acceptodds
Under review as a conference paper at ICLR 2027

Choosing a Side is not Enough: Epistemic Transparency Under Knowledge Conflict

Abstract

LLMs are increasingly used in professional knowledge work, where their outputs can influence decisions with real-world consequences. When a model encounters information that contradicts its prior knowledge, it should communicate this conflict transparently to the user. Existing work on knowledge conflicts focuses on which source, parametric or contextual, a model chooses, rather than whether models disclose the presence of the conflict. Moreover, existing work primarily studies multiple-choice questions under simplified counterfactual evidence generation schemes. In this work, we address the first gap by scoring model responses against a taxonomy of epistemic transparency, rather than measuring accuracy alone. We address the second gap by proposing a general framework for generating knowledge conflicts grounded in advanced domain knowledge, which we instantiate in this work as a set of real-world tasks based on counterfactual evidence. These contributions allow us to measure epistemic transparency under realistic knowledge conflicts, as well as the impact of the format and plausibility of the evidence, and the real-world stakes of the task. We find substantial differences in transparency across five evaluated frontier models, with none fully matching our definition of optimal, risk-sensitive epistemic alignment. We hope that our work will ground future discussion of epistemic transparency in real-world applications of LLMs, and help establish a risk framework for epistemic transparency across domains and deployment scenarios.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Epistemic transparency, LLM agency, Knowledge conflict, Human-computer interaction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_fzzjOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

-studies epistemic transparency under knowledge conflict: when contextual evidence contradicts knowledge that a model appears to already have. -proposes a taxonomy of behaviors beyond answer accuracy -builds a conflict-generation pipeline from advanced-domain questions, varies evidence format and task stakes, and evaluates five frontier models in both multiple-choice and more realistic task settings.

Critical strengths

moving from “which source did the model follow?” to “did the model make the conflict visible to the user?” is a key shift, especially for real deployments. I consider it a key strength.

experimental setup is broader than most prior work in this area. The authors go beyond MCQs, vary stakes and document format, and evaluate several models with repeated samples, which gives the paper real empirical substance.

Critical weaknesses

the evaluation stack largely depends on LLM judges, including conflict scoring and plausibility scoring, and some plausibility analyses are based on the same family of models being studied. This is not fatal, but it limits how hard I can lean on the fine-grained empirical conclusions.

dataset is still fairly small and heavily synthetic. The final MMLU-based set appears to come from 86 retained questions, and the paper does not provide a strong human-validation study showing that the generated conflicts and the taxonomy labels match expert judgment at a meaningful scale.

Questions for authors

Can you add a human evaluation on a nontrivial subset of items for the main taxonomy labels, especially conflict detection and attribution explicitness?

How sensitive are the main trends to judge choice?

2 comments

We appreciate the reviewer’s recognition of the shift our work represents, and their feedback has led to a revision that we believe meaningfully strengthens the paper’s empirical foundations.

Can you add a human evaluation on a nontrivial subset of items for the main taxonomy labels, especially conflict detection and attribution explicitness?

We ran a small human audit of the LLM judge (9 auditors scoring 5-6 model responses each, see full results in Appendix B.5). On conflict detection and compliance, the two central dimensions for our main results, annotators agree strongly with each other (Krippendorff's α = 1.00 and 0.81) and the judge closely matches the human majority (Cohen's κ = 0.72 and 0.77); the same holds for correction (α = 0.82, κ = 0.88). Agreement is weaker on attribution explicitness (α = 0.03) and reader agency preservation (α = 0.17), and we have too few refusals to evaluate propose-alternative robustly. Our interpretation is that the weaker agreement reflects these behaviors being more subjective and harder to define. The goal of this work is to show that epistemic transparency can be measured under realistic conflicts, rather than to rank models precisely on every dimension. Defining attribution and reader agency more precisely, ideally with subject-matter experts, is left for future work.

How sensitive are the main trends to judge choice?

We re-scored our 23 human-validated responses with a second judge from a different model family (GPT-5.4) using the identical prompt, taxonomy and central-claim hint. The two judges agree closely on conflict detection, compliance and correction (Cohen's κ between 0.83 and 0.91). Agreement is lower on attribution explicitness and reader agency preservation (κ of about 0.5 to 0.6). These are the same two dimensions where human annotators also agreed least, and we already flag results on them as more subjective. The sample is small, so we see this as a consistency check rather than a precise measure of reliability. We have added the full comparison to Appendix B.6.

