Pith. sign in

REVIEW 2 major objections 4 minor 2 cited by

Dynamic and Parametric Retrieval-Augmented Generation

T0 review · 2 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This tutorial argues that the next RAG systems will retrieve on the fly during generation and inject retrieved knowledge into model parameters rather than into the prompt.

desk verdict A useful tutorial survey of Dynamic and Parametric RAG that overstates the case for parametric injection; the 'more effective' claim is asserted, not demonstrated. read the letter →

arxiv 2506.06704 v1 pith:AYCRFXG3 submitted 2025-06-07 cs.CL cs.IR

classification cs.CLcs.IR
keywords Retrieval-AugmentedGenerationDynamicRAGParametricknowledgeinjectionLoRAadaptersmulti-hopreasoningLLMuncertaintyestimationhypernetwork
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 tutorial makes the case that the standard retrieve-then-generate RAG pipeline—retrieve once, append documents to the prompt, then generate—is the main bottleneck for knowledge-intensive LLM applications. It identifies two complementary fixes: Dynamic RAG, which decides during generation when and what to retrieve based on the model's own evolving state, and Parametric RAG, which converts each retrieved document into plug-in parameters instead of context text. The paper surveys representative systems in each line and presents them as a single research agenda for the field. If the framing holds, future RAG systems will treat retrieval as an on-demand, generation-time operation and external knowledge as something to be written into the model's weights rather than pasted into the input.

What carries the argument

The argument turns on two mechanisms. The first is the document-to-parameter transformation, a mapping from a retrieved document D to a plug-in parameter module P whose insertion into the model internalizes the knowledge in D; it is implemented either by offline fine-tuning of document-specific LoRA modules or by an online hypernetwork that generates parameters from semantic embeddings. The second is the generation-time trigger, a set of signals—reflection tokens, low-confidence token probabilities, attention and entropy patterns, or self-aware token-level uncertainty—that decide when retrieval should happen and what query should be issued. The conceptual load-bearing distinction is that in-context knowledge reaches the model only through attention key-value computation, whereas factual knowledge is encoded in feed-forward network parameters, so injecting knowledge into parameters is claimed to make external knowledge behave like the model's own memory rather than like extra text it must attend to.

What would settle it

Run PRAG or DyPRAG against a strong in-context RAG baseline with the same base model and retriever on knowledge-intensive question answering; if prompting with the retrieved passages matches or beats parameter injection on accuracy, robustness, and latency, the paradigm-shift claim is undercut. A second decisive check is a conflict test in which the prompt supplies a fact that contradicts the model's parametric memory, measuring whether parameter-injected knowledge shifts answers more reliably than in-context knowledge does.

Watch

Extended reading notes

Core claim

The central claim is that RAG must move beyond static retrieval and in-context knowledge injection. Dynamic RAG accomplishes the first move by interleaving retrieval with generation: reflection tokens, token-level confidence scores, attention distributions, and self-aware uncertainty estimates trigger retrieval on demand, and the LLM or an external module formulates queries from the current context. Parametric RAG accomplishes the second move by introducing a document-to-parameter transformation that maps a document D into a plug-in parameter module P; once P is inserted into the LLM, the model can use the external knowledge the same way it uses its internal parametric knowledge. The paper distinguishes two implementations: training document-specific LoRA adapters on synthetic QA data, as in PRAG, and generating parameters on the fly from document embeddings via a hypernetwork, as in DyPRAG. It concludes that Parametric RAG represents a paradigm shift from context-level to parameter-level knowledge injection and provides a more effective way to integrate external information into LLMs.

Load-bearing premise

The load-bearing premise is that LLMs use in-context information more weakly than their internal parametric knowledge, so writing retrieved knowledge into the weights is inherently more effective than putting it in the prompt; the tutorial relies on cited evidence for this rather than on its own controlled measurement.

Editorial extensions

If this is right

  • Systems that decide when to retrieve during generation will outperform one-shot retrieval on multi-hop reasoning and long-form generation, because the model's information needs evolve as it generates.
  • Retrieved knowledge can be packaged as plug-in parameter modules, allowing documents to be injected without repeatedly paying the computational cost of long contexts.
  • Hypernetwork-generated parameter modules make parametric injection scalable at test time, removing the need to fine-tune and store a separate adapter for every document.
  • Dynamic RAG and Parametric RAG are complementary halves of one agenda: one decides when and what to fetch, the other decides how to absorb what is fetched.
  • Making external knowledge parametric should improve factual grounding over grounded-in-context approaches, especially when retrieved content conflicts with what the model already believes.

