Pith. sign in

REVIEW 3 major objections 5 minor 37 references

Current LLMs update beliefs as much for unreliable sources as for reliable ones, and as much for claims that move them away from the truth as toward it.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 13:19 UTC pith:UU5GLA3O

load-bearing objection A genuinely useful benchmark for source/truth discernment in LLMs; the near-chance headline is plausible but needs one stability analysis before it is fully secure. the 3 major comments →

arxiv 2607.19355 v1 pith:UU5GLA3O submitted 2026-05-22 cs.AI cs.CLcs.CY

Information Discernment in Large Language Models

Joshua Ashkinaze , Laura Kurek , Alina Faisal , Tongyuan Miao , Mariam Joseph , Ceren Budak , Eric Gilbert This is my paper
classification cs.AI cs.CLcs.CY
keywords information discernmentsource discernmenttruth discernmentknowledge conflictsretrieval-augmented generationLLM belief updatingmisinformationbenchmark
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper asks whether LLMs treat external information the way a careful reader would: updating more for reliable sources and more when a claim moves them closer to the truth. Across 13 models and nearly 670K trials, it finds models do neither: update magnitude correlates near zero with source reliability and with truth improvement, and models actually weight source popularity twice as strongly as reliability. A pre-registered user study of 299 LLM users confirms that people endorse these norms and say violations would reduce their trust and usage intent. The paper releases a benchmark for measuring this 'information discernment' and shows that simple prompt modifications can partially fix the deficit.

Core claim

The paper's central claim is that current LLMs systematically fail at information discernment: when told '[Source] says the answer is C', models shift their numeric answers just as much for a fringe outlet as for a highly reliable one, and just as much when the claim moves them toward the ground truth as when it moves them away. The exception is 'correct defense': models usually stick with an initially correct answer. Bigger and newer models show better truth discernment but no better source discernment, and models integrate external information best on datasets where their priors are already accurate. The failures are measured with Spearman correlations between update magnitude and reliabil

What carries the argument

The experimental machinery is a before/after probe: ask a model a numeric question to get a prior P1, present a claim C = T × f from a source S with known reliability and popularity, elicit posterior P2, and define update magnitude Δ = |P1 − P2|. Source discernment is the Spearman correlation ρ(Δ, reliability); truth discernment is ρ(Δ, I) where I = |P1 − T| − |C − T| is how much the claim improves on the prior. The framework also defines a 'reliability propensity' probability and a 'correct defense' stickiness measure, and a decomposition of low source discernment into mis-estimating reliability vs failing to apply it (LLMs estimate reliability moderately, ρ=0.41, but applying their own est

Load-bearing premise

The experiment treats the single numeric token a model emits when asked 'Give a numeric answer and nothing else' as its genuine belief, before and after the claim; if those tokens are noisy or arbitrary under knowledge conflict, the near-zero discernment correlations would be an artifact of the probe rather than proof that models fail to discriminate.

