appears to be a typo; it should be R@10 or F1 depending on the metric being plotted.
  3. [Sec. 4.5] The explanation for why CII hurts retrieval is imprecise: LCII is applied to inputs of the form q'_n ∘ p+_n, whereas LR is applied to q'_n, so the interference is via shared parameters rather than a 'changed input query form' for the retrieval objective itself.
  4. [Sec. 4.1 and Table 8] The phrase 'five conversational search datasets' overstates the evaluation: only four datasets have both retrieval and generation results, while FaithDial is used only for the reliability analysis and has no retrieval collection.
  5. [Sec. 3.5 and Sec. 4.1] The loss weight α=0.5, the top-k=10 RAG setting, and the training-data mixture ratio are fixed without sensitivity analysis; given that the Limitations section acknowledges this, a small sweep or at least a table with a few values would strengthen the robustness claims.

Circularity Check

1 steps flagged · score 7.0 of 10

TopiOCQA is both a fine-tuning corpus and a reported test set, so the headline mutual-improvement claim is partly circular; only OR-QuAC and INSCIT remain as clean held-out evidence.

  1. fitted input called prediction [Section 3.4 (Data Discrepancy Mitigation); Section 4.1 (Evaluation Datasets, Training data); Tables 2-5]
    "Thus, we include the conversational search data (Adlakha et al., 2022) to meet this requirement to mitigate the data discrepancy issue. ... We use the ad-hoc search dataset MSMARCO ..., the The Question About the World subset ... and the whole conversational search dataset TopiOCQA for fine-tuning the unified model. ... We conduct the main evaluation on four widely-used conversational search datasets, including TopiOCQA ..."

    The same dataset TopiOCQA is used twice: Section 4.1 names it in the training data as 'the whole conversational search dataset TopiOCQA' and also names it among the evaluation datasets; Tables 2-5 then report TopiOCQA as a test column. Since the model is fine-tuned on the full TopiOCQA, its TopiOCQA scores (e.g., retrieval NDCG@3 42.6 vs ChatRetriever 40.1; RAG F1 29.6 vs 28.3) measure fit to the training input, not held-out prediction. The DDM mechanism is exactly the addition of this same data (Section 3.4), so the 'w/ DDM > w/o DDM' gain on TopiOCQA in Tables 4 and 5 is forced by construction: one model has seen the test set, the other has not. This contaminates the strongest evidence for the central mutual-improvement claim.

full rationale

The only genuinely circular element is the treatment of TopiOCQA as a held-out test set after the full dataset is used for fine-tuning. Section 4.1 explicitly lists 'the whole conversational search dataset TopiOCQA' in the training data and also names TopiOCQA among the evaluation datasets; Tables 2-5 then report TopiOCQA as a test column. The DDM mechanism in Section 3.4 is precisely the addition of this dataset, so the DDM ablation on TopiOCQA compares a model trained on the test set with one not trained on it, making the improvement by construction rather than by the mechanism. This inflates the headline claim that the unified model mutually improves both tasks and outperforms baselines. Once TopiOCQA is removed, the central claim still has independent support on OR-QuAC and INSCIT (e.g., retrieval NDCG@3 43.5 vs 41.9 and 36.2 vs 35.1; RAG F1 17.8 vs 17.3 and 33.2 vs 30.3), so the paper is only partially circular. I do not find any self-citation chain or uniqueness-theorem smuggling that would force the result; the remaining steps are ordinary fine-tuning and evaluation.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The ledger shows a modest methodological contribution: one new loss (LCII), one data-mixture choice, and standard retrieval/generation losses. No invented physical entities. The most consequential items are the ad hoc assumptions: TopiOCQA as training data and the unverified benefit of LCII.