Reading between the lines

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

  • Editorial inference: if parametric injection is as effective as the tutorial claims, the document-to-parameter transformation could be applied to knowledge bases as a whole, turning retrieval-augmented generation into a form of continual memory update rather than per-query context stuffing.
  • Editorial inference: the tutorial's own premise suggests a direct comparative test it does not report—measure answer shifts when the prompt and the parameters disagree, to quantify whether parametric injection really grounds outputs more reliably than in-context injection.
  • Editorial inference: the uncertainty signals used for dynamic retrieval (token-level confidence, attention, predictive entropy) are likely reusable for hallucination detection and query rewriting, so the dynamic-retrieval mechanism may transfer to neighboring LLM reliability problems.
  • Editorial inference: a natural synthesis the tutorial does not develop is a fully test-time system in which a dynamic trigger calls a hypernetwork to generate parameters on the fly, combining when-to-retrieve with how-to-inject without offline per-document training.
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

2 major / 4 minor

Summary. This manuscript is a tutorial proposal for SIGIR 2025, aimed at introducing two emerging RAG paradigms: Dynamic RAG, which adaptively decides when and what to retrieve during generation, and Parametric RAG, which injects retrieved knowledge directly into model parameters rather than into the prompt. The paper motivates these paradigms by arguing that standard RAG suffers from static one-shot retrieval and suboptimal in-context knowledge injection. It describes representative methods (Self-RAG, FLARE, DRAGIN, SeaKR for Dynamic RAG; PRAG and DyPRAG for Parametric RAG), provides a lecture schedule, and outlines supporting materials. The central scientific assertion is that parameter-level injection is a 'more effective solution' than in-context injection (§1.3), which the tutorial frames as a paradigm shift.

Significance. If the framing is correct, this tutorial addresses a genuine gap: prior SIGIR and ACL tutorials covered standard RAG but did not synthesize the rapidly growing literature on dynamic retrieval timing and parametric knowledge injection. The tutorial's organization of these works into two complementary axes is timely and potentially useful for IR researchers. The authors are well placed academically, and the planned supporting materials (slides, manuscript, GitHub repository) could provide practical value. However, the significance is dampened by the manuscript's unsupported comparative claim that Parametric RAG is inherently more effective than in-context injection. Because this claim is load-bearing for the tutorial's thesis, the current version is better viewed as a proposal with an unverified motivating premise rather than a settled survey.

major comments (2)
  1. [Section 1.3 and Abstract] The claim that Parametric RAG is "a more effective solution for integrating external information into LLMs" is stated as established fact, but this manuscript presents no direct experimental comparison between parameter-level injection and in-context injection. The works cited in §1.1 ([20,25,53]) support only that long contexts degrade attention and that factual knowledge is localized in feed-forward parameters; they do not show that LoRA-style parameter injection outperforms prompting with the same passages. The tutorial should either report the quantitative comparisons from PRAG [38] and DyPRAG [42] or explicitly label the effectiveness claim as a motivating hypothesis and an open research question.
  2. [Section 1.1] The logical step from "LLMs encode factual knowledge in FFN parameters" to "in-context knowledge use is weaker and less reliable" is an unsupported inference. The sentence "This fundamental difference ... leads to a weaker and less reliable use of external knowledge" is presented categorically, but none of the cited references directly compares the reliability of parameter-injected versus prompt-injected knowledge. This premise underpins the entire Parametric RAG section, so the tutorial should either soften the wording to indicate that this is a hypothesis (e.g., "may lead to") or cite a head-to-head empirical study that substantiates the superiority claim.
minor comments (4)
  1. [Section 1.1, References] Reference [9] (AlphaEdit) is a knowledge-editing paper, not a primary study of factual knowledge localization in feed-forward networks; the tutorial would be better served by citing the original interpretability work on factual recall (e.g., Geva et al.) rather than a pre-print that builds on it.
  2. [Section 1.2] The verbal descriptions of Self-RAG, FLARE, DRAGIN, and SeaKR are clear but hard to retain without a side-by-side comparison; adding a small table summarizing when-retrieval-triggering mechanisms, query formulation, training requirements, and representative results would improve the tutorial's pedagogical value.
  3. [Section 4] The schedule lists two separate Q&A sessions without labels; renaming them "Mid-tutorial Q&A" and "Final Q&A" would eliminate confusion for attendees planning around the break.
  4. [References] Reference [27] is a LessWrong blog post, which is non-archival and may not meet the citation standards of a SIGIR tutorial; consider replacing it with a peer-reviewed source on neuron-level factual recall or explicitly flagging it as a non-archival resource.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: this tutorial proposal contains no derivations or predictions that reduce to its own inputs, and its two categories are anchored by independent prior systems despite prominent self-citations.

full rationale

This document is a SIGIR tutorial proposal rather than a research paper with a derivation chain. It contains no equations, no fitted parameters, and no predicted quantities that could reduce by construction to its inputs. The central claim that parameter-level knowledge injection is more effective than in-context injection (§1.3) rests on an empirical premise in §1.1: that LLMs encode factual knowledge in their parameters and therefore use in-context information more weakly. That premise is supported by external citations [9, 20, 25, 27, 53] and is not directly measured in this manuscript. This is an evidentiary or framing weakness, not circularity: the cited studies are independent and do not define the tutorial's conclusion into existence. The two tutorial categories are also not reducible to the authors' own work. Dynamic RAG is anchored by Self-RAG [3], FLARE [18], and SeaKR [51] as well as the authors' DRAGIN [37]; Parametric RAG is anchored by both the authors' PRAG [38] and the independent DyPRAG [42]. No uniqueness theorem, ansatz, or load-bearing conclusion is imported solely through self-citation. The unsupported inference in §1.1 should be weighed as a correctness risk, but it does not meet the standard for circularity. Therefore the circularity score is 0.

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

The ledger is light because the paper is a tutorial, not a derivation. It rests on three domain and framing assumptions: internal parametric knowledge is the primary locus of LLM knowledge, static retrieval is insufficient for complex generation, and Dynamic and Parametric RAG are the key complementary directions. No free parameters or invented entities appear.

assumptions (3)
  • domain assumption In-context knowledge injection is less effective and less reliable than parameter-level integration because LLMs primarily encode factual knowledge in parameters (Section 1.1).
    This is invoked to motivate Parametric RAG, citing works such as [9, 27, 53]; the manuscript itself provides no direct measurement.
  • domain assumption Static one-shot retrieval fails for multi-hop reasoning because information needs evolve during generation (Section 1.1).
    Motivates Dynamic RAG; based on cited empirical results, not on experiments in this tutorial.
  • ad hoc to paper Dynamic RAG and Parametric RAG are the two representative and complementary new paradigms for RAG (Sections 1.2 and 1.3).
    This taxonomy is the tutorial's organizational premise and is not derived from a systematic survey or evidence; it reflects the authors' own categorization of the literature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dynamic and Parametric Retrieval-Augmented Generation." pith.science (2026). https://pith.science/paper/AYCRFXG3

@misc{pith2026250606704,
  author       = {Pith},
  title        = {Pith review of: Dynamic and Parametric Retrieval-Augmented Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AYCRFXG3}},
  note         = {Machine review of arXiv:2506.06704}
}
read the original abstract

Retrieval-Augmented Generation (RAG) has become a foundational paradigm for equipping large language models (LLMs) with external knowledge, playing a critical role in information retrieval and knowledge-intensive applications. However, conventional RAG systems typically adopt a static retrieve-then-generate pipeline and rely on in-context knowledge injection, which can be suboptimal for complex tasks that require multihop reasoning, adaptive information access, and deeper integration of external knowledge. Motivated by these limitations, the research community has moved beyond static retrieval and in-context knowledge injection. Among the emerging directions, this tutorial delves into two rapidly growing and complementary research areas on RAG: Dynamic RAG and Parametric RAG. Dynamic RAG adaptively determines when and what to retrieve during the LLM's generation process, enabling real-time adaptation to the LLM's evolving information needs. Parametric RAG rethinks how retrieved knowledge should be injected into LLMs, transitioning from input-level to parameter-level knowledge injection for enhanced efficiency and effectiveness. This tutorial offers a comprehensive overview of recent advances in these emerging research areas. It also shares theoretical foundations and practical insights to support and inspire further research in RAG.

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. MetaSyn: A Benchmark for LLM Agents on Meta-Analysis Articles from Nature Portfolio

    cs.CL 2026-06 unverdicted novelty 8.0 of 10

    MetaSyn is a stage-level benchmark of 442 meta-analyses showing LLM agents retrieve up to 90.9% of eligible studies but include at most 52.7% in their final reports.

  2. RWGBench: Evaluating Scholarly Positioning in Related Work Generation

    cs.DL 2026-05 unverdicted novelty 7.0 of 10

    RWGBench measures related-work generation by citation choices, and shows citation-focused metrics expose failures that text-similarity and LLM-judge scores miss.

Reference graph

Works this paper leans on

59 extracted references · 15 canonical work pages · cited by 2 Pith papers

  1. [20]

    Mosh Levy, Alon Jacoby, and Yoav Goldberg. 2024. Same task, more tokens: the impact of input length on the reasoning performance of large language models. arXiv preprint arXiv:2402.14848(2024)

  2. [25]

    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 (2024), 157–173

  3. [38]

    Weihang Su, Yichen Tang, Qingyao Ai, Junxi Yan, Changyue Wang, Hongning Wang, Ziyi Ye, Yujia Zhou, and Yiqun Liu. 2025. Parametric Retrieval-Augmented Generation.arXiv preprint arXiv:2501.15915(2025)

  4. [42]

    Yuqiao Tan, Shizhu He, Huanxuan Liao, Jun Zhao, and Kang Liu. 2025. Better wit than wealth: Dynamic Parametric Retrieval Augmented Generation for Test-time Knowledge Enhancement.arXiv preprint arXiv:2503.23895(2025)

  5. [1]

    Qingyao Ai. 2025. Homepage of Qingyao Ai. https://qingyaoai.github.io/. Ac- cessed: 2025-03-28

  6. [2]

    Akari Asai, Sewon Min, Zexuan Zhong, and Danqi Chen. 2023. Retrieval-based language models and applications. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 6: Tutorial Abstracts). 41–46

  7. [3]

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi

  8. [4]

    Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Ruther- ford, Katie Millican, George Bm Van Den Driessche, Jean-Baptiste Lespiau, Bog- dan Damoc, Aidan Clark, et al. 2022. Improving language models by retrieving from trillions of tokens. InInternational conference on machine learning. PMLR, 2206–2240

Show all 59 references
  1. [5]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners.Advances in neural information processing systems33 (2020), 1877–1901

  2. [6]

    Deng Cai, Yan Wang, Lemao Liu, and Shuming Shi. 2022. Recent advances in retrieval-augmented text generation. InProceedings of the 45th international ACM SIGIR conference on research and development in information retrieval. 3417–3419

  3. [7]

    Xuesong Chen, Ziyi Ye, Xiaohui Xie, Yiqun Liu, Xiaorong Gao, Weihang Su, Shuqi Zhu, Yike Sun, Min Zhang, and Shaoping Ma. 2022. Web search via an efficient and effective brain-machine interface. InProceedings of the fifteenth ACM international conference on web search and data...

  4. [8]

    Qian Dong, Qingyao Ai, Hongning Wang, Yiding Liu, Haitao Li, Weihang Su, Yiqun Liu, Tat-Seng Chua, and Shaoping Ma. 2025. Decoupling Knowledge and Context: An Efficient and Effective Retrieval Augmented Generation Framework via Cross Attention. InProceedings of the ACM on Web ...

  5. [9]

    Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Shi Jie, Xiang Wang, Xiangnan He, and Tat-Seng Chua. 2024. Alphaedit: Null-space constrained knowledge editing for language models.arXiv preprint arXiv:2410.02355(2024)

  6. [10]

    Yan Fang, Jingtao Zhan, Qingyao Ai, Jiaxin Mao, Weihang Su, Jia Chen, and Yiqun Liu. 2024. Scaling laws for dense retrieval. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1339–1349

  7. [11]

    Luyu Gao and Jamie Callan. 2021. Condenser: a pre-training architecture for dense retrieval.arXiv preprint arXiv:2104.08253(2021)

  8. [12]

    Suchin Gururangan, Ana Marasović, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A Smith. 2020. Don’t stop pretraining: Adapt language models to domains and tasks.arXiv preprint arXiv:2004.10964(2020)

  9. [13]

    Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. 2020. Retrieval augmented language model pre-training. InInternational conference on machine learning. PMLR, 3929–3938

  10. [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. InInternational Conference on Learning Representations

  11. [15]

    Gautier Izacard and Edouard Grave. 2020. Leveraging passage retrieval with generative models for open domain question answering.arXiv preprint arXiv:2007.01282(2020)

  12. [16]

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of hallucination in natural language generation.Comput. Surveys55, 12 (2023), 1–38

  13. [17]

    Zhengbao Jiang, Luyu Gao, Jun Araki, Haibo Ding, Zhiruo Wang, Jamie Callan, and Graham Neubig. 2022. Retrieval as attention: End-to-end learning of retrieval and reading within a single transformer.arXiv preprint arXiv:2212.02027(2022)

  14. [18]

    Zhengbao Jiang, Frank F Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie Callan, and Graham Neubig. 2023. Active retrieval augmented generation.arXiv preprint arXiv:2305.06983(2023)

  15. [19]

    Vladimir Karpukhin, Barlas Oğuz, 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)

  16. [21]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing S...

  17. [22]

    Haitao Li, Jia Chen, Weihang Su, Qingyao Ai, and Yiqun Liu. 2023. Towards better web search performance: pre-training, fine-tuning and learning to rank. arXiv preprint arXiv:2303.04710(2023)

  18. [23]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Cheng- gang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437(2024). 4

  19. [24]

    Huanshuo Liu, Hao Zhang, Zhijiang Guo, Kuicai Dong, Xiangyang Li, Yi Quan Lee, Cong Zhang, and Yong Liu. 2024. CtrlA: Adaptive Retrieval-Augmented Generation via Probe-Guided Control.arXiv preprint arXiv:2405.18727(2024)

  20. [26]

    Yixiao Ma, Yueyue Wu, Weihang Su, Qingyao Ai, and Yiqun Liu. 2023. CaseEn- coder: A Knowledge-enhanced Pre-trained Model for Legal Case Encoding.arXiv preprint arXiv:2305.05393(2023)

  21. [27]

    Neel Nanda, Senthooran Rajamanoharan, János Kramár, and Rohin Shah. 2023. Fact Finding: Attempting to Reverse-Engineer Factual Recall on the Neuron Level. https://www.lesswrong.com/posts/iGuwZTHWb6DFY3sKB/fact-finding- attempting-to-reverse-engineer-factual-recall Accessed: 2025-01-24

  22. [28]

    Vipula Rawte, Amit Sheth, and Amitava Das. 2023. A survey of hallucination in large foundation models.arXiv preprint arXiv:2309.05922(2023)

  23. [29]

    Stephen Robertson, Hugo Zaragoza, et al . 2009. The probabilistic relevance framework: BM25 and beyond.Foundations and Trends®in Information Retrieval 3, 4 (2009), 333–389

  24. [30]

    Alireza Salemi and Hamed Zamani. 2024. Towards a search engine for machines: Unified ranking for multiple retrieval-augmented large language models. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 741–751

  25. [31]

    Weihang Su. 2025. Homepage of Weihang Su. https://oneal2000.github.io/. Accessed: 2025-03-28

  26. [32]

    Weihang Su, Qingyao Ai, Xiangsheng Li, Jia Chen, Yiqun Liu, Xiaolong Wu, and Shengluan Hou. 2024. Wikiformer: Pre-training with structured information of wikipedia for ad-hoc retrieval. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 19026–19034

  27. [33]

    Weihang Su, Qingyao Ai, Yueyue Wu, Yixiao Ma, Haitao Li, and Yiqun Liu. 2023. Caseformer: Pre-training for Legal Case Retrieval.arXiv preprint arXiv:2311.00333 (2023)

  28. [34]

    Weihang Su, Yiran Hu, Anzhe Xie, Qingyao Ai, Quezi Bing, Ning Zheng, Yun Liu, Weixing Shen, and Yiqun Liu. 2024. STARD: A Chinese Statute Retrieval Dataset Derived from Real-life Queries by Non-professionals. InFindings of the Association for Computational Linguistics: EMNLP 2...

  29. [35]

    Weihang Su, Xiangsheng Li, Yiqun Liu, Min Zhang, and Shaoping Ma. 2023. Thuir2 at ntcir-16 session search (ss) task.arXiv preprint arXiv:2307.00250(2023)

  30. [36]

    Weihang Su, Yichen Tang, Qingyao Ai, Changyue Wang, Zhijing Wu, and Yiqun Liu. 2024. Mitigating entity-level hallucination in large language models. In Proceedings of the 2024 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the...

  31. [37]

    Weihang Su, Yichen Tang, Qingyao Ai, Zhijing Wu, and Yiqun Liu. 2024. DRAGIN: Dynamic Retrieval Augmented Generation based on the Real-time Information Needs of Large Language Models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Vo...

  32. [39]

    Weihang Su, Changyue Wang, Qingyao Ai, Yiran Hu, Zhijing Wu, Yujia Zhou, and Yiqun Liu. 2024. Unsupervised Real-Time Hallucination Detection based on the Internal States of Large Language Models. InFindings of the Association for Computational Linguistics: ACL 2024, Lun-Wei Ku...