What would settle it

Run a version of the experiment where prior and posterior are elicited as 10 independent samples (or as full probability distributions) for a subset of trials; if average update magnitude then correlates strongly with source reliability (say ρ > 0.2), the near-zero correlation in the single-token probe would be an artifact, not a genuine absence of source discernment. Conversely, finding any unprompted model that scores above 0.2 on source discernment across datasets would directly refute the claim that all models perform near chance.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim is right, retrieval-augmented systems built on today's LLMs can be led astray by low-quality sources, since belief updating does not track source quality and instead tracks popularity.
  • Scaling model size and recency will not by itself fix source discernment: the paper finds no improvement on that axis across matched pairs of bigger/newer models.
  • Simple inference-time instructions—explicitly rating source reliability, reminding the model to defend a correct prior, chain-of-thought—improve facets of discernment and are cheap, actionable mitigations.
  • External knowledge integration is most effective where it is least needed: models update well from evidence only on datasets where their priors are already accurate, limiting how much web augmentation can equalize knowledge gaps.
  • If the user study's stated norms reflect real adoption behavior, low discernment is likely to erode trust and usage of LLM-based search as these systems replace traditional search.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A direct testable implication is that near-zero source discernment is not a hard ceiling: since models can rate reliability moderately well when explicitly asked, fine-tuning or prompting strategies that bind reliability judgments to update sizes may close the gap.
  • The probe uses a single numeric token as the model's belief; if that token is noisy under knowledge conflict, the true belief-update distributions might be more discerning than reported—re-running with full probability distributions or repeated sampling would settle this.
  • The popularity-over-reliability result suggests a 'trust equals fame' shortcut learned from pretraining; if so, interventions that break the popularity–reliability confound in training data (not just prompts) may be needed to fix source discernment.
  • Because the design isolates belief updating from retrieval, the results bound the best-case performance of RAG: retrieval improvements alone cannot correct failures at the updating stage.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a normative framework, 'information discernment', with three axioms: source discernment (update more for higher-reliability sources), truth discernment (update more when a claim moves the prior closer to ground truth), and correct defense (do not abandon a correct prior). The authors introduce Learn2Discern (L2D), a benchmark combining numeric-answer questions from existing and new datasets with 132 real web sources stratified by reliability and popularity, and claims generated by multiplicative perturbations of the ground truth. They report experiments across 13 LLMs and nearly 670K trials, finding near-chance source and truth discernment, stronger sensitivity to popularity than reliability, no consistent improvement in source discernment with scale/recency, and some benefit from inference-time prompting interventions. A pre-registered, quota-matched user study (n=299) is used to validate that LLM users endorse the axioms and report that violations reduce trust and usage intent. The paper releases the dataset, prompts, and survey.

Significance. If the central empirical claim holds, the paper identifies a practically important blind spot in current LLMs' use of external information, with direct relevance to RAG and misinformation propagation. The contribution is potentially significant: the benchmark is large and diverse, the measurement framework is simple and interpretable, the user study anchors the metrics in stated user preferences, and the data are released for reuse. The scale (13 models, 670K trials, 95% CIs), the pre-registration of the user study, and the explicit attention to contamination via novel GSS/World Bank items are clear strengths. However, the headline 'near chance' conclusion rests on an unvalidated assumption that a single numeric token elicited from an LLM is a stable, faithful point estimate of its belief. The absence of a test-retest or calibration control means the central null result is not yet separable from measurement noise. The paper also makes several categorical claims that are not fully supported by its own tables.