free parameters (3)
  • loss balance alpha = 0.5
    Equation (1) weights LCII by alpha; set to 0.5 by hand with no sensitivity analysis. It directly controls the trade-off between the three objectives.
  • training data mixture ratio = not reported
    The proportions of MS MARCO, UltraChat QAW subset, and TopiOCQA are chosen by hand; the paper notes only a 'fixed ratio' and does not report it.
  • top-k for RAG inference = 10
    At inference, top-10 retrieved passages are fed to generation; k is chosen without ablation.
assumptions (4)
  • standard math InfoNCE contrastive loss and dot-product similarity over the last-token embedding provide valid dense retrieval supervision for a decoder-only LLM.
    Used in Section 3.2.1; standard DPR/RepLLaMA practice.
  • domain assumption Session-masked generation, conditioning only on query representation V_q', preserves generation ability while retrieval fine-tuning is applied.
    Adopted from Mao et al. (2024a) in Section 3.2.2; assumed to prevent catastrophic forgetting.
  • ad hoc to paper The context identification instruction loss LCII reduces inconsistency between retrieval and generation.
    Proposed in Section 3.3; its own ablation (Section 4.5, Table 4) shows the loss hurts retrieval, so the assumption is empirically questionable.
  • ad hoc to paper Adding TopiOCQA as joint training data mitigates data discrepancy without making TopiOCQA results unusable for evaluation.
    Assumed in Sections 3.4 and 4.1; contradicted by standard eval practice since TopiOCQA is also a test set.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UniConv: Unifying Retrieval and Response Generation for Large Language Models in Conversations." pith.science (2026). https://pith.science/paper/FFHWH64E

@misc{pith2026250707030,
  author       = {Pith},
  title        = {Pith review of: UniConv: Unifying Retrieval and Response Generation for Large Language Models in Conversations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FFHWH64E}},
  note         = {Machine review of arXiv:2507.07030}
}
read the original abstract

The rapid advancement of conversational search systems revolutionizes how information is accessed by enabling the multi-turn interaction between the user and the system. Existing conversational search systems are usually built with two different models. This separation restricts the system from leveraging the intrinsic knowledge of the models simultaneously, which cannot ensure the effectiveness of retrieval benefiting the generation. The existing studies for developing unified models cannot fully address the aspects of understanding conversational context, managing retrieval independently, and generating responses. In this paper, we explore how to unify dense retrieval and response generation for large language models in conversation. We conduct joint fine-tuning with different objectives and design two mechanisms to reduce the inconsistency risks while mitigating data discrepancy. The evaluations on five conversational search datasets demonstrate that our unified model can mutually improve both tasks and outperform the existing baselines.

Figures

Figures reproduced from arXiv: 2507.07030 by the authors.

Figure 1
Figure 1. Overview of our UniConv. Three learning ob [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The performance of different systems to fine [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The performance of unified (Uni.) and separated (Sep.) models on dense retrieval (NDCG@3) and [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

71 extracted references · 43 canonical work pages

  1. [1]

    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 INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Vaibhav Adlakha, Shehzaad Dhuliawala, Kaheer Suleman, Harm de Vries, and Siva Reddy. 2022. Topiocqa: Open-domain conversational question answering with topic switching. Transactions of the Association for Computational Linguistics, 10:468--483

    work page 2022

  4. [4]

    Raviteja Anantha, Svitlana Vakulenko, Zhucheng Tu, Shayne Longpre, Stephen Pulman, and Srinivas Chappidi. 2021. Open-domain question answering goes conversational via question rewriting. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 520--534

    2021

  5. [5]

    AnthropicAI. 2023. Introducing claude

    work page 2023

  6. [6]

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2023. Self-rag: Learning to retrieve, generate, and critique through self-reflection. arXiv preprint arXiv:2310.11511

    arXiv 2023

  7. [7]

    Payal Bajaj, Daniel Campos, Nick Craswell, Li Deng, Jianfeng Gao, Xiaodong Liu, Rangan Majumder, Andrew McNamara, Bhaskar Mitra, Tri Nguyen, et al. 2016. Ms marco: A human generated machine reading comprehension dataset. arXiv preprint arXiv:1611.09268

    arXiv 2016

  8. [8]