I appreciate the response. I will maintain my positive review.

Reviewer_Q9T3OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how LLMs communicate conflicts between external documents and their own knowledge. It proposes a six-dimensional taxonomy of epistemic transparency and a pipeline for generating counterfactual documents from MMLU-Pro across three domains, document formats, and task stakes. Five frontier models are evaluated using an LLM judge. Results show that models often follow conflicting evidence without acknowledging the conflict. Some models recognize conflicts in their reasoning but fail to communicate them to users. The study also finds that higher evidence plausibility is associated with lower conflict detection, and no model consistently demonstrates the expected risk-sensitive behavior across all conditions.

Critical strengths
  1. The paper examines whether LLMs disclose knowledge conflicts rather than simply choosing the correct answer. This provides useful insights into how models handle and communicate conflicting information.
  2. The paper verifies models' prior knowledge before introducing conflicts and evaluates five models across different domains, document formats, and task stakes. This supports a more comprehensive evaluation of model behavior.
Critical weaknesses
  1. I’m not convinced that the evaluation metrics are sufficiently reliable. The paper uses an LLM judge to evaluate all six transparency dimensions, but I could not find any human validation or agreement statistics. How can we be sure that subjective metrics such as "reader agency preservation" and "attribution explicitness" are evaluated consistently? Furthermore, the plausibility scores are provided by the same models being evaluated. I’m concerned that this could introduce circularity and weaken the reliability of the main findings.
  2. Why do the authors assume that lower compliance indicates better risk-sensitive alignment? If a model completes a task while correcting the incorrect information, shouldn't this also be considered appropriate behavior? Furthermore, Figure 1B compares models based on whether their results follow the expected directions, but I could not find statistical significance tests supporting these comparisons. I think the authors need to better justify their evaluation criteria and provide stronger evidence for their conclusions.
Questions for authors
  1. Which model is used as the LLM judge? Could the authors provide human annotation results and agreement statistics for all six transparency metrics, covering both final responses and reasoning traces?
  2. How many conflict documents receive each plausibility score (0–3)? For documents rated 3, have the authors checked whether they still contain the intended knowledge conflict?
  3. Could the authors provide statistical significance tests for the comparisons in Figures 1B and 5? Which of the six directional comparisons are statistically significant for each model?
  4. Why does the main analysis compare only casual and medium document formats, even though formal documents were also generated?
3 comments
AuthorsOpenReview ↗

Response (part1)

We thank the reviewer for recognizing that our proposed framework enables a “more comprehensive evaluation of model behavior”. We are grateful for their constructive feedback, which has substantially strengthened the revised manuscript, and which we address point by point below.

Critical Weaknesses:

  1. [No human validation, plausibility circularity]

We ran a human audit of the LLM judge, see our response below.

We evaluate the plausibility using the model(s) evaluated on purpose: our goal is to understand the impact of self-assessed plausibility on behavior. However, for completeness, we repeat the plausibility analysis using 3 held-out models (Grok 4.7, Llama 4 Maverick, and Mistral Large 4). We observe near-identical results when evaluating plausibility using an ensemble of the 5 study models, 3 held-out models, or all 8 models. See the updated Appendix B.2.1 for additional results.

  1. [Compliance definition, behavior of interest, Figure 1B statistical significance]

In this work we posit that a desirable behavior under knowledge conflict is to either refuse the task, or to complete the task and simultaneously flag the conflict. We only consider compliance problematic when it is not complemented by conflict detection. We consider both silent adoption and silent correction to be undesirable when it comes to epistemic transparency: even if the model silently corrects the wrong claim while completing the task, the user doesn’t become aware that the given evidence is problematic, does not get a chance to verify it, and might rely on it again in the future. We consider explicit correction to be desirable in addition to detection, hence measure it as one of the epistemic transparency traits.

To the reviewer’s second point on statistical significance, we have added the analysis and results in the response to their question below (Question #3).

Questions For Authors:

  1. [Missing details, missing human validation]

Opus 4.8 was used as the transparency scorer. We’ve updated the paper to make it clear.