major comments (3)
  1. [§5 (Algorithm 1), §6 (Δ definition), Appendix F.1–F.2] The core metrics all derive from Δ = |P1 − P2|, where P1 and P2 are single numeric tokens elicited by 'Give a numeric answer and nothing else.' The paper gives no test-retest reliability, no calibration evidence, and no control condition in which P1 and P2 are elicited on the same question without a source/claim. On uncertain numeric items, output instability will contribute to Δ; this attenuates ρ(Δ, S) and ρ(Δ, I) toward zero and also corrupts the taxonomy in Table 3 (a noisy P2 would turn 'correct hold' into 'sycophantic cave'). The Reliable-prompt control (Figure 6, ρ(Δ, R_LLM)=0.46) shows that the probe can detect a signal under a different prompting regime, but it does not validate the baseline P1/P2. Please add a no-information stability control and, if possible, a calibration check, or apply a measurement-error correction before reporting near-chance discernment.
  2. [§9, Table 2] The claim that 'all models perform near chance across all metrics except Correct Defense' is contradicted by Table 2. For example, Claude-3.5-Sonnet has truth discernment −0.25, Gemini-2.0-Flash −0.07, and GPT-3.5-Turbo −0.043. A Spearman correlation of −0.25 is not 'near chance'; it indicates systematic updating away from the truth. Conversely, source discernment values around 0.02–0.06 may be statistically nonzero at N≈670K, so 'near chance' needs a formal justification. Please report equivalence tests (e.g., TOST with pre-specified bounds) or reframe the headline as 'small / inconsistent' rather than 'near chance'.
  3. [§9, Appendix J (Tables 11–12)] The text states that bigger and newer models 'did not improve on reliability-related metrics (source discernment and reliability propensity).' This is not supported by the authors' own tables. Table 11 shows gpt-4o-mini→gpt-4o source discernment +0.028 (p<.001) and gpt-4.1-mini→gpt-4.1 +0.022 (p<.001); Table 12 shows gpt-3.5-turbo→gpt-4o +0.032 (p<.001). The pattern is mixed rather than absent. If the intended claim is 'no *consistent* or *practically meaningful* improvement', that should be stated with an effect-size criterion, and the abstract's 'not source discernment' should be softened accordingly.
minor comments (5)
  1. [Appendix F.3] The CoT prompt contains a template artifact: 'FINAL_ANSWER: Give a numeric answer and nothing else.' is fine, but the RATIONALE line includes '$produce' which appears to be a LaTeX/formatting error. Please fix.
  2. [Figure 4] The captions are garbled: the same '(a) OECD vs Non-OECD' and '(b) Pre- vs Post-Internet' labels appear under both panels, while the text refers to 'update quality ratio' and 'median relative prior error.' Relabel the subfigures so they match the text.
  3. [Table 3] The 'Direction' column uses Good/Neutral/Bad but the definitions of 'Appropriate Update' vs 'Inappropriate Update' depend on the prior being wrong; this is clear in the descriptions, but adding a sentence on how ties are treated (e.g., when P2 equals P1 within tolerance) would improve precision.
  4. [Appendix I.2] Typo: 'Proceedure' should be 'Procedure'. Also, the survey wave handling for Source Discernment (Wave 2 only) is described; consider adding one sentence in the main text to preempt concerns about using a subset of waves for one of the three axioms.
  5. [§10, Limitations] The paper already acknowledges that source credibility may be claim-dependent; this is important because the source-discernment metric uses a claim-independent PC1 reliability label. I suggest adding a sentence in the discussion noting that the near-zero source discernment should be interpreted relative to this global, topic-independent notion of reliability.

Circularity Check

0 steps flagged

No significant circularity: empirical metrics are direct correlations against external labels, with only an explicitly hypothetical oracle simulation.

full rationale

The paper's central empirical metrics (source discernment rho(Delta, S), truth discernment rho(Delta, I), and the corresponding propensities) are direct Spearman correlations between elicited update magnitudes Delta=|P1-P2| and externally defined quantities: PC1 source reliability from Lin et al. and improvement I=|P1-T|-|C-T| computed from ground truth T and perturbation C=T*F. No parameter is fitted to the same data that is then used to test the headline claims. The only optimization in the paper is the Appendix H oracle simulation, which fits the trust-weight slope a to a target hypothetical rho(alpha, I); the paper explicitly labels this 'Hypothetical' and 'Oracle', and it is not used to fit any empirical constant. The three axioms are author-postulated normative conditions validated by a separate pre-registered user study, not derived from model behavior. The Reliable-prompt decomposition (rho(Delta, R_LLM)=0.46 vs. rho(Delta, PC1)=0.089) is an explicit sensitivity analysis, not a claimed prediction. The only self-citation (ref [1], Ashkinaze et al., Deep Value Benchmark) is used only to justify paired model-size comparisons and is not load-bearing for the paper's conclusions. The single-token P1/P2 elicitation is a measurement-validity concern (no test-retest or calibration evidence is reported), but that is a correctness-risk issue rather than a circular reduction of the derivation to its inputs. No specific equation or fitted parameter was found that makes a stated prediction equivalent to an input by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The central claims rest on three families of assumptions: (1) that numeric output equals belief; (2) that PC1 domain-reliability labels are valid, claim-independent source quality signals for arbitrary news-source/question pairings; and (3) that multiplying ground truth by {1/3,1/2,1,2,3} creates a meaningful falsity continuum. The benchmark also has hand-set design constants (perturbation factors, reliability cut at 0.7, popularity tiers, winsorization) that shape the measured correlations. No physical entity is introduced; the 'information discernment' construct is new but externally anchored by the user survey.

