RAG System for Supporting Japanese Litigation Procedures: Faithful Response Generation Complying with Legal Norms

Atsushi Keyaki; Hiroaki Yamada; Mihoko Sumida; Ryutaro Ohara; Yuya Ishihara

REVIEW 3 major objections 5 minor 19 references

RAG System for Supporting Japanese Litigation Procedures: Faithful Response Generation Complying with Legal Norms

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Legal RAG must obey three norms, not just answer accurately.

desk verdict A clear, honest position paper that derives three RAG evaluation axes from Japanese civil-procedure norms; the derivation is the contribution and it holds, but the third axis is a desideratum rather than an implementable requirement. read the letter →

arxiv 2511.22858 v1 pith:T2WUTR26 submitted 2025-11-28 cs.CL cs.IR

Yuya Ishihara , Atsushi Keyaki , Hiroaki Yamada , Ryutaro Ohara , Mihoko Sumida This is my paper · ORCID

classification cs.CLcs.IR

keywords Retrieval-AugmentedGenerationLegalnormsJapanesecivilprocedureMedicallitigationFaithfulnessDataattributionTemporalvalidityExpertknowledge

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

Working toward an AI that helps judges in Japanese medical lawsuits, this paper argues that a retrieval-augmented generation (RAG) system must satisfy three legal-norm requirements, not just answer accurately. First, the system may only draw on expert knowledge that is community-validated and equally accessible to both parties, respecting the civil-procedure prohibition on a judge's private knowledge. Second, every answer must be traceable to and faithful within the retrieved context; a correct-sounding answer drawn from the model's pretraining memory is norm-violating. Third, the system must use the knowledge valid at the legally relevant time—the standard of care when the alleged negligence happened versus the latest knowledge at the close of oral proceedings—managed through timestamps and citation networks. The paper proposes how each requirement could be evaluated, laying groundwork for a 'norm-compliant RAG' design.

What carries the argument

The carrying mechanism is a RAG pipeline equipped with three norm-enforcing components. A source-control layer restricts the document database to expert-community-validated sources and filters retrieval results so both parties have equal access. A faithfulness layer uses prompting constraints, chain-of-verification steps, and claim-level natural-language-inference scoring to force generated claims to originate in the retrieved context. A temporal layer attaches publication and validity timestamps to sources and navigates citation networks so that the system can select knowledge valid at the dispute-relevant time. Together these components turn ordinary retrieval-augmented generation into wha

What would settle it

Find a real medical negligence case where the standard of care at the time of the alleged act is contested between two streams of expert knowledge and no document carries a verifiable validity timestamp; if such a case is normatively unresolvable by retrieval, the temporal requirement (3) lacks a principled implementation. Alternatively, run an experiment where a RAG system answers a negligence query using only documents from the wrong decade, and show that standard faithfulness/accuracy metrics still score it as correct—demonstrating that existing evaluation misses the paper's dimension.

Watch

Extended reading notes

Core claim

The central claim is that a legal RAG system supporting expert testimony in Japanese medical litigation can be designated norm-compliant only if it meets three linked conditions: (1) controlled knowledge sourcing—retrieval is restricted to documents that have passed critical verification by the relevant expert community and are equally available to both parties; (2) attribution and faithfulness—generated responses must derive from the retrieved context and be faithful to it, with any reliance on pre-training knowledge treated as inappropriate and any contradiction of the source treated as a failure; and (3) issue-specific reference time—the system must know when each source was published and

Load-bearing premise

The load-bearing premise is that the legally relevant state of medical and legal knowledge at any given moment can be captured as retrievable documents with accurate publication and validity timestamps—so that which knowledge was 'valid at the time of the act' can be decided by a machine; the paper offers no evidence such timestamped, community-validated corpora exist or that validity periods can be automatically determined.

Editorial extensions

If this is right

If the three aspects are accepted, evaluations of legal RAG systems must add source-scope, attribution, and temporal-validity metrics alongside accuracy and factuality.
Responses that are factually correct but rely on the model's pretraining knowledge would be ruled norm-violating, changing the error taxonomy of legal AI.
The same norm-compliance lens applies to other expert-commissioner domains such as construction, intellectual property, and system-development litigation.
The timestamp requirement implies that RAG development for law requires building and maintaining document corpora with validated publication and validity dates, not just better retrieval ranking.

Reading between the lines

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

A natural testable extension: on a dataset of medical negligence questions, compare standard RAG faithfulness scores against human judgments about whether the answer would be admissible in court; the paper predicts mismatches whenever the answer is grounded in retrieved context that is temporally or procedurally inappropriate.
Implicit but not stated: if the 'equally available to both parties' requirement is taken seriously, the retrieval index must be a shared, inspectable artifact—this turns the RAG system from a judge's private assistant into an element of the adversarial procedure, with implications for transparency and audit.
The paper's own acknowledgment that hallucinations cannot be fully eliminated suggests the strongest version of the claim is about procedural legitimacy rather than total reliability; even partial substitution of expert commissioners may be acceptable only if a human expert verifies the final output.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: this is a workshop-grade position paper, and it's a decent one. The core move is reframing three familiar RAG concerns — source control, faithfulness/attribution, temporal validity — as consequences of Japan's adversarial civil procedure. That reframing is real and not something I've seen done before. The paper reads as careful: it grounds the source-control requirement in Japanese scholarship on judges' use of private knowledge, and the reference-time requirement in the distinction between the standard of care at the time of the alleged act versus knowledge at the close of oral proceedings.

What it does well: the requirements follow from the stated norms without circularity; the paper is explicit that implementation and evaluation are future work; and it doesn't oversell the feasibility of replacing expert commissioners with LLMs. It even cites work showing RAG reduces but doesn't eliminate hallucinations. That's honest.

Soft spots, in order of softness. First, requirement (3) is the load-bearing one and the least specified. The paper says valid time periods could be managed 'by utilizing timestamps and citation networks,' but the standard of care at time T is a legal construct, not something that falls out of publication dates. A source can be published but not yet accepted, or authoritative after being superseded. The paper itself notes that contradictory expert knowledge can coexist during transitions, but doesn't propose a representation for that coexistence. The stress-test note is right about this. I'd only push back on one implication: because the paper is a requirements proposal, not a claimed solution, a hard-to-operationalize axis is a weakness, not a fatal flaw. The requirements still have value as evaluation targets.

Second, the 'approximately twenty percent of data becomes outdated within five years' statistic in Section 2.2 comes from an acknowledged personal communication with Professor Yonemura but has no citation. For a paper that leans on it to motivate the need for retrieval, that's a thin reed.

Third, the novelty claim — no prior research on legal-norm conformance in RAG — is asserted with 'to the best of our knowledge.' It's plausible, but the paper doesn't document a systematic search of Japanese-language literature. Given the authors cite Japanese legal scholarship, they know the terrain; it's a minor gap.

Overall, the central argument holds at the level it's actually stated. The paper doesn't claim to have built the system. If the title is read as promising a system, that's misleading; but the body is clear that this is a design discussion.

Who gets value: people building legal RAG evaluation benchmarks, especially in civil-law jurisdictions, and researchers thinking about how procedural norms should constrain retrieval. It deserves a serious referee for a workshop or a short paper venue; it would not carry a full journal on its own.

My recommendation: engage with it, but treat it as a requirements paper. If a referee report is requested, accept the assignment. The paper is honest, citeable, and raises a real question about whether RAG evaluation should be norm-grounded.

Referee Report

3 major / 5 minor

Summary. The paper proposes a set of three requirements for a Retrieval-Augmented Generation (RAG) system intended to support Japanese medical litigation procedures while complying with legal norms. Drawing on the adversarial principle, the prohibition on judges' use of privately acquired expert knowledge, and the need for issue-specific reference times, the authors argue that a norm-compliant legal RAG must (1) restrict retrieval to sources that have been critically validated by the relevant expert community and made equally available to both parties, (2) generate responses that are attributable to and faithful to the retrieved context, and (3) use sources with timestamps appropriate to the legal question at issue. The paper presents these three aspects as the core contribution, with implementation and evaluation left to future work.

Significance. If validated, the proposed framework would move evaluation of legal RAG systems beyond raw accuracy and factuality, directing attention to procedural compliance and source temporal validity. The derivation is grounded in Japanese legal scholarship (e.g., [5], [6], [7]) and is internally coherent: the three requirements do follow from the adversarial principle and the issue-specific reference-time doctrine. The paper is honest about its scope, explicitly deferring methods and experiments to future work, and it does not overclaim empirical results. Its main value is as a conceptual standpoint for a domain where evaluation frameworks are still nascent.

major comments (3)

[Section 4.3 / Section 2.3] Requirement (3) is under-specified in a way that is load-bearing for the entire framework. The paper asserts that source validity can be managed 'by utilizing timestamps and citation networks' (Sec. 4.3), but it does not explain how legal validity at a given time is to be determined. The standard of care at time T is a legal construct, not a document-level metadata property: a source may be published but not yet accepted, or superseded but still authoritative. The paper itself acknowledges that contradictory expert knowledge can coexist during a transitional period (Sec. 2.3), yet offers no method to represent, retrieve, or reason about such coexisting sources. Without a concrete model of temporal validity and a procedure for deriving the reference time from the legal issue, requirement (3) cannot be evaluated or implemented. The statement that implementation is future work does not miti
[Section 4.1] The evaluation of requirement (1) is circular. The paper states that outputs derived from sources that deviate from 'the predefined scope' can be labeled inappropriate, but the 'predefined scope' is never defined. Section 2.1 gives a general principle—sources must be critically validated by the relevant expert community and equally available to both parties—but does not specify how this principle translates into a concrete corpus or retrieval filter. In particular, the paper does not address how to treat party-submitted evidence, which is a central element of the adversarial process. Without a definition of admissible source scope, requirement (1) cannot be tested, leaving the first axis of conformance ungrounded.
[Section 4.2] Requirement (2), while plausible, is also under-specified for the legal setting. The paper mentions prompting and chain-of-verification as possible enforcement mechanisms, and Ragas as an evaluation metric, but it does not address situations where the retrieved context contains conflicting authorities—an acknowledged possibility in Sec. 2.3. A faithful response may nevertheless be legally non-compliant if it silently selects one line of authority at the expense of another, or if it omits a minority view that is relevant to the issue. The paper should clarify whether faithfulness is to be interpreted as claim-level entailment with respect to the entire retrieved context, and how conflicting sources should be handled within the framework.

minor comments (5)

[Section 2.2] The statistic 'approximately twenty percent of such data becomes outdated within five years' is attributed to a professor but has no citation. Since this is used as motivation, please provide a source or clearly label it as an informal estimate.
[Figure 2] The components 'Maintenance of timestamp' and 'Identifying reference time' are shown in the retrieval module but are not defined in the text. Please explain what these components do and how they relate to requirements (2) and (3).
[Section 4.2] The phrase 'chain-of-verification steps' is used without a citation. If this refers to a specific method (e.g., self-verification [11] or a similar approach), please add a reference.
[References] Reference [16] appears to list the same authors as [13] but with a different title ('Reducing Hallucination in Structured Outputs via Retrieval-Augmented Generation'). Please verify the title, venue, and authors; the citation may be incorrect.
[Abstract and Section 4.3] The abstract states that the retrieval module must reference external knowledge with 'appropriate timestamps corresponding to the issues at hand,' while requirement (3) in Sec. 4.3 focuses on the 'valid time period' of sources. These are distinct concepts: one is about the issue's reference time, the other about a source's validity interval. Please clarify the relationship between them.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the requirements are grounded in external legal scholarship and the paper makes no empirical prediction to reduce.

full rationale

This is a position/requirements paper, not an empirical derivation. The three proposed aspects of 'conformance to the norm' are justified from external legal scholarship ([5], [6], [7]) and from procedural norms, not from the paper's own conclusion. No load-bearing self-citations appear: the cited legal authorities are external, and the authors do not invoke their own prior theorems, uniqueness results, or fitted parameters. The evaluative statements, such as labeling outputs from outside the predefined source scope as 'inappropriate' (Section 4.1) or treating responses based on inappropriate timestamps as 'unsuitable' (Section 4.3), are definitional components of the proposed evaluation framework rather than derived predictions; they do not claim to have been validated or fitted. The paper explicitly defers implementation and experiments to future work (Conclusion: 'Our future work includes proposing methods that satisfy each requirement, refining evaluation metrics, implementing them, and conducting experiments.'), so there is no 'prediction' whose equivalence to the inputs could be tested. The skeptic concern about timestamps and validity periods being under-specified is a feasibility or correctness risk, not circularity. Consequently, no circular step can be exhibited, and the appropriate score is 0.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

The paper introduces no new entities and no fitted parameters. Its central claim rests on domain assumptions drawn from Japanese civil procedure and legal scholarship; the least supported are the unsourced 20% decay statistic and the substitution-of-experts premise. These are assumptions the reader must accept without independent evidence specific to the system being built.

assumptions (6)

domain assumption Adversarial principle: judges must decide only on claims and evidence submitted by the parties and may not rely on privately acquired knowledge.
Section 2.1; grounds requirement (1) that retrieval sources must be restricted to party-accessible, community-validated knowledge.
domain assumption Expert knowledge may be used without evidentiary presentation only if it has been critically verified by the expert community and the parties can contest it.
Section 2.1 citing [5]; this admissibility criterion is what the system's source control operationalizes.
domain assumption Judges must continually update their understanding of statutes, judicial precedents, and social norms.
Section 2.2; motivates requirement (2) on faithfulness and attribution, since pre-training knowledge becomes obsolete.
domain assumption Negligence is judged by knowledge valid at the time of the act, while causal questions use the most up-to-date knowledge at the close of oral proceedings.
Section 2.3 citing [7]; motivates requirement (3) on issue-specific reference time.
ad hoc to paper Approximately 20% of authoritative medical data becomes outdated within five years.
Section 2.2; attributed to Prof. Yonemura as an informal estimate with no published source; used to justify the timeliness requirements.
domain assumption A RAG-based LLM can substitute for expert commissioners in litigation support.
Sections 1 and 2; the framing premise of the whole paper, asserted without demonstration and in tension with the cited observation that hallucinations cannot be fully eliminated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of RAG System for Supporting Japanese Litigation Procedures: Faithful Response Generation Complying with Legal Norms." pith.science (2026). https://pith.science/paper/T2WUTR26

@misc{pith2026251122858,
  author       = {Pith},
  title        = {Pith review of: RAG System for Supporting Japanese Litigation Procedures: Faithful Response Generation Complying with Legal Norms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T2WUTR26}},
  note         = {Machine review of arXiv:2511.22858}
}

read the original abstract

This study discusses the essential components that a Retrieval-Augmented Generation (RAG)-based LLM system should possess in order to support Japanese medical litigation procedures complying with legal norms. In litigation, expert commissioners, such as physicians, architects, accountants, and engineers, provide specialized knowledge to help judges clarify points of dispute. When considering the substitution of these expert roles with a RAG-based LLM system, the constraint of strict adherence to legal norms is imposed. Specifically, three requirements arise: (1) the retrieval module must retrieve appropriate external knowledge relevant to the disputed issues in accordance with the principle prohibiting the use of private knowledge, (2) the responses generated must originate from the context provided by the RAG and remain faithful to that context, and (3) the retrieval module must reference external knowledge with appropriate timestamps corresponding to the issues at hand. This paper discusses the design of a RAG-based LLM system that satisfies these requirements.

Figures

Figures reproduced from arXiv: 2511.22858 by the authors.

**Figure 1.** process of civil litigation witnesses, and offer explanations to judges to aid in assessing the reliability of evidence and witness testimonies presented by the parties. Particularly, because fact-finding by judges in medical litigation requires domain-specific medical expertise, identifying the matters that should be subject to expert testimony demands extensive referencing and analysis of a large volume of legal a… view at source ↗

**Figure 2.** Overview of the a RAG-based LLM System consider the expert knowledge that was valid at the relevant point in time. Based on the above, this study addresses the realization of a “norm-compliant RAG” system, focusing on: (1)controlling knowledge sources in compliance with procedural requirements concerning the use of expert knowledge, (2)attribution and faithfulness of generated responses to their information sources;… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 3 linked inside Pith

[5]

Okanari, Saibankan no shichi riyou no kinshi (the prohibition of judge ’s use of private knowledge), Hougaku Zasshi :(Journal of Law) of Osaka City University 68 (2021) 1 – 66

G. Okanari, Saibankan no shichi riyou no kinshi (the prohibition of judge ’s use of private knowledge), Hougaku Zasshi :(Journal of Law) of Osaka City University 68 (2021) 1 – 66

2021
[6]

E. Sugiyama, Saibankan niyoru senmonchishiki no shushu to riyou (collection and use of expert knowledge by the judge, symposium: The discipline of civil judges in the exercise of their powers), Minso Zasshi (Journal of Civil Procedure) 69 (2023) 103 – 114

2023
[7]

Y. Shirai, Mijukuji Moumakusyou to Ishi no Kashitu (Misdiagnosis of retinopathy of prematurity and Doctor’s Negligence), Hanrei kara Manabu Minji-Jijitsu Nintei: Jurisuto Zoukan (Special Edition of Journal: Jurist: Learning from Case Law: Civil Fact-Finding) (2006) 252–256

2006
[1]

Petroni, T

F. Petroni, T. Rocktäschel, S. Riedel, P. Lewis, A. Bakhtin, Y. Wu, A. Miller, Language Models as Knowledge Bases?, in: Proc. of the EMNLP-IJCNLP 2019, 2019

2019
[2]

J. Wei, Y. Tay, R. Bommasani, C. Raffel, B. Zoph, S. Borgeaud, D. Yogatama, M. Bosma, D. Zhou, D. Metzler, E. H. Chi, T. Hashimoto, O. Vinyals, P. Liang, J. Dean, W. Fedus, Emergent abilities of large language models, Transactions on Machine Learning Research (TMLR) (2022) 2835–8856

2022
[3]

Magesh, F

V. Magesh, F. Surani, M. Dahl, M. Suzgun, C. D. Manning, D. E. Ho, Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, Journal of Empirical Legal Studies 22 (2025) 216–242

2025
[4]

Y.-W. Chu, K. Zhang, C. Malon, M. R. Min, Reducing Hallucinations of Medical Multimodal Large Language Models with Visual Retrieval-Augmented Generation , in: Proc. of the AAAI 2025, 2025

2025
[8]

Albalak, Y

A. Albalak, Y. Elazar, S. M. Xie, S. Longpre, N. Lambert, X. Wang, N. Muennighoff, B. Hou, L. Pan, H. Jeong, C. Raffel, S. Chang, T. Hashimoto, W. Y. Wang, A Survey on Data Selection for Language Models, arXiv:2402.16827, 2024

arXiv 2024

Show all 19 references

[9]

K. Li, O. Patel, F. Viégas, H. Pfister, M. Wattenberg, Inference-Time Intervention: Eliciting Truthful Answers from a Language Model, in: Proc. of the NeurIPS 2023, 2023

2023
[10]

Manakul, A

P. Manakul, A. Liusie, M. Gales, SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models, in: Proc. of the EMNLP 2023, 2023

2023
[11]

Zhang, B

X. Zhang, B. Peng, Y. Tian, J. Zhou, L. Jin, L. Song, H. Mi, H. Meng, Self-Alignment for Factuality: Mitigating Hallucinations in LLMs via Self-Evaluation, in: Proc. of the ACL 2024, 2024

2024
[12]

Y. Wang, R. G. Reddy, Z. M. Mujahid, A. Arora, A. Rubashevskii, J. Geng, O. M. Afzal, L. Pan, N. Borenstein, A. Pillai, I. Augenstein, I. Gurevych, P. Nakov, Factcheck-Bench: Fine-Grained Evaluation Benchmark for Automatic Fact-checkers, in: Proc. of the Findings of the EMNLP ...