    Yiruo Cheng, Kelong Mao, Ziliang Zhao, Guanting Dong, Hongjin Qian, Yongkang Wu, Tetsuya Sakai, Ji-Rong Wen, and Zhicheng Dou. 2024. Coral: Benchmarking multi-turn conversational retrieval-augmentation generation. arXiv preprint arXiv:2410.23090

    arXiv 2024

Show all 71 references
  1. [9]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171--4186

    2019

  2. [10]

    Ning Ding, Yulin Chen, Bokai Xu, Yujia Qin, Shengding Hu, Zhiyuan Liu, Maosong Sun, and Bowen Zhou. 2023. Enhancing chat language models by scaling high-quality instructional conversations. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processi...

    2023

  3. [11]

    Nouha Dziri, Ehsan Kamalloo, Sivan Milton, Osmar Zaiane, Mo Yu, Edoardo M Ponti, and Siva Reddy. 2022. Faithdial: A faithful benchmark for information-seeking dialogue. Transactions of the Association for Computational Linguistics, 10:1473--1490

    2022

  4. [12]

    Ahmed Elgohary, Denis Peskov, and Jordan Boyd-Graber. 2019. Can you unpack that? learning to rewrite questions-in-context. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language P...

    2019

  5. [13]

    Jianfeng Gao, Chenyan Xiong, Paul Bennett, and Nick Craswell. 2022. Neural approaches to conversational information retrieval. arXiv preprint arXiv:2201.05176

    2022 arXiv

  6. [14]

    Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations

    2022

  7. [15]

    Yunah Jang, Kang-il Lee, Hyunkyung Bae, Seungpil Won, Hwanhee Lee, and Kyomin Jung. 2023. Itercqr: Iterative conversational query reformulation without human supervision. arXiv preprint arXiv:2311.09820

    2023 arXiv

  8. [16]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de Las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, L \' e lio Renard Lavaud, Marie - Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, ...

    2023 doi arXiv

  9. [17]

    Zhuoran Jin, Pengfei Cao, Yubo Chen, Kang Liu, and Jun Zhao. 2023. Instructor: Instructing unsupervised conversational dense retrieval with large language models. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 6649--6675

    2023

  10. [18]

    Jeff Johnson, Matthijs Douze, and Herv \'e J \'e gou. 2019. Billion-scale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535--547

    2019

  11. [19]

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361

    2020 arXiv

  12. [20]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)...

    2020

  13. [21]

    Sungdong Kim and Gangwoo Kim. 2022. Saving dense retriever from shortcut dependency in conversational search. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10278--10287. Association for Computational Linguistics

    2022

  14. [22]

    Yilong Lai, Jialong Wu, Congzhi Zhang, Haowen Sun, and Deyu Zhou. 2024. Adacqr: Enhancing query reformulation for conversational search via sparse and dense retrieval alignment. In COLING

    2024

  15. [23]

    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, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Proc...

    2020

  16. [24]

    Ming Li, Pei Chen, Chenguang Wang, Hongyu Zhao, Yijun Liang, Yupeng Hou, Fuxiao Liu, and Tianyi Zhou. 2024 a . Mosaic-it: Free compositional data augmentation improves instruction tuning. arXiv preprint arXiv:2405.13326

    2024 arXiv

  17. [25]

    Xiaoxi Li, Yujia Zhou, and Zhicheng Dou. 2024 b . Unigen: A unified generative framework for retrieval and question answering with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 8688--8696

    2024

  18. [26]

    Sheng-Chieh Lin, Jheng-Hong Yang, and Jimmy Lin. 2021. Contextualized query embeddings for conversational search. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1004--1015

    2021

  19. [27]

    Zihan Liu, Wei Ping, Rajarshi Roy, Peng Xu, Mohammad Shoeybi, and Bryan Catanzaro. 2024. Chatqa: Building gpt-4 level conversational qa models. arXiv preprint arXiv:2401.10225

    2024 arXiv

  20. [28]

    Simon Lupart, Mohammad Aliannejadi, and Evangelos Kanoulas. 2025. Disco meets llms: A unified approach for sparse retrieval and contextual distillation in conversational search. In SIGIR

    2025

  21. [29]

    Xueguang Ma, Liang Wang, Nan Yang, Furu Wei, and Jimmy Lin. 2024. Fine-tuning llama for multi-stage text retrieval. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2421--2425

    2024

  22. [30]

    Kelong Mao, Chenlong Deng, Haonan Chen, Fengran Mo, Zheng Liu, Tetsuya Sakai, and Zhicheng Dou. 2024 a . Chatretriever: Adapting large language models for generalized and robust conversational dense retrieval. In Proceedings of the 2024 Conference on Empirical Methods in Natur...

    2024

  23. [31]

    Kelong Mao, Zhicheng Dou, Haonan Chen, Fengran Mo, and Hongjin Qian. 2023 a . Large language models know your contextual search intent: A prompting framework for conversational search. In Findings of the Association for Computational Linguistics: EMNLP 2023

    2023

  24. [32]

    Kelong Mao, Zhicheng Dou, Bang Liu, Hongjin Qian, Fengran Mo, Xiangli Wu, Xiaohua Cheng, and Zhao Cao. 2023 b . Search-oriented conversational query editing. In Findings of the Association for Computational Linguistics: ACL 2023, pages 4160--4172

    2023

  25. [33]

    Kelong Mao, Zhicheng Dou, Hongjin Qian, Fengran Mo, Xiaohua Cheng, and Zhao Cao. 2022. Convtrans: Transforming web search sessions for conversational dense retrieval. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 2935--2946

    2022

  26. [34]

    Kelong Mao, Zheng Liu, Hongjin Qian, Fengran Mo, Chenlong Deng, and Zhicheng Dou. 2024 b . Rag-studio: Towards in-domain adaptation of retrieval augmented generation through self-alignment. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 725--735

    2024

  27. [35]

    Kelong Mao, Hongjin Qian, Fengran Mo, Zhicheng Dou, Bang Liu, Xiaohua Cheng, and Zhao Cao. 2023 c . Learning denoised and interpretable session representation for conversational search. In Proceedings of the ACM Web Conference 2023, pages 3193--3202

    2023

  28. [36]

    Chuan Meng, Pengjie Ren, Zhumin Chen, Christof Monz, Jun Ma, and Maarten de Rijke. 2020 a . RefNet : A reference-aware network for background based conversation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 8496--8503

    2020

  29. [37]

    Chuan Meng, Pengjie Ren, Zhumin Chen, Zhaochun Ren, Tengxiao Xi, and Maarten de Rijke. 2021. Initiative-aware self-supervised learning for knowledge-grounded conversations. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information...

    2021

  30. [38]

    Chuan Meng, Pengjie Ren, Zhumin Chen, Weiwei Sun, Zhaochun Ren, Zhaopeng Tu, and Maarten de Rijke. 2020 b . DukeNet : A dual knowledge interaction network for knowledge-grounded conversation. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Develop...

    2020

  31. [39]

    Chuan Meng, Francesco Tonolini, Fengran Mo, Nikolaos Aletras, Emine Yilmaz, and Gabriella Kazai. 2025. Bridging the gap: From ad-hoc to proactive search in conversations. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval

    2025

  32. [40]

    Fengran Mo, Abbas Ghaddar, Kelong Mao, Mehdi Rezagholizadeh, Boxing Chen, Qun Liu, and Jian-Yun Nie. 2024 a . Chiq: Contextual history enhancement for improving query rewriting in conversational search. In Proceedings of the 2024 Conference on Empirical Methods in Natural Lang...