free parameters (4)
  • Perturbation factors F = {1/3, 1/2, 1, 2, 3}
    Claims are generated as C = T x F; the chosen scales define 'moderately' vs 'highly' false claims (Section 5) and directly shape the improvement distribution used to compute truth discernment.
  • Reliability cut and popularity tier cutoffs = reliability split at 0.7; 3 popularity tiers from tri-modal distribution
    Used to stratify the 132 sources into 6 strata (Section 3, Appendix B); changes here alter which sources are 'high' vs 'low' reliability and thus the source-discernment measurement.
  • Winsorization thresholds = 5th and 95th percentiles
    Update magnitudes are winsorized at 5/95% (Appendix E), attenuating extreme updates and affecting the Spearman and pairwise metrics.
  • Oracle alpha (a) = minimize_scalar over a in (-20,20) per target rho
    Appendix H fits alpha(a) to hit a target Spearman correlation in the hypothetical error-reduction simulation; this is a simulation knob, not a data fit.
axioms (4)
  • domain assumption PC1 source reliability scores from Lin et al. (2023) are valid, claim-independent labels for source reliability.
    Source discernment is defined as correlation of update magnitude with PC1; if reliability is claim-dependent or PC1 is not what models should use, the benchmark's source-discernment conclusion is affected. The paper acknowledges this in Limitations (Section 10).
  • domain assumption A single numeric output elicited by 'Give a numeric answer and nothing else' is a faithful measure of the model's belief.
    Prior and posterior beliefs P1, P2 are read directly from model outputs (Algorithm 1, Section 5); no calibration or test-retest analysis is given, so |P1-P2| may conflate belief revision with output noise.
  • domain assumption Perturbing ground truth as C = T x F creates claims whose falsity is meaningful and whose improvement I = |P1-T| - |C-T| is the correct normative quantity for truth discernment.
    The claim generation assumes scalar multiplication preserves a continuum of falsity; the authors justify it as similar to FEVER (Section 10), but it is an assumption about how LLMs should update.
  • ad hoc to paper The three proposed axioms are the correct normative criteria for information discernment.
    SD, TD, and Correct Defense are introduced in Section 6; the user study (n=299) provides external endorsement, but they are not derived from prior theory and other reasonable axioms could be chosen.
invented entities (1)
  • Information discernment construct (source discernment + truth discernment) independent evidence
    purpose: Normative evaluation framework for LLM belief updating under knowledge conflicts
    It is a new conceptual entity; the user study provides external evidence that users endorse the axioms and that violations affect trust and usage intent, giving a falsifiable handle outside the model benchmark.

pith-pipeline@v1.3.0-alltime-deepseek · 31710 in / 14862 out tokens · 158674 ms · 2026-08-02T13:19:07.534499+00:00 · methodology

0 comments
read the original abstract

LLMs are increasingly used with external knowledge sources like the internet. Do they weigh information appropriately -- updating more for reliable sources (source discernment) and more when claims bring priors closer to the truth (truth discernment)? We formalize this as information discernment and introduce Learn2Discern (L2D), an experimental framework and benchmark grounded in three normative axioms with interpretable metrics. To establish external validity, a pre-registered, quota-matched user study (n=299) confirms that real LLM users endorse all three axioms and report that violations reduce their trust and usage intent. Across 13 models and nearly 670K trials, we find consistent failures across both dimensions: models perform near chance on source and truth discernment, rely on source popularity twice as much as source reliability, and update roughly equally whether a claim improves or worsens their position relative to the ground truth. Models integrate external knowledge most effectively on datasets where their priors are already the most accurate. Newer and larger models improve truth discernment but not source discernment, a blind spot that model complexity does not address. We identify simple inference-time interventions that improve both forms of discernment. We release our dataset and survey as a testbed for a core alignment property that scales in importance as LLMs replace traditional search.

Figures

Figures reproduced from arXiv: 2607.19355 by Alina Faisal, Ceren Budak, Eric Gilbert, Joshua Ashkinaze, Laura Kurek, Mariam Joseph, Tongyuan Miao.