We ran a small human audit of the LLM judge (9 auditors scoring 5-6 model responses each, see full results in Appendix B.5). On conflict detection and compliance, the two central dimensions for our main results, annotators agree strongly with each other (Krippendorff's α = 1.00 and 0.81) and the judge closely matches the human majority (Cohen's κ = 0.72 and 0.77); the same holds for correction (α = 0.82, κ = 0.88). Agreement is weaker on attribution explicitness (α = 0.03) and reader agency preservation (α = 0.17), and we have too few refusals to evaluate propose-alternative robustly. Our interpretation is that the weaker agreement reflects these behaviours being more subjective and harder to define. The goal of this work is to show that epistemic transparency can be measured under realistic conflicts, rather than to rank models precisely on every dimension. Defining attribution and reader agency more precisely, ideally with subject-matter experts, is left for future work.

  1. [Statistics of plausibility scores and examples]

We provide the breakdown of plausibility scores from the ensemble of models (rounded mean) for 86 stimuli (conflicting evidence in medium format only):

Plausibility ScorePercentage
010.5%
123.3%
223.3%
343.0%

We have reviewed a sample of documents scored 3, and can confirm they all contain the intended knowledge conflict. However, even in the presence of this conflict, the models still consider the overall document to be empirically plausible, even if they later raise the contradiction when asked to complete a task. As seen in Figure 13, GPT-5.4, for example, flags conflicts in ~50% of documents that it itself assigns a plausibility score of 3. This points to calibration as a potential explanation for a large number of documents with a plausibility score 3: a model knows a conflict is present, but is not considering this conflict to be significant enough to score the whole document as implausible.

AuthorsOpenReview ↗

Response (part2)

  1. [Statistical tests for headline results]

Following the reviewer’s request, we tested each directional prediction with a topic-cluster bootstrap (10,000 resamples; paired for stakes and format, independent for domain). A contrast counts as aligned or misaligned only if the 95% CI excludes zero in the expected or opposite direction. Across 30 model-by-contrast tests, 6 are significantly aligned, 24 are not significant, and none are misaligned. The aligned results are Sonnet 5 compliance on all three axes, GPT-5.4 compliance for stakes and domain, and Qwen detection for stakes (full stats table in the paper; Table 5).

ModelDomain complianceDomain detectionFormat complianceFormat detectionStakes complianceStakes detectionScore
GPT-5.4−0.077 [−0.168, −0.002]−0.071 [−0.289, +0.147]−0.058 [−0.145, +0.014]−0.014 [−0.130, +0.087]−0.087 [−0.174, −0.014]+0.072 [+0.000, +0.174]2/6
Gemini 3.5 Flash0.000 [0.000, 0.000]−0.043 [−0.287, +0.207]0.000 [0.000, 0.000]+0.072 [−0.043, +0.217]0.000 [0.000, 0.000]+0.058 [−0.116, +0.232]0/6
Claude Sonnet 5−0.282 [−0.512, −0.045]−0.179 [−0.500, +0.158]−0.304 [−0.507, −0.101]−0.262 [−0.548, +0.015]−0.442 [−0.609, −0.275]−0.061 [−0.327, +0.182]3/6
Qwen 3.6 27B+0.006 [−0.033, +0.042]+0.149 [−0.106, +0.397]0.000 [−0.043, +0.043]+0.116 [−0.029, +0.275]−0.014 [−0.043, +0.000]+0.174 [+0.058, +0.319]1/6
DeepSeek V4 Pro+0.017 [−0.029, +0.062]−0.060 [−0.282, +0.168]−0.014 [−0.043, +0.000]0.000 [−0.087, +0.087]−0.014 [−0.043, +0.000]−0.058 [−0.203, +0.116]0/6

We have updated the paper to reflect this more conservative reading: 20% tests are significantly aligned, and none are misaligned. This is still concerning, since a risk-sensitive model should shift toward caution as domain risk, evidence credibility, and stakes rise, and for most models and axes it does not, which is our original point. Our main contribution is the framework that makes these questions testable. It generates conflicts across stakes levels, formats, and domains, and we expect more data and a wider range of domains, scenarios, and formats to reveal clearer patterns.

  1. [Analysis limited to 2 document formats]

As noted in our response to Reviewer EzZ3, our goal is to convey that the framework supports conflict generation at multiple stakes levels, formats, and domains, and that our experiments use a simple instantiation of it. We compare casual vs. medium because this contrast is enough to test whether credibility shifts behavior, and adding the formal format would add complexity without changing our main point. We have clarified this in the paper.

Thanks for the clarifications, and I will consider raising my rating.