    2024

  33. [41]

    Fengran Mo, Kelong Mao, Ziliang Zhao, Hongjin Qian, Haonan Chen, Yiruo Cheng, Xiaoxi Li, Yutao Zhu, Zhicheng Dou, and Jian-Yun Nie. 2024 b . A survey of conversational search. arXiv preprint arXiv:2410.15576

    2024 arXiv

  34. [42]

    Fengran Mo, Kelong Mao, Yutao Zhu, Yihong Wu, Kaiyu Huang, and Jian-Yun Nie. 2023 a . Convgqr: Generative query reformulation for conversational search. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4998--5012

    2023

  35. [43]

    Fengran Mo, Chuan Meng, Mohammad Aliannejadi, and Jian-Yun Nie. 2025. Conversational search: From fundamentals to frontiers in the llm era. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval

    2025

  36. [44]

    Fengran Mo, Jian-Yun Nie, Kaiyu Huang, Kelong Mao, Yutao Zhu, Peng Li, and Yang Liu. 2023 b . Learning to relate to previous turns in conversational search. In 29th ACM SIGKDD Conference On Knowledge Discover and Data Mining (SIGKDD)

    2023

  37. [45]

    Fengran Mo, Chen Qu, Kelong Mao, Yihong Wu, Zhan Su, Kaiyu Huang, and Jian-Yun Nie. 2024 c . Aligning query representation with rewritten query and relevance judgments in conversational search. In Proceedings of the 33rd ACM International Conference on Information and Knowledg...

    2024

  38. [46]

    Fengran Mo, Chen Qu, Kelong Mao, Tianyu Zhu, Zhan Su, Kaiyu Huang, and Jian-Yun Nie. 2024 d . History-aware conversational dense retrieval. In Findings of the Association for Computational Linguistics: ACL 2024, pages 13366--13378

    2024

  39. [47]

    Fengran Mo, Bole Yi, Kelong Mao, Chen Qu, Kaiyu Huang, and Jian-Yun Nie. 2024 e . Convsdg: Session data generation for conversational search. In Companion Proceedings of the ACM on Web Conference 2024, pages 1634--1642

    2024

  40. [48]

    Fengran Mo, Longxiang Zhao, Kaiyu Huang, Yue Dong, Degen Huang, and Jian-Yun Nie. 2024 f . How to leverage personal textual knowledge for personalized conversational information retrieval. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Man...

    2024

  41. [49]

    Niklas Muennighoff, Hongjin Su, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Amanpreet Singh, and Douwe Kiela. 2024. Generative representational instruction tuning. arXiv preprint arXiv:2402.09906

    2024 arXiv

  42. [50]

    Zhao, Yi Luan, Keith B

    Jianmo Ni, Chen Qu, Jing Lu, Zhuyun Dai, Gustavo Hern \' a ndez \' A brego, Ji Ma, Vincent Y. Zhao, Yi Luan, Keith B. Hall, Ming - Wei Chang, and Yinfei Yang. 2022. https://doi.org/10.18653/V1/2022.EMNLP-MAIN.669 Large dual encoders are generalizable retrievers . In Proceeding...

    2022 doi

  43. [51]

    https://platform.openai.com/docs/models/gpt-3-5-turbo

    OpenAI. https://platform.openai.com/docs/models/gpt-3-5-turbo

  44. [52]

    Zhenyu Pan, Haozheng Luo, Manling Li, and Han Liu. 2024. Conv-coa: Improving open-domain question answering in large language models via conversational chain-of-action. arXiv preprint arXiv:2405.17822

    2024 arXiv

  45. [53]

    Chen Qu, Liu Yang, Cen Chen, Minghui Qiu, W Bruce Croft, and Mohit Iyyer. 2020. Open-retrieval conversational question answering. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages 539--548

    2020

  46. [54]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1--67

    2020

  47. [55]

    Pengjie Ren, Zhumin Chen, Zhaochun Ren, Evangelos Kanoulas, Christof Monz, and Maarten De Rijke. 2021. Conversations with search engines: Serp-based conversational response generation. ACM Transactions on Information Systems (TOIS), 39(4):1--29

    2021

  48. [56]

    Alireza Salemi and Hamed Zamani. 2024. Evaluating retrieval quality in retrieval-augmented generation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2395--2400

    2024

  49. [57]

    I Sutskever. 2014. Sequence to sequence learning with neural networks. arXiv preprint arXiv:1409.3215

    2014 arXiv

  50. [58]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton - Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu,...