  33. [40]

    Weihang Su, Changyue Wang, Anzhe Xie, Qingyao Ai, Yiran Hu, and Yiqun Liu. 2024. LegalAID: A Large Language Model for the Chinese Legal Field. https://github.com/oneal2000/LegalAID

  34. [41]

    Weihang Su, Baoqing Yue, Qingyao Ai, Yiran Hu, Jiaqi Li, Changyue Wang, Kaiyuan Zhang, Yueyue Wu, and Yiqun Liu. 2025. JuDGE: Benchmarking Judgment Document Generation for Chinese Legal System. InProceedings of the 48th International ACM SIGIR Conference on Research and Develo...

  35. [43]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971(2023)

  36. [44]

    Yiteng Tu, Weihang Su, Yujia Zhou, Yiqun Liu, and Qingyao Ai. 2025. RbFT: Ro- bust Fine-tuning for Retrieval-Augmented Generation against Retrieval Defects. arXiv preprint arXiv:2501.18365(2025)

  37. [45]

    Changyue Wang, Weihang Su, Qingyao Ai, and Yiqun Liu. 2024. Knowledge Editing through Chain-of-Thought.arXiv preprint arXiv:2412.17727(2024)

  38. [46]

    Changyue Wang, Weihang Su, Qingyao Ai, and Yiqun Liu. 2025. Decoupling Reasoning and Knowledge Injection for In-Context Knowledge Editing.arXiv preprint arXiv:2506.00536(2025)

  39. [47]

    Changyue Wang, Weihang Su, Qingyao Ai, and Yiqun Liu. 2025. Joint Evaluation of Answer and Reasoning Consistency for Hallucination Detection in Large Reasoning Models.arXiv preprint arXiv:2506.04832(2025)

  40. [48]

    Changyue Wang, Weihang Su, Hu Yiran, Qingyao Ai, Yueyue Wu, Cheng Luo, Yiqun Liu, Min Zhang, and Shaoping Ma. 2024. LeKUBE: A Legal Knowledge Update BEnchmark.arXiv preprint arXiv:2407.14192(2024)

  41. [49]

    Song Wang, Yaochen Zhu, Haochen Liu, Zaiyi Zheng, Chen Chen, and Jundong Li

  42. [50]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al . 2024. Qwen2. 5 Technical Report.arXiv preprint arXiv:2412.15115(2024)

  43. [51]

    Surveys 57, 3 (2024), 1–37

    Knowledge editing for large language models: A survey.Comput. Surveys 57, 3 (2024), 1–37

  44. [52]

    Ziyi Ye, Xiaohui Xie, Qingyao Ai, Yiqun Liu, Zhihong Wang, Weihang Su, and Min Zhang. 2024. Relevance Feedback with Brain Signals.ACM Transactions on Information Systems42, 4 (2024), 1–37

  45. [53]

    Zijun Yao, Weijian Qi, Liangming Pan, Shulin Cao, Linmei Hu, Weichuan Liu, Lei Hou, and Juanzi Li. 2024. Seakr: Self-aware knowledge retrieval for adaptive retrieval augmented generation.arXiv preprint arXiv:2406.19215(2024)

  46. [54]

    ChengXiang Zhai. 2008. Statistical language models for information retrieval. Synthesis lectures on human language technologies1, 1 (2008), 1–141

  47. [55]

    Zeping Yu and Sophia Ananiadou. 2024. Neuron-Level Knowledge Attribution in Large Language Models. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (Eds.). Association for Computational ...

  48. [56]

    Jingtao Zhan, Jiaxin Mao, Yiqun Liu, Jiafeng Guo, Min Zhang, and Shaoping Ma

  49. [57]

    Jingtao Zhan. 2025. Homepage of Jingtao Zhan. https://jingtaozhan.github.io/. Accessed: 2025-03-28

  50. [2021]

    InProceed- ings of the 44th international ACM SIGIR conference on research and development in information retrieval

    Optimizing dense retrieval model training with hard negatives. InProceed- ings of the 44th international ACM SIGIR conference on research and development in information retrieval. 1503–1512. 5

  51. [2024]

    InInternational Conference on Learning Representations (ICLR)

    Self-Reflective Retrieval-Augmented Generation (Self-RAG). InInternational Conference on Learning Representations (ICLR)

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.