Figure 1
Figure 1. Figure 1: Results from user study [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Metrics colored from worst possible (blue) to best possible (orange). Dashed boxes indicate [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Left: models are more sensitive to source popularity than reliability. Right: higher truth [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: External evidence integration performance by dataset characteristics. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Discernment metrics by model and prompt with 95% CIs. [PITH_FULL_IMAGE:figures/full_fig_p028_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Source discernment under the Reliable prompt condition: observed source discernment [PITH_FULL_IMAGE:figures/full_fig_p028_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 3 canonical work pages

  1. [1]

    Deep Value Benchmark: Measuring Whether Models Generalize Deep Values or Shallow Preferences

    Joshua Ashkinaze, Hua Shen, Sai Avula, Eric Gilbert, and Ceren Budak. Deep Value Benchmark: Measuring Whether Models Generalize Deep Values or Shallow Preferences. October 2025. URL:https://openreview.net/forum?id=bzxlOyjWbU

    2025

  2. [2]

    Kevin Aslett, Zeve Sanderson, William Godel, Nathaniel Persily, Jonathan Nagler, and Joshua A. Tucker. Online searches to evaluate misinformation can increase its perceived veracity.Na- ture, 625(7995):548–556, 2024. URL: https://www.ncbi.nlm.nih.gov/pmc/articles/ PMC10794132/,doi:10.1038/s41586-023-06883-y

    work page doi:10.1038/s41586-023-06883-y 2024

  3. [3]

    Mohamad Adam Bujang. An elaboration on sample size determination for correlations based on effect sizes and confidence interval width: a guide for researchers.Restorative Dentistry & Endodontics, 49(2):e21, May 2024. URL: https://www.ncbi.nlm.nih.gov/pmc/articles/ PMC11148401/,doi:10.5395/rde.2024.49.e21

    work page doi:10.5395/rde.2024.49.e21 2024

  4. [4]

    MAIN- RAG: Multi-Agent Filtering Retrieval-Augmented Generation

    Chia-Yuan Chang, Zhimeng Jiang, Vineeth Rakesh, Menghai Pan, Chin-Chia Michael Yeh, Guanchu Wang, Mingzhi Hu, Zhichao Xu, Yan Zheng, Mahashweta Das, and Na Zou. MAIN- RAG: Multi-Agent Filtering Retrieval-Augmented Generation. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar, editors,Proceedings of the 63rd Annual Meeting of t...

    2025

  5. [5]

    Combating misinformation in the age of LLMs: Opportunities and challenges.AI Magazine, 45(3):354–368, 2024

    Canyu Chen and Kai Shu. Combating misinformation in the age of LLMs: Opportunities and challenges.AI Magazine, 45(3):354–368, 2024. URL: https://onlinelibrary.wiley.com/ doi/abs/10.1002/aaai.12188,doi:10.1002/aaai.12188

    doi:10.1002/aaai.12188 2024

  6. [6]

    A Survey on RAG Meeting LLMs: Towards Retrieval-Augmented Large Language Models

    Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. A Survey on RAG Meeting LLMs: Towards Retrieval-Augmented Large Language Models. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD ’24, pages 6491–6501, New York, NY , USA, 2024. Association for Computing Machiner...

    arXiv 2024

  7. [7]

    WikiContradict: A Benchmark for Evaluating LLMs on Real-World Knowledge Conflicts from Wikipedia, 2024

    Yufang Hou, Alessandra Pascale, Javier Carnerero-Cano, Tigran Tchrakian, Radu Mari- nescu, Elizabeth Daly, Inkit Padhi, and Prasanna Sattigeri. WikiContradict: A Benchmark for Evaluating LLMs on Real-World Knowledge Conflicts from Wikipedia, 2024. URL: https://arxiv.org/abs/2406.13805

    Pith/arXiv arXiv 2024

  8. [8]

    Weld, and Luke Zettlemoyer

    Mandar Joshi, Eunsol Choi, Daniel S. Weld, and Luke Zettlemoyer. TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension, May 2017. URL: http://arxiv.org/abs/1705.03551,doi:10.48550/arXiv.1705.03551

    Pith/arXiv arXiv doi:10.48550/arxiv.1705.03551 2017

  9. [9]

    Internet-Augmented Dialogue Generation, July 2021

    Mojtaba Komeili, Kurt Shuster, and Jason Weston. Internet-Augmented Dialogue Generation, July 2021. URL: http://arxiv.org/abs/2107.07566, doi:10.48550/arXiv.2107.07566

    Pith/arXiv arXiv doi:10.48550/arxiv.2107.07566 2021

  10. [10]

    Studying Large Language Model Behaviors Under Context-Memory Conflicts With Real Documents

    Evgenii Kortukov, Alexander Rubinstein, Elisa Nguyen, and Seong Joon Oh. Studying Large Language Model Behaviors Under Context-Memory Conflicts With Real Documents. 2024

    2024

  11. [11]

    Learning Question Classifiers

    Xin Li and Dan Roth. Learning Question Classifiers. InCOLING 2002: The 19th Interna- tional Conference on Computational Linguistics, 2002. URL: https://www.aclweb.org/ anthology/C02-1150

    2002

  12. [12]

    Birds have four legs?! Nu- merSense: Probing Numerical Commonsense Knowledge of Pre-trained Language Mod- els, September 2020

    Bill Yuchen Lin, Seyeon Lee, Rahul Khanna, and Xiang Ren. Birds have four legs?! Nu- merSense: Probing Numerical Commonsense Knowledge of Pre-trained Language Mod- els, September 2020. arXiv:2005.00683 [cs]. URL: http://arxiv.org/abs/2005.00683, doi:10.48550/arXiv.2005.00683

    Pith/arXiv arXiv doi:10.48550/arxiv.2005.00683 2020

  13. [13]

    Reducing misinformation sharing at scale using digital accuracy prompt ads, February 2024

    Hause Lin, Haritz Garro, Nils Wernerfelt, Jesse C Shore, Adam Hughes, Daniel Deisenroth, Nathaniel Barr, Adam J Berinsky, Dean Eckles, Gordon Pennycook, and David G Rand. Reducing misinformation sharing at scale using digital accuracy prompt ads, February 2024. URL:https://osf.io/preprints/psyarxiv/u8anb_v1/,doi:10.31234/osf.io/u8anb

    work page doi:10.31234/osf.io/u8anb 2024

  14. [14]

    High level of correspondence across different news domain quality rating sets.PNAS Nexus, 2(9):pgad286, September 2023

    Hause Lin, Jana Lasser, Stephan Lewandowsky, Rocky Cole, Andrew Gully, David G Rand, and Gordon Pennycook. High level of correspondence across different news domain quality rating sets.PNAS Nexus, 2(9):pgad286, September 2023. doi:10.1093/pnasnexus/pgad286. 10

    doi:10.1093/pnasnexus/pgad286 2023

  15. [15]

    The Moon is Made of Marshmallows

    Yifei Ming, Senthil Purushwalkam, Shrey Pandit, Zixuan Ke, Xuan-Phi Nguyen, Caiming Xiong, and Shafiq Joty. FaithEval: Can Your Language Model Stay Faithful to Context, Even If "The Moon is Made of Marshmallows". In Y . Yue, A. Garg, N. Peng, F. Sha, and R. Yu, editors,International Conference on Learning Representations, volume 2025, pages 29430–29456, 2...

    2025

  16. [16]

    General Social Survey, 2025

    NORC. General Social Survey, 2025. URL:https://gss.norc.org/

    2025

  17. [17]

    Attacking Open- domain Question Answering by Injecting Misinformation

    Liangming Pan, Wenhu Chen, Min-Yen Kan, and William Yang Wang. Attacking Open- domain Question Answering by Injecting Misinformation. In Jong C. Park, Yuki Arase, Baotian Hu, Wei Lu, Derry Wijaya, Ayu Purwarianti, and Adila Alfa Krisnadhi, editors, Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference...

    2023

  18. [18]

    On the Risk of Misinformation Pollution with Large Language Models, October 2023

    Yikang Pan, Liangming Pan, Wenhu Chen, Preslav Nakov, Min-Yen Kan, and William Yang Wang. On the Risk of Misinformation Pollution with Large Language Models, October 2023. URL:http://arxiv.org/abs/2305.13661,doi:10.48550/arXiv.2305.13661

    Pith/arXiv arXiv doi:10.48550/arxiv.2305.13661 2023

  19. [19]

    Who‘s Who: Large Lan- guage Models Meet Knowledge Conflicts in Practice

    Quang Hieu Pham, Hoang Ngo, Anh Tuan Luu, and Dat Quoc Nguyen. Who‘s Who: Large Lan- guage Models Meet Knowledge Conflicts in Practice. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen, editors,Findings of the Association for Computational Linguistics: EMNLP 2024, pages 10142–10151, Miami, Florida, USA, November 2024. Association for Com- putational L...

    doi:10.18653/v1/2024.findings-emnlp.593 2024

  20. [20]

    Evaluation of Reliability Criteria for News Publishers with Large Language Models, February 2025

    Manuel Pratelli, John Bianchi, Fabio Pinelli, and Marinella Petrocchi. Evaluation of Reliability Criteria for News Publishers with Large Language Models, February 2025. URL: http: //arxiv.org/abs/2412.15896,doi:10.1145/3717867.3717924

    Pith/arXiv arXiv 2025

  21. [21]

    NLP Evaluation in trouble: On the Need to Measure LLM Data Contamination for each Benchmark, October 2023

    Oscar Sainz, Jon Ander Campos, Iker García-Ferrero, Julen Etxaniz, Oier Lopez de Lacalle, and Eneko Agirre. NLP Evaluation in trouble: On the Need to Measure LLM Data Contamination for each Benchmark, October 2023. URL: http://arxiv.org/abs/2310.18018, doi:10. 48550/arXiv.2310.18018

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.18018 2023

  22. [22]

    The spread of low-credibility content by social bots.Nature Communications, 9(1):4787, November 2018

    Chengcheng Shao, Giovanni Luca Ciampaglia, Onur Varol, Kai-Cheng Yang, Alessandro Flammini, and Filippo Menczer. The spread of low-credibility content by social bots.Nature Communications, 9(1):4787, November 2018. URL: https://www.nature.com/articles/ s41467-018-06930-7,doi:10.1038/s41467-018-06930-7

    doi:10.1038/s41467-018-06930-7 2018

  23. [23]

    Bowman, Newton Cheng, Esin Durmus, Zac Hatfield-Dodds, Scott R

    Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Amanda Askell, Samuel R. Bowman, Newton Cheng, Esin Durmus, Zac Hatfield-Dodds, Scott R. Johnston, Shauna Kravec, Timothy Maxwell, Sam McCandlish, Kamal Ndousse, Oliver Rausch, Nicholas Schiefer, Da Yan, Miranda Zhang, and Ethan Perez. Towards Understanding Sycophancy in Language Models, October 202...

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.13548 2023

  24. [24]

    Mayer, and Padhraic Smyth

    Mark Steyvers, Heliodoro Tejeda, Aakriti Kumar, Catarina Belem, Sheer Karny, Xinyue Hu, Lukas W. Mayer, and Padhraic Smyth. What large language models know and what people think they know.Nature Machine Intelligence, 7(2):221–231, February 2025. URL: https:// www.nature.com/articles/s42256-024-00976-7,doi:10.1038/s42256-024-00976-7

    doi:10.1038/s42256-024-00976-7 2025

  25. [25]

    FEVER: a large-scale dataset for Fact Extraction and VERification, December 2018

    James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. FEVER: a large-scale dataset for Fact Extraction and VERification, December 2018. arXiv:1803.05355 [cs]. URL:http://arxiv.org/abs/1803.05355,doi:10.48550/arXiv.1803.05355

    Pith/arXiv arXiv doi:10.48550/arxiv.1803.05355 2018

  26. [26]

    Astute RAG: Overcom- ing Imperfect Retrieval Augmentation and Knowledge Conflicts for Large Language Models, May 2025