    2023 doi arXiv

  51. [59]

    Nikos Voskarides, Dan Li, Pengjie Ren, Evangelos Kanoulas, and Maarten de Rijke. 2020. Query resolution for conversational search with limited supervision. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages...

    2020

  52. [60]

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. https://doi.org/10.48550/ARXIV.2401.00368 Improving text embeddings with large language models . CoRR, abs/2401.00368

    2024 doi arXiv

  53. [61]

    Zeqiu Wu, Yi Luan, Hannah Rashkin, David Reitter, and Gaurav Singh Tomar. 2022. Conqrr: Conversational query rewriting for retrieval with reinforcement learning. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP)

    2022

  54. [62]

    Zeqiu Wu, Ryu Parish, Hao Cheng, Sewon Min, Prithviraj Ammanabrolu, Mari Ostendorf, and Hannaneh Hajishirzi. 2023. Inscit: Information-seeking conversations with mixed-initiative interactions. Transactions of the Association for Computational Linguistics, 11:453--468

    2023

  55. [63]

    Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N Bennett, Junaid Ahmed, and Arnold Overwijk. 2020. Approximate nearest neighbor negative contrastive learning for dense text retrieval. In International Conference on Learning Representations

    2020

  56. [64]

    Fanghua Ye, Meng Fang, Shenghui Li, and Emine Yilmaz. 2023. Enhancing conversational search: Large language model-aided informative query rewriting. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 5985--6006

    2023

  57. [65]

    Linhao Ye, Zhikai Lei, Jianghao Yin, Qin Chen, Jie Zhou, and Liang He. 2024. Boosting conversational question answering with fine-grained retrieval-augmentation and self-check. arXiv preprint arXiv:2403.18243

    2024 arXiv

  58. [66]

    Chanwoong Yoon, Gangwoo Kim, Byeongguk Jeon, Sungdong Kim, Yohan Jo, and Jaewoo Kang. 2024. Ask optimal questions: Aligning large language models with retriever's preference in conversational search. arXiv preprint arXiv:2402.11827

    2024 arXiv

  59. [67]

    Shi Yu, Zhenghao Liu, Chenyan Xiong, Tao Feng, and Zhiyuan Liu. 2021. Few-shot conversational dense retrieval. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 829--838

    2021

  60. [68]

    Yue Yu, Wei Ping, Zihan Liu, Boxin Wang, Jiaxuan You, Chao Zhang, Mohammad Shoeybi, and Bryan Catanzaro. 2024. Rankrag: Unifying context ranking with retrieval-augmented generation in llms. arXiv preprint arXiv:2407.02485

    2024 arXiv

  61. [69]

    Hamed Zamani, Johanne R Trippas, Jeff Dalton, Filip Radlinski, et al. 2023. Conversational information seeking. Foundations and Trends in Information Retrieval , 17(3-4):244--456

    2023

  62. [70]

    Jinghan Zhang, Xiting Wang, Fengran Mo, Yeyang Zhou, Wanfu Gao, and Kunpeng Liu. 2025. Entropy-based exploration conduction for multi-step reasoning. arXiv preprint arXiv:2503.15848

    2025 arXiv

  63. [71]

    Jintian Zhang, Cheng Peng, Mengshu Sun, Xiang Chen, Lei Liang, Zhiqiang Zhang, Jun Zhou, Huajun Chen, and Ningyu Zhang. 2024. Onegen: Efficient one-pass unified generation and retrieval for llms. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 4...

    2024

Pith tools

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