2024
[13]

Lewis, E

P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Küttler, M. Lewis, W. tau Yih, T. Rocktäschel, S. Riedel, D. Kiela, Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, in: Proc. of the NIPS 2020, 2020

2020
[14]

Y. Gao, Y. Xiong, X. Gao, K. Jia, J. Pan, Y. Bi, Y. Dai, J. Sun, M. Wang, H. Wang, Retrieval-Augmented Generation for Large Language Models: A Survey, arXiv:2312.10997, 2023

2023 arXiv
[15]

S. T. I. Tonmoy, S. M. M. Zaman, V. Jain, A. Rani, V. Rawte, A. Chadha, A. Das, A Comprehensive Survey of Hallucination Mitigation Techniques in Large Language Models, arXiv:2401.01313, 2024

2024 arXiv
[16]

Lewis, E

P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Küttler, M. Lewis, W. tau Yih, T. Rocktäschel, S. Riedel, D. Kiela, Reducing Hallucination in Structured Outputs via Retrieval- Augmented Generation, in: Proc. of the NAACL 2024, 2024

2024
[17]

Zhang, J

W. Zhang, J. Zhang, Hallucination Mitigation for Retrieval-Augmented Large Language Models: A Review, Mathematics 13 (2025)

2025
[18]

Pruthi, F

G. Pruthi, F. Liu, S. Kale, M. Sundararajan, Estimating Training Data Influence by Tracing Gradient Descent, in: Proc. of the NeurIPS 2020, 2020

2020
[19]

T. A. Chang, D. Rajagopal, T. Bolukbasi, L. Dixon, I. Tenney, Scalable Influence and Fact Tracing for Large Language Model Pretraining, in: Proc. of the ICLR 2025, 2025

2025

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [5]

Okanari, Saibankan no shichi riyou no kinshi (the prohibition of judge ’s use of private knowledge), Hougaku Zasshi :(Journal of Law) of Osaka City University 68 (2021) 1 – 66

G. Okanari, Saibankan no shichi riyou no kinshi (the prohibition of judge ’s use of private knowledge), Hougaku Zasshi :(Journal of Law) of Osaka City University 68 (2021) 1 – 66

2021

[2] [6]

E. Sugiyama, Saibankan niyoru senmonchishiki no shushu to riyou (collection and use of expert knowledge by the judge, symposium: The discipline of civil judges in the exercise of their powers), Minso Zasshi (Journal of Civil Procedure) 69 (2023) 103 – 114

2023

[3] [7]

Y. Shirai, Mijukuji Moumakusyou to Ishi no Kashitu (Misdiagnosis of retinopathy of prematurity and Doctor’s Negligence), Hanrei kara Manabu Minji-Jijitsu Nintei: Jurisuto Zoukan (Special Edition of Journal: Jurist: Learning from Case Law: Civil Fact-Finding) (2006) 252–256

2006

[4] [1]

Petroni, T

F. Petroni, T. Rocktäschel, S. Riedel, P. Lewis, A. Bakhtin, Y. Wu, A. Miller, Language Models as Knowledge Bases?, in: Proc. of the EMNLP-IJCNLP 2019, 2019

2019

[5] [2]

J. Wei, Y. Tay, R. Bommasani, C. Raffel, B. Zoph, S. Borgeaud, D. Yogatama, M. Bosma, D. Zhou, D. Metzler, E. H. Chi, T. Hashimoto, O. Vinyals, P. Liang, J. Dean, W. Fedus, Emergent abilities of large language models, Transactions on Machine Learning Research (TMLR) (2022) 2835–8856

2022

[6] [3]

Magesh, F

V. Magesh, F. Surani, M. Dahl, M. Suzgun, C. D. Manning, D. E. Ho, Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, Journal of Empirical Legal Studies 22 (2025) 216–242

2025

[7] [4]

Y.-W. Chu, K. Zhang, C. Malon, M. R. Min, Reducing Hallucinations of Medical Multimodal Large Language Models with Visual Retrieval-Augmented Generation , in: Proc. of the AAAI 2025, 2025

2025

[8] [8]

Albalak, Y

A. Albalak, Y. Elazar, S. M. Xie, S. Longpre, N. Lambert, X. Wang, N. Muennighoff, B. Hou, L. Pan, H. Jeong, C. Raffel, S. Chang, T. Hashimoto, W. Y. Wang, A Survey on Data Selection for Language Models, arXiv:2402.16827, 2024

arXiv 2024