    Fei Wang, Xingchen Wan, Ruoxi Sun, Jiefeng Chen, and Sercan Ö Arık. Astute RAG: Overcom- ing Imperfect Retrieval Augmentation and Knowledge Conflicts for Large Language Models, May 2025. URL: http://arxiv.org/abs/2410.07176, doi:10.48550/arXiv.2410.07176

    Pith/arXiv arXiv doi:10.48550/arxiv.2410.07176 2025

  27. [27]

    ClashEval: Quantifying the tug-of-war between an LLM’s internal prior and external evidence, 2025

    Kevin Wu, Eric Wu, and James Zou. ClashEval: Quantifying the tug-of-war between an LLM’s internal prior and external evidence, 2025. URL:https://arxiv.org/abs/2404.10198

    Pith/arXiv arXiv 2025

  28. [28]

    Lin, Shujian Yang, Tianqi Zhang, Weiyan Shi, Tianwei Zhang, Zhixuan Fang, Wei Xu, and Han Qiu

    Rongwu Xu, Brian S. Lin, Shujian Yang, Tianqi Zhang, Weiyan Shi, Tianwei Zhang, Zhixuan Fang, Wei Xu, and Han Qiu. The Earth is Flat because...: Investigating LLMs’ Belief towards 11 Misinformation via Persuasive Conversation, May 2024. URL: http://arxiv.org/abs/2312. 09085,doi:10.48550/arXiv.2312.09085

    Pith/arXiv arXiv doi:10.48550/arxiv.2312.09085 2024

  29. [29]

    Knowledge Conflicts for LLMs: A Survey

    Rongwu Xu, Zehan Qi, Zhijiang Guo, Cunxiang Wang, Hongru Wang, Yue Zhang, and Wei Xu. Knowledge Conflicts for LLMs: A Survey. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen, editors,Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 8541–8565, Miami, Florida, USA, November 2024. Association for Computationa...

    doi:10.18653/v1/2024.emnlp-main.486 2024

  30. [30]

    Accuracy and Political Bias of News Source Credibility Ratings by Large Language Models, February 2025

    Kai-Cheng Yang and Filippo Menczer. Accuracy and Political Bias of News Source Credibility Ratings by Large Language Models, February 2025. URL: http://arxiv.org/abs/2304. 00228,doi:10.1145/3717867.3717903

    arXiv 2025

  31. [31]

    Benchmarking and Defending Against Indirect Prompt Injection Attacks on Large Language Models, January 2025

    Jingwei Yi, Yueqi Xie, Bin Zhu, Emre Kiciman, Guangzhong Sun, Xing Xie, and Fangzhao Wu. Benchmarking and Defending Against Indirect Prompt Injection Attacks on Large Language Models, January 2025. URL: http://arxiv.org/abs/2312.14197, doi:10.1145/3690624. 3709179

    Pith/arXiv arXiv 2025

  32. [32]

    twenty-two

    Wenxuan Zhou, Sheng Zhang, Hoifung Poon, and Muhao Chen. Context-faithful Prompting for Large Language Models, October 2023. URL: http://arxiv.org/abs/2303.11315, doi:10.48550/arXiv.2303.11315. 12 Appendix Table of Contents A Data Processing Details 14 A.1 Data Cleaning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 A.2 Commo...

    Pith/arXiv arXiv doi:10.48550/arxiv.2303.11315 2023

  33. [33]

    Your prior belief was your previous answer

  34. [34]

    You've received new evidence from [SOURCE]

  35. [35]

    Consider the reliability of [SOURCE] as the strength of this evidence

  36. [36]

    commitment check

    Update your belief proportionally to [SOURCE]'s reliability. For highly reliable sources, update significantly. For moderately reliable sources, update moderately. For less reliable sources, update minimally. Follow the following format: BAYESIAN UPDATING: Applying the BAYESIAN UPDATING PROCEDURE, we... FINAL_ANSWER: Give a numeric answer and nothing else...

    1948

  37. [37]

    We obtained informed consent before participants continued

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or ...