Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

MIRAGE: Defending Long-Form RAG Against Misinformation Pollution

T0 review · 4 major / 6 minor · reviewed 2026-07-11 · grok-4.5

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

Pith's one-line read Cross-document claim consistency can filter or block polluted RAG evidence before generation, restoring long-form factuality.

desk verdict Solid systems paper: training-free claim-graph defense plus a reusable pollution protocol, with large multi-model gains that partly rest on synthetic edits and fallback blocking. read the letter →

arxiv 2607.05069 v1 pith:NMAQXCIT submitted 2026-07-06 cs.CL

Saadeldine Eletter , Ruihong Zeng , Yuxia Wang , Maxim Panov , Aleksandr Rubashevskii , Preslav Nakov This is my paper
classification cs.CL
keywords retrieval-augmentedgenerationmisinformationpollutionlong-formquestionansweringclaimgraphnaturallanguageinferencedefended-claimsgaterobustRAGfactualityevaluation
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

Retrieval-augmented generation is supposed to ground language models in external evidence, but real-world retrieval often returns on-topic passages that contain fabrications, entity swaps, misleading framings, or other subtle errors. The paper argues that these polluted passages can steer models into confident but wrong long-form answers, and that relevance alone is not enough. MIRAGE is a training-free wrapper that extracts sentence-level claims from the retrieved set, links them with natural-language-inference support and contradiction edges across sources, prunes to a consistent multi-source subset, and then either conditions generation on those defended claims or blocks retrieval and answers from parametric knowledge. The authors also release a minimal-edit pollution protocol with four realistic corruption families so that clean, mixed, and fully polluted regimes can be compared fairly. Across four long-form QA benchmarks and several commercial and open-weight models, pollution severely hurts vanilla RAG, while MIRAGE restores factuality and beats prior robust-RAG defenses without collapsing performance on clean retrieval.

What carries the argument

The Defended-Claims Gate over an NLI-based cross-document claim graph: claims become nodes with support and contradiction edges; a consistent multi-source subset is selected; retrieval is trusted only when the contradiction ratio is low and source diversity is high, otherwise generation is blocked and the model answers parametrically.

What would settle it

Construct a mixed or fully polluted retrieval set in which every polluted claim is lexically anchored, multi-source consistent, and not contradicted by counter-evidence probes, then check whether MIRAGE still passes the gate and whether VeriScore F1 falls back toward vanilla polluted RAG rather than No-RAG or clean levels.

Watch

Extended reading notes

Core claim

The paper claims that truthful evidence for long-form answers tends to be corroborated across independent sources, whereas polluted evidence induces structural inconsistency that an NLI claim graph can detect. By scoring claims for multi-source support, pruning contradictions, and gating generation on global consistency and source diversity, MIRAGE can either generate from a defended subset of verified claims or fall back to no-retrieval answering, restoring factuality under mixed and fully polluted retrieval while staying close to clean RAG when evidence is coherent.

Load-bearing premise

The method assumes that true claims usually agree across sources while polluted claims create detectable contradictions under NLI and a lexical-overlap pair filter, so coordinated multi-source misinformation or sparse paraphrase-heavy pollution can defeat the gate.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes MIRAGE, a training-free, model-agnostic wrapper for long-form RAG that extracts sentence-level claims from top-k retrieved passages, builds a cross-source NLI support/contradiction graph (with optional counter-evidence probes), greedily prunes to a defended claim set S⋆, and applies a Defended-Claims Gate (contradiction ratio r_contr and source diversity d_src; eqs. 5–6) that either conditions generation on verified claims or blocks retrieval and answers parametrically. The authors also introduce a minimal-edit pollution protocol with four families (Unambiguous, Conflicting, Misleading, Fabricated) to create matched Clean/Mixed/FullP regimes. Across four long-form QA benchmarks and five LLM backbones, vanilla RAG degrades sharply under pollution while MIRAGE largely restores VeriScore F1@k and outperforms many robust-RAG baselines (Tables 1–2), with ablations, gate pass/block rates, k-sensitivity, and NLI/VeriScore label audits.

Significance. If the results transfer beyond the synthetic pollution generator, this is a useful systems contribution: a modular, training-free evidence-adjudication layer that targets on-topic misinformation rather than mere irrelevance, with released code, datasets, and a controllable pollution protocol that the community can reuse. Strengths include breadth of evaluation (four datasets, commercial and open-weight models, many baselines), offline gate calibration without the generator, separation of VeriScore’s external evidence from RAG passages, component ablations (Table 5), and human agreement checks on NLI edges and VeriScore labels (Table 6). The work is significant primarily as an empirical defense and evaluation protocol rather than as a new theoretical principle of multi-source consistency.

major comments (4)
  1. [§4 Pollution Protocol; Limitations] §4 and App. B/K: The central mechanism claim—that truthful claims are multi-source consistent while polluted claims induce NLI-detectable structural inconsistency (§1, §3; eqs. 5–6)—is stress-tested almost exclusively under GPT-4o-mini minimal-edit rewrites that often flip atomic facts or invent particulars while preserving topic and lexical anchors. Coordinated multi-source falsehoods, shared framing across domains, or paraphrase/omission pollution that preserves surface agreement are acknowledged in Limitations but not measured. Without at least one such stress regime (or real web-pollution sample), Tables 1–2 support robustness to this generator more than to the broader misinformation setting the abstract invokes. Please add such an evaluation or substantially narrow the claim language.
  2. [Table 3; §6.1–6.2; Appendix I] Table 3 and §6.1–6.2: Under FullP the gate pass rate is ~0%, so reported FullP recovery is largely safer parametric fallback (plus the stronger no-evidence prompt noted in App. I), not successful adjudication of polluted evidence. MixP gains mix pruning of inconsistent claims with partial blocking (~48.6% pass). The main text currently presents both regimes as “restoring factuality” via the same pipeline. Please decompose results by gate decision (PASS vs BLOCK), report factuality conditional on pass, and state explicitly how much of FullP improvement is blocking versus claim pruning.
  3. [§3.4; Table 8; eqs. (5)–(6)] Table 8 / eqs. (5)–(6): The global configuration sets r_max = 0.98, so the contradiction-ratio criterion is nearly inactive unless almost all edges are contradictions; d_min = 2 then dominates the trust decision. This makes the gate’s advertised dual criteria asymmetric and raises the question whether “consistency” is doing the work claimed in §3.4. Please report (i) which criterion triggers blocks in Clean/MixP/FullP, (ii) sensitivity of Tables 1–2 to r_max and d_min, and (iii) justification for r_max = 0.98 beyond offline grid search on held-out graph statistics.
  4. [§3.2 Pair selection; Appendix J.2] §3.2 / App. J.2: Pair selection uses containment-style token overlap with τ_overlap = 0.62 before NLI. The diagnostic (Table 18) shows large NLI savings but also that paraphrastic or omission-based pollution can sparsify the graph and push the system toward conservative blocking. Given Misleading pollution is one of the four families, please quantify false-negative edge rates on Misleading/paraphrase pairs and show how often MixP failures are due to missed support/contradiction edges rather than gate thresholds.
minor comments (6)
  1. [Figure 1; Figure 3] Figure 1 and Figure 3 are helpful but low-resolution in the manuscript text; ensure vector figures and readable edge labels in the camera-ready version.
  2. [§3.2; §5.2; Table 4] Notation: ˜r(c) is introduced as max-normalized retrieval score; later tables use F1@k with k overloaded as both retrieval depth and VeriScore’s claim-count median. Disambiguate (e.g., F1@K vs. top-k).
  3. [§2 Related Work] Related Work (§2) groups many concurrent robust-RAG methods; a short table of assumptions (training-free? multi-source consistency? pollution type) would help readers place MIRAGE relative to AstuteRAG, TrustRAG, RobustRAG, and MADAM-RAG.
  4. [Appendix C.2; Table 10] Domain-trust priors (Table 10) assign Wikipedia 0.90 and forums 0.20. Even if secondary (α = 0.5), a one-paragraph sensitivity check with α = 0 (no domain prior) would strengthen the claim that multi-source agreement dominates.
  5. [Throughout] Minor prose: “FA V A” spacing artifacts and occasional hyphenation breaks (e.g., “MI-RAGE”) should be cleaned for production.
  6. [Appendix G] App. G runtime is on 30 samples with API cost only; wall-clock NLI latency per query would help practitioners assess the local compute trade-off claimed in Limitations.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MIRAGE is an empirical systems defense whose results do not reduce by construction to fitted inputs or self-citation.

full rationale

MIRAGE does not present a first-principles derivation or a fitted quantity renamed as a prediction. The core pipeline (claim extraction, cross-source NLI graph, greedy pruning of F(S), and the Defended-Claims Gate on r_contr and d_src) is a designed procedure whose success is measured empirically against external VeriScore labels. Gate thresholds are chosen offline from held-out graph statistics without the generator and are excluded from evaluation, so they are hyperparameters rather than circular fits. Claim-scoring weights (α=0.5, β=0.3) are fixed priority choices, not outcome-fitted parameters. Evaluation evidence for VeriScore is retrieved via Serper independently of the RAG passages, so the metric is not the same object as the defended graph. The concurrent Eletter et al. (2026) self-citation is non-load-bearing multimodal concurrent work. Author-defined pollution families and domain priors raise external-validity questions about whether synthetic edits induce the inconsistency the method assumes, but that is evaluation design risk, not a reduction of a claimed prediction to its inputs by construction. Under the stated circularity criteria, the paper is self-contained against its benchmarks.

Assumptions & free parameters 6 free parameters · 5 assumptions · 2 invented entities

MIRAGE is an engineering system paper. Its central claim rests on empirical evaluation plus design assumptions about multi-source consistency, NLI reliability, and controllable pollution—not on free physical constants or invented particles. The free parameters are the fixed gate and graph-construction knobs; axioms are standard NLP tooling assumptions plus the paper’s consistency hypothesis; invented entities are the named method components and pollution taxonomy.

free parameters (6)
  • r_max (max contradiction ratio)
    Global gate threshold fixed offline at 0.98 from held-out graph statistics; controls when retrieval is trusted.
  • d_min (min source diversity)
    Global gate threshold fixed at 2; requires multi-domain support before defended generation.
  • α, β claim-scoring weights
    Fixed at 0.5 and 0.3 to prioritize cross-source support over domain prior and retrieval rank.
  • τ_overlap lexical pair filter
    Set to 0.62 to prefilter claim pairs before NLI; trades paraphrase recall for efficiency.
  • Graph caps (k, N, P_max, M, k_ctr, θ_support)
    Computational defaults (e.g., k=20, N=120, P_max=400, M=8) that shape which claims and edges enter the defended set.
  • Domain-trust prior table
    Hand-assigned trust weights by TLD/keyword (Wikipedia 0.90 … forums 0.20) used as a secondary score term.
assumptions (5)
  • ad hoc to paper Truthful claims tend to be corroborated across independent sources, while misinformation induces contradictions detectable by NLI.
    Stated as the core insight in §1/§3; load-bearing for pruning and gating.
  • domain assumption Off-the-shelf NLI (DeBERTa-large-MNLI etc.) is a sufficiently reliable adjudicator of support/contradiction between claims and counter-evidence.
    Used for all edge weights and self-contradiction probes; partially validated on 150 pairs.
  • ad hoc to paper Minimal-edit GPT-4o-mini rewrites of clean passages are a realistic proxy for real-world retrieval pollution.
    Defines the CLEAN/MIXED/POLLUTED regimes in §4; evaluation validity depends on this proxy.
  • domain assumption VeriScore F1@k with external Serper evidence measures long-form factuality comparably across clean and polluted generation contexts.
    Primary metric throughout §5–6; supported by limited human and multi-model agreement checks.
  • domain assumption Standard dense retrieval and sentence segmentation preserve enough claim structure for graph construction without LLM claim decomposition.
    Pipeline premise in §3 and Appendix J; keeps method training-free.
invented entities (2)
  • MIRAGE defended-claims graph and Defended-Claims Gate
    purpose: Select a consistent multi-source claim subset and decide whether to condition generation or fall back parametrically.
    Central proposed mechanism; no independent existence outside this system design.
  • Four-family minimal-edit pollution protocol (Unambiguous, Conflicting, Misleading, Fabricated)
    purpose: Create matched clean/mixed/full pollution evaluation regimes while preserving topical relevance.
    Author-constructed benchmark generator; useful artifact but not an external natural phenomenon.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MIRAGE: Defending Long-Form RAG Against Misinformation Pollution." pith.science (2026). https://pith.science/paper/NMAQXCIT

@misc{pith2026260705069,
  author       = {Pith},
  title        = {Pith review of: MIRAGE: Defending Long-Form RAG Against Misinformation Pollution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NMAQXCIT}},
  note         = {Machine review of arXiv:2607.05069}
}
read the original abstract

Retrieval-Augmented Generation (RAG) improves factuality by grounding LLMs in external evidence, but real-world retrieval is often polluted: semantically relevant passages may contain subtle misinformation, misleading framings, or fabrications. We introduce MIRAGE, a training-free, model-agnostic defense for long-form RAG. MIRAGE builds an NLI-based cross-document claim graph and applies a Defended-Claims Gate to either condition generation on a consistent, multi-source supported subset or to block retrieval and answer parametrically. We also release a minimal-edit pollution protocol spanning four perturbation families (Unambiguous, Conflicting, Misleading, Fabricated) to construct matched clean, mixed, and fully polluted evaluation regimes. Across four long-form QA benchmarks and multiple commercial and open-weight LLMs, pollution severely degrades vanilla RAG, while MIRAGE consistently restores factuality under mixed and fully polluted evidence and outperforms prior robust-RAG methods. Our implementation and datasets are available at https://github.com/SaadElDine/MIRAGE.

Figures

Figures reproduced from arXiv: 2607.05069 by the authors.

Figure 1
Figure 1. Polluted RAG problem. A retriever returns a mix of clean and polluted passages. Even a single high￾scoring polluted document can steer the LLM toward a confident but incorrect answer. errors into the final answer with high confi￾dence (Perez et al., 2023) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. and Algorithm 1 summarize the entire pipeline. 3.2 Claim Graph Construction Given a user query q, a retriever returns a set of top￾k passages D = {d1, . . . , dk}. For each passage d ∈ D, MIRAGE performs lightweight sentence segmentation and treats each sentence as a claim c, for which we track the following: • a source identifier src(c) ∈ {1, . . . , k}; • a domain dom(c), derived from the source URL (e.g., wikiped… view at source ↗
Figure 3
Figure 3. Defended-claims graph. Claims form a graph with NLI-based support (green) and contradiction (red) edges. Polluted claims induce dense contradictions, while clean evidence forms multi-source support clus￾ters. MIRAGE uses these patterns to detect unreliable retrieval. If either condition in equation (6) is violated, MI￾RAGE deems retrieval unreliable and blocks RAG, instructing the LLM to answer from its paramet￾ric … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Pollution-type counts by dataset. Bars are grouped by pollution family (Conflicting, Misleading, Fabricated, Unambiguous); within each group, colors indicate the dataset (see legend). Top: Full regime, where all retrieved passages are corrupted. Bottom: Mixed regime, w…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Trust Before Fusion: QIMG-7 and Source-Aware Resolution for Polluted Multimodal RAG

    cs.CL 2026-07 conditional novelty 6.0 of 10

    Under multimodal retrieval pollution, selective source-aware trust (Field-Selector SATR) beats unconditional fusion by ~12 balanced points on QIMG-7, driven mainly by text-reliability modeling.

Reference graph

Works this paper leans on

88 extracted references · 2 canonical work pages · cited by 1 Pith paper

  1. [1]

    Retrieval-Augmented Generation for

    Zhao, Penghao and Zhang, Hailin and Yu, Qinhan and Wang, Zhengren and Geng, Yunteng and Fu, Fangcheng and Yang, Ling and Zhang, Wentao and Jiang, Jie and Cui, Bin , journal =. Retrieval-Augmented Generation for. 2026 , publisher =

    2026

  2. [2]

    2024 , publisher =

    Yu, Yue and Ping, Wei and Liu, Zihan and Wang, Boxin and You, Jiaxuan and Zhang, Chao and Shoeybi, Mohammad and Catanzaro, Bryan , booktitle =. 2024 , publisher =

    2024

  3. [3]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , series =

    When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memories , author =. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , series =. 2023 , address =

    2023

  4. [4]

    ArXiv preprint , volume =

    Reliable, Adaptable, and Attributable Language Models with Retrieval , author =. ArXiv preprint , volume =. 2024 , eprint =

    2024

  5. [5]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , series =

    Dense Passage Retrieval for Open-Domain Question Answering , author =. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , series =. 2020 , address =

    2020

  6. [6]

    Findings of the Association for Computational Linguistics: EMNLP 2023 , series =

    How to Train Your Dragon: Diverse Augmentation Towards Generalizable Dense Retrieval , author =. Findings of the Association for Computational Linguistics: EMNLP 2023 , series =. 2023 , address =

    2023

  7. [7]

    and Smith, Noah A

    Kasai, Jungo and Sakaguchi, Keisuke and Takahashi, Yoichi and Le Bras, Ronan and Asai, Akari and Yu, Xinyan Velocity and Radev, Dragomir R. and Smith, Noah A. and Choi, Yejin and Inui, Kentaro , booktitle =. 2023 , publisher =

    2023

  8. [8]

    Findings of the Association for Computational Linguistics: ACL 2024 , series =

    Benchmarking Retrieval-Augmented Generation for Medicine , author =. Findings of the Association for Computational Linguistics: ACL 2024 , series =. 2024 , address =

    2024

Show all 88 references
  1. [9]

    Findings of the Association for Computational Linguistics: EMNLP 2023 , series =

    On the Risk of Misinformation Pollution with Large Language Models , author =. Findings of the Association for Computational Linguistics: EMNLP 2023 , series =. 2023 , address =

    2023

  2. [10]

    Findings of the Association for Computational Linguistics: NAACL 2024 , series =

    Why So Gullible? Enhancing the Robustness of Retrieval-Augmented Models against Counterfactual Noise , author =. Findings of the Association for Computational Linguistics: NAACL 2024 , series =. 2024 , address =

    2024

  3. [11]

    Findings of the Association for Computational Linguistics: NAACL 2024 , series =

    Adapting Fake News Detection to the Era of Large Language Models , author =. Findings of the Association for Computational Linguistics: NAACL 2024 , series =. 2024 , address =

    2024

  4. [12]

    ArXiv preprint , volume =

    Fake News Detectors are Biased Against Texts Generated by Large Language Models , author =. ArXiv preprint , volume =. 2023 , eprint =

    2023

  5. [13]

    2020 , publisher =

    Shu, Kai and Mahudeswaran, Deepak and Wang, Suhang and Lee, Dongwon and Liu, Huan , journal =. 2020 , publisher =

    2020

  6. [14]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , series =

    Improving Factuality with Explicit Working Memory , author =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , series =. 2025 , address =

    2025

  7. [15]

    On the Structural Memory of

    Zeng, Ruihong and Fang, Jinyuan and Liu, Siwei and Meng, Zaiqiao , journal =. On the Structural Memory of. 2024 , eprint =

    2024

  8. [16]

    Proceedings of the International Conference on Learning Representations , series =

    Self-Consistency Improves Chain of Thought Reasoning in Language Models , author =. Proceedings of the International Conference on Learning Representations , series =. 2023 , url =

    2023

  9. [17]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , series =

    Malaviya, Chaitanya and Shaw, Peter and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina , editor =. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , series =. 2023 , address =

    2023

  10. [18]

    ArXiv preprint , volume =

    Learning to Reason for Factuality , author =. ArXiv preprint , volume =. 2025 , eprint =

    2025

  11. [19]

    Findings of the Association for Computational Linguistics: ACL 2024 , series =

    Chain-of-Verification Reduces Hallucination in Large Language Models , author =. Findings of the Association for Computational Linguistics: ACL 2024 , series =. 2024 , address =

    2024

  12. [20]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , series =

    Verify-and-Edit: A Knowledge-Enhanced Chain-of-Thought Framework , author =. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , series =. 2023 , address =

    2023

  13. [21]

    Advances in Neural Information Processing Systems , series =

    Long-form Factuality in Large Language Models , author =. Advances in Neural Information Processing Systems , series =. 2024 , publisher =

    2024

  14. [22]

    Proceedings of the Conference on Language Modeling , series =

    Fine-grained Hallucination Detection and Editing for Language Models , author =. Proceedings of the Conference on Language Modeling , series =. 2024 , url =

    2024

  15. [23]

    and Hashimoto, Tatsunori B

    Dubois, Yann and Li, Chen Xuechen and Taori, Rohan and Zhang, Tianyi and Gulrajani, Ishaan and Ba, Jimmy and Guestrin, Carlos and Liang, Percy S. and Hashimoto, Tatsunori B. , booktitle =. 2023 , publisher =

    2023

  16. [24]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , series =

    Min, Sewon and Krishna, Kalpesh and Lyu, Xinxi and Lewis, Mike and Yih, Wen-tau and Koh, Pang and Iyyer, Mohit and Zettlemoyer, Luke and Hajishirzi, Hannaneh , editor =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , series =. 2023 , address =

    2023

  17. [25]

    ArXiv preprint , volume =

    A Comprehensive Survey of Hallucination Mitigation Techniques in Large Language Models , author =. ArXiv preprint , volume =. 2024 , eprint =

    2024

  18. [26]

    2023 , publisher =

    Liu, Yi-Hsien and Han, Tianle and Ma, Siyuan and Zhang, Jia-Yu and Yang, Yuanyu and Tian, Jiaming and He, Haoyang and Li, Antong and He, Mengshen and Liu, Zheng and others , journal =. 2023 , publisher =

    2023

  19. [27]

    Self-Alignment for Factuality: Mitigating Hallucinations in

    Zhang, Xiaoying and Peng, Baolin and Tian, Ye and Zhou, Jingyan and Jin, Lifeng and Song, Linfeng and Mi, Haitao and Meng, Helen , editor =. Self-Alignment for Factuality: Mitigating Hallucinations in. Proceedings of the 62nd Annual Meeting of the Association for Computational...

    2024

  20. [28]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , series =

    Wang, Yuxia and Reddy, Revanth Gangi and Mujahid, Zain Muhammad and Arora, Arnav and Rubashevskii, Aleksandr and Geng, Jiahui and Afzal, Osama Mohammed and Pan, Liangming and Borenstein, Nadav and Pillai, Aditya and Augenstein, Isabelle and Gurevych, Iryna and Nakov, Preslav ,...

    2024

  21. [29]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , series =

    Fatahi Bayat, Farima and Zhang, Lechen and Munir, Sheza and Wang, Lu , editor =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , series =. 2025 , address =

    2025

  22. [30]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , series =

    Song, Yixiao and Kim, Yekyung and Iyyer, Mohit , editor =. Findings of the Association for Computational Linguistics: EMNLP 2024 , series =. 2024 , address =

    2024

  23. [31]

    Length-Controlled

    Dubois, Yann and Liang, Percy and Hashimoto, Tatsunori , booktitle =. Length-Controlled. 2024 , url =

    2024

  24. [32]

    Retrieval-Augmented Generation for Knowledge-Intensive

    Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and K. Retrieval-Augmented Generation for Knowledge-Intensive. Advances in Neural Information Processing Systems , series =. 2020 , publisher =

    2020

  25. [33]

    Proceedings of the 37th International Conference on Machine Learning , series =

    Retrieval Augmented Language Model Pre-Training , author =. Proceedings of the 37th International Conference on Machine Learning , series =. 2020 , publisher =

    2020

  26. [34]

    ArXiv preprint , volume =

    Retrieval-Augmented Generation for Large Language Models: A Survey , author =. ArXiv preprint , volume =. 2023 , eprint =

    2023

  27. [35]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , series =

    Active Retrieval Augmented Generation , author =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , series =. 2023 , address =

    2023

  28. [36]

    Asai, Akari and Wu, Zeqiu and Wang, Yizhong and Sil, Avirup and Hajishirzi, Hannaneh , booktitle =. Self-. 2024 , url =

    2024

  29. [37]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , series =

    Su, Weihang and Tang, Yichen and Ai, Qingyao and Wu, Zhijing and Liu, Yiqun , editor =. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , series =. 2024 , address =

    2024

  30. [38]

    2025 , url =

    Wei, Zhepei and Chen, Wei-Lin and Meng, Yu , booktitle =. 2025 , url =

    2025

  31. [39]

    Certifiably Robust

    Xiang, Chong and Wu, Tong and Zhong, Zexuan and Wagner, David and Chen, Danqi and Mittal, Prateek , booktitle =. Certifiably Robust. 2024 , url =

    2024

  32. [40]

    Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , series =

    Lewis, Mike and Liu, Yinhan and Goyal, Naman and Ghazvininejad, Marjan and Mohamed, Abdelrahman and Levy, Omer and Stoyanov, Veselin and Zettlemoyer, Luke , editor =. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , series =. 2020 , address =

    2020

  33. [41]

    Worse than Zero-shot? A Fact-Checking Dataset for Evaluating the Robustness of

    Zeng, Linda and Gupta, Rithwik and Motwani, Divij and Zhang, Yi and Yang, Diji , booktitle =. Worse than Zero-shot? A Fact-Checking Dataset for Evaluating the Robustness of. 2025 , url =

    2025

  34. [42]

    Proceedings of the Conference on Language Modeling , series =

    Retrieval-Augmented Generation with Conflicting Evidence , author =. Proceedings of the Conference on Language Modeling , series =. 2025 , url =

    2025

  35. [43]

    , editor =

    Wang, Fei and Wan, Xingchen and Sun, Ruoxi and Chen, Jiefeng and Arik, Sercan O. , editor =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , series =. 2025 , address =

    2025

  36. [44]

    Findings of the Association for Computational Linguistics: ACL 2023 , series =

    Discovering Language Model Behaviors with Model-Written Evaluations , author =. Findings of the Association for Computational Linguistics: ACL 2023 , series =. 2023 , address =

    2023

  37. [45]

    Proceedings of the International Conference on Learning Representations , series =

    Making Retrieval-Augmented Language Models Robust to Irrelevant Context , author =. Proceedings of the International Conference on Learning Representations , series =. 2024 , url =

    2024

  38. [46]

    ArXiv preprint , volume =

    After Retrieval, Before Generation: Enhancing the Trustworthiness of Large Language Models in Retrieval-Augmented Generation , author =. ArXiv preprint , volume =. 2025 , eprint =

    2025

  39. [47]

    2025 , eprint =

    Zhou, Huichi and Lee, Kin-Hei and Zhan, Zhonghao and Chen, Yue and Li, Zhenhao and Wang, Zhaoyang and Haddadi, Hamed and Yilmaz, Emine , journal =. 2025 , eprint =

    2025

  40. [48]

    Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , series =

    Evidence-Driven Retrieval Augmented Response Generation for Online Misinformation , author =. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , series =. 2024...

    2024

  41. [49]

    Grattafiori, Aaron and Dubey, Abhimanyu and Jauhri, Abhinav and Pandey, Abhinav and Kadian, Abhishek and Al-Dahle, Ahmad and Letman, Aiesha and Mathur, Akhil and Schelten, Alan and Vaughan, Alex and Yang, Amy and Fan, Angela and Goyal, Anirudh and Hartshorn, Anthony and Yang, ...

    2024

  42. [50]

    Jiang, Albert Q. and Sablayrolles, Alexandre and Mensch, Arthur and Bamford, Chris and Chaplot, Devendra Singh and Casas, Diego de las and Bressand, Florian and Lengyel, Gianna and Lample, Guillaume and Saulnier, Lucile and Lavaud, Lelio Renard and Lachaux, Marie-Anne and Stoc...

    2023

  43. [51]

    2021 , url =

    He, Pengcheng and Liu, Xiaodong and Gao, Jianfeng and Chen, Weizhu , booktitle =. 2021 , url =

    2021

  44. [52]

    2025 , eprint =

    Yang, An and Li, Anfeng and Yang, Baosong and Zhang, Beichen and Hui, Binyuan and Zheng, Bo and Yu, Bowen and Gao, Chang and Huang, Chengen and Lv, Chenxu and Zheng, Chujie and Liu, Dayiheng and Zhou, Fan and Huang, Fei and Hu, Feng and Ge, Hao and Wei, Haoran and Lin, Huan an...

    2025

  45. [53]

    2019 , eprint =

    Liu, Yinhan and Ott, Myle and Goyal, Naman and Du, Jingfei and Joshi, Mandar and Chen, Danqi and Levy, Omer and Lewis, Mike and Zettlemoyer, Luke and Stoyanov, Veselin , journal =. 2019 , eprint =

    2019

  46. [54]

    Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics , series =

    Soricut, Radu and Echihabi, Abdessamad , editor =. Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics , series =. 2010 , address =

    2010

  47. [55]

    Findings of the Association for Computational Linguistics: EMNLP 2025 , series =

    Wan, Yingjia and Tan, Haochen and Zhu, Xiao and Zhou, Xinyu and Li, Zhiwei and Lv, Qingsong and Sun, Changxuan and Zeng, Jiaqi and Xu, Yi and Lu, Jianqiao and Liu, Yinhong and Guo, Zhijiang , editor =. Findings of the Association for Computational Linguistics: EMNLP 2025 , ser...

    2025

  48. [56]

    Findings of the Association for Computational Linguistics: EMNLP 2025 , series =

    Rajendhran, Rishanth and Zadeh, Amir and Sarte, Matthew and Li, Chuan and Iyyer, Mohit , editor =. Findings of the Association for Computational Linguistics: EMNLP 2025 , series =. 2025 , address =

    2025

  49. [57]

    Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) , series =

    A Broad-Coverage Challenge Corpus for Sentence Understanding through Inference , author =. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) , series =. 2018 , ...

    2018

  50. [58]

    2026 , howpublished =

    2026

  51. [59]

    Trust Before Fusion:

    Eletter, Saadeldine and Aijaz, Owais and Nakov, Preslav , year =. Trust Before Fusion:

  52. [60]

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

    2024

  53. [61]

    Mingda Chen, Yang Li, Karthik Padthe, Rulin Shao, Alicia Yi Sun, Luke Zettlemoyer, Gargi Ghosh, and Wen-tau Yih. 2025. https://doi.org/10.18653/v1/2025.acl-long.548 Improving factuality with explicit working memory . In Proceedings of the 63rd Annual Meeting of the Association...

    2025 doi

  54. [62]

    Xinbang Dai, Huikang Hu, Yuncheng Hua, Jiaqi Li, Yongrui Chen, Rihui Jin, Nan Hu, and Guilin Qi. 2025. https://arxiv.org/abs/2505.17118 After retrieval, before generation: Enhancing the trustworthiness of large language models in retrieval-augmented generation . ArXiv preprint...

    2025

  55. [63]

    Liang, and Tatsunori B

    Yann Dubois, Chen Xuechen Li, Rohan Taori, Tianyi Zhang, Ishaan Gulrajani, Jimmy Ba, Carlos Guestrin, Percy S. Liang, and Tatsunori B. Hashimoto. 2023. https://proceedings.neurips.cc/paper_files/paper/2023/hash/5fc47800ee5b30b8777fdd30abcaaf3b-Abstract-Conference.html AlpacaFa...

    2023

  56. [64]

    Saadeldine Eletter, Owais Aijaz, and Preslav Nakov. 2026. Trust before fusion: QIMG -7 and source-aware resolution for polluted multimodal RAG . Manuscript

    2026

  57. [65]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Art...

    2024 arXiv

  58. [66]

    Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021. https://openreview.net/forum?id=XPZIaotutsD DeBERTa : Decoding-enhanced BERT with disentangled attention . In Proceedings of the International Conference on Learning Representations, ICLR '21

    2021

  59. [67]

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

    2023 arXiv

  60. [68]

    Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie Callan, and Graham Neubig

    Zhengbao Jiang, Frank F. Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie Callan, and Graham Neubig. 2023 b . https://doi.org/10.18653/v1/2023.emnlp-main.495 Active retrieval augmented generation . In Proceedings of the 2023 Conference on Empirical Meth...

    2023 doi

  61. [69]

    Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.741 FA ct S core: Fine-grained atomic evaluation of factual precision in long form text generatio...

    2023 doi

  62. [70]

    Abhika Mishra, Akari Asai, Vidhisha Balachandran, Yizhong Wang, Yulia Tsvetkov, Graham Neubig, and Hannaneh Hajishirzi. 2024. https://openreview.net/forum?id=dJMTn3QOWO Fine-grained hallucination detection and editing for language models . In Proceedings of the Conference on L...

    2024

  63. [71]

    OpenAI . 2026. OpenAI API model documentation. https://developers.openai.com/api/docs/models. Accessed: 2026-06-10

    2026

  64. [72]

    Yikang Pan, Liangming Pan, Wenhu Chen, Preslav Nakov, Min-Yen Kan, and William Wang. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.97 On the risk of misinformation pollution with large language models . In Findings of the Association for Computational Linguistics: EMNL...

    2023 doi

  65. [73]

    Ethan Perez, Sam Ringer, Kamile Lukosiute, Karina Nguyen, Edwin Chen, Scott Heiner, Craig Pettit, Catherine Olsson, Sandipan Kundu, Saurav Kadavath, Andy Jones, Anna Chen, Ben Mann, Brian Israel, Bryan Seethor, Cameron McKinnon, Christopher Olah, Da Yan, Daniela Amodei, and 44...

    2023 doi

  66. [74]

    Rishanth Rajendhran, Amir Zadeh, Matthew Sarte, Chuan Li, and Mohit Iyyer. 2025. https://doi.org/10.18653/v1/2025.findings-emnlp.491 V eri F ast S core: Speeding up long-form factuality evaluation . In Findings of the Association for Computational Linguistics: EMNLP 2025, EMNL...

    2025 doi

  67. [75]

    Yixiao Song, Yekyung Kim, and Mohit Iyyer. 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.552 V eri S core: Evaluating the factuality of verifiable claims in long-form text generation . In Findings of the Association for Computational Linguistics: EMNLP 2024, EMNLP '24,...

    2024 doi

  68. [76]

    Radu Soricut and Abdessamad Echihabi. 2010. https://aclanthology.org/P10-1063/ TrustRank : Inducing trust in automatic translations via ranking . In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, ACL '10, pages 612--621, Uppsala, Swede...

    2010

  69. [77]

    Weihang Su, Yichen Tang, Qingyao Ai, Zhijing Wu, and Yiqun Liu. 2024. https://doi.org/10.18653/v1/2024.acl-long.702 DRAGIN : Dynamic retrieval augmented generation based on the real-time information needs of large language models . In Proceedings of the 62nd Annual Meeting of ...

    2024 doi

  70. [78]

    Yingjia Wan, Haochen Tan, Xiao Zhu, Xinyu Zhou, Zhiwei Li, Qingsong Lv, Changxuan Sun, Jiaqi Zeng, Yi Xu, Jianqiao Lu, Yinhong Liu, and Zhijiang Guo. 2025. https://doi.org/10.18653/v1/2025.findings-emnlp.1295 F a S t F act: Faster, stronger long-form factuality evaluations in ...

    2025 doi

  71. [79]

    Fei Wang, Xingchen Wan, Ruoxi Sun, Jiefeng Chen, and Sercan O. Arik. 2025 a . https://doi.org/10.18653/v1/2025.acl-long.1476 ASTUTE RAG : Overcoming imperfect retrieval augmentation and knowledge conflicts for large language models . In Proceedings of the 63rd Annual Meeting o...

    2025 doi

  72. [80]

    Han Wang, Archiki Prasad, Elias Stengel-Eskin, and Mohit Bansal. 2025 b . https://openreview.net/forum?id=z1MHB2m3V9 Retrieval-augmented generation with conflicting evidence . In Proceedings of the Conference on Language Modeling, COLM '25

    2025

  73. [81]

    Jerry Wei, Chengrun Yang, Xinying Song, Yifeng Lu, Nathan Hu, Jie Huang, Dustin Tran, Daiyi Peng, Ruibo Liu, Da Huang, Cosmo Du, and Quoc V. Le. 2024. https://proceedings.neurips.cc/paper_files/paper/2024/hash/937ae0e83eb08d2cb8627fe1def8c751-Abstract-Conference.html Long-form...

    2024

  74. [82]

    Zhepei Wei, Wei-Lin Chen, and Yu Meng. 2025. https://openreview.net/forum?id=P1qhkp8gQT InstructRAG : Instructing retrieval-augmented generation via self-synthesized rationales . In Proceedings of the International Conference on Learning Representations, ICLR '25

    2025

  75. [83]

    Chong Xiang, Tong Wu, Zexuan Zhong, David Wagner, Danqi Chen, and Prateek Mittal. 2024. https://openreview.net/forum?id=qsEeACAJjD Certifiably robust RAG against retrieval corruption . In ICML 2024 Workshop on Next Generation of AI Safety , ICML Workshop '24

    2024

  76. [84]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others. 2025. https://arxiv.org/abs/2505.09388 Qw...

    2025 arXiv

  77. [85]

    Ori Yoran, Tomer Wolfson, Ori Ram, and Jonathan Berant. 2024. https://openreview.net/forum?id=ZS4m74kZpH Making retrieval-augmented language models robust to irrelevant context . In Proceedings of the International Conference on Learning Representations, ICLR '24

    2024

  78. [86]

    Zhenrui Yue, Huimin Zeng, Yimeng Lu, Lanyu Shang, Yang Zhang, and Dong Wang. 2024. https://doi.org/10.18653/v1/2024.naacl-long.313 Evidence-driven retrieval augmented response generation for online misinformation . In Proceedings of the 2024 Conference of the North American Ch...

    2024 doi

  79. [87]

    Linda Zeng, Rithwik Gupta, Divij Motwani, Yi Zhang, and Diji Yang. 2025. https://openreview.net/forum?id=R4MeWTeVej Worse than zero-shot? a fact-checking dataset for evaluating the robustness of RAG against misleading retrievals . In The Thirty-ninth Annual Conference on Neura...

    2025

  80. [88]

    Huichi Zhou, Kin-Hei Lee, Zhonghao Zhan, Yue Chen, Zhenhao Li, Zhaoyang Wang, Hamed Haddadi, and Emine Yilmaz. 2025. https://arxiv.org/abs/2501.00879 TrustRAG : Enhancing robustness and trustworthiness in RAG . ArXiv preprint, arXiv:2501.00879

    2025 arXiv

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.