Reviewer_EzZ3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Prior work on knowledge conflicts mostly measures which source a model follows, not whether it discloses the conflict. The authors argue this, and add that multiple-choice and entity-substitution stimuli cannot vary the conditions under which disclosure matters. From a risk-minimization view of a user who unknowingly supplies a false document, they derive a six-part taxonomy of "epistemic transparency":

  • Always scored: Compliance and Conflict Detection.
  • Scored only on refusal: Propose Alternative.
  • Scored only on compliance: Correction, Attribution Explicitness and Reader Agency Preservation.

The stimuli come from a pipeline over MMLU‑Pro, covering psychology, economics and health:

  1. An LLM screens questions and picks the most plausible wrong option; 142 of 150 questions are judged usable.
  2. A question is kept only if all five target models answer it correctly ≥80% of the time; 86 remain.
  3. A generator that is not told which answer is correct writes up either the right answer (baseline) or the wrong one (conflict).
  4. A formatter renders each write-up three ways: a LinkedIn-style post, an encyclopedic article, or a formal paper.
  5. A separate LLM writes a one-line task request at each of two stakes levels, e.g. "expand my personal study notes" (low) vs. "update our ED trauma protocol" (high), with the document held fixed.

Section 4 is the multiple-choice study: medium format, no task request, compared against ConflictQA and FaithEval. It records the chosen answer and whether the justification and the reasoning trace flag the conflict, as judged by an LLM. Section 5 is a full factorial over domain × conflict/baseline × format × stakes × model. An LLM judge scores all six taxonomy dimensions on both the response and the reasoning trace. The five models also rate each conflict document's "empirical plausibility" on a 0-3 scale after the fact. The paper concludes four things:

  • Disclosure is rare, and silent adoption is the main failure.
  • Models differ a lot in strategy.
  • Transparency is concentrated on implausible fabrications.
  • No model fully shows the expected rise in caution.

It closes by proposing that a well-calibrated model should comply less and flag more as risk, credibility and stakes increase.

Critical strengths

The question is legitimately interesting and timely. As LLMs work over user-supplied documents in professional settings, whether a model surfaces a contradiction matters more than which side it quietly picks, and "silent vs. flagged adoption" is useful. Several design choices are better than usual in this paper:

  • Knowledge is verified per question across all evaluated models, instead of assuming the benchmark key is what the model knows. The point about ConflictQA's GPT‑4-era memory answers (A.3) is well taken.
  • Conflict and baseline documents use the same prompts, and the generator is not told which answer is correct. This gives a matched no conflict control (Fig. 12).
  • Stakes are isolated cleanly: the document stays fixed and only the one-line task request changes.
  • Response and reasoning trace are scored separately. This exposes the "noticed in reasoning, dropped from the answer" gap, which matters for anyone monitoring reasoning.

The stimuli are clearly more persuasive than ConflictQA and FaithEval, which is a contribution in its own right. All prompts are printed verbatim, and the limitations section is candid.

Critical weaknesses

The headline result is not supported. The abstract and Fig. 1B ("none fully aligned", 4/6 … 2/6) rest on a stipulated target and a scoring rule that does not hold up:

  1. The six comparisons are never defined. They appear to be 3 axes × {compliance↓, detection↑}, each scored by the sign of a point estimate with no uncertainty. The paper's own text contradicts the figure. §5.2 says only Sonnet and GPT‑5.4 comply less under high stakes ("other models show little to no difference") and that detection is "similar across stakes levels". Yet Fig. 1B credits Qwen and DeepSeek with the expected compliance drop on stakes, and GPT‑5.4, Gemini and Qwen with the expected detection rise.
  2. Detection is measured only among compliant responses. Figs. 5 and 9–12 report it conditional on Compliance = True, although Table 2 defines it as independent of compliance. Conditioning on an outcome of the manipulation produces exactly Sonnet's pattern of lower compliance and lower detection on all three axes. If Sonnet's refusals are the cases where it flagged the conflict, dropping them lowers detection among what remains. Its 3/6 may be an artifact.
  3. Raw compliance is the wrong target. GPT‑5.4 drafts the protocol while correcting the false claim, and it is penalized for not refusing more. The score therefore mixes refusal policy with transparency. The risk-relevant quantity is uncritical compliance: complies, does not flag, does not correct.
  4. The credibility axis runs backwards. It asks models to defer less to more credible-looking evidence, against ordinary evidential reasoning. Risk-sensitivity justifies flagging more, not complying less.
  5. Levels and domains are dropped without explanation. The formal format and the psychology domain are generated but left out of the directional tests (medium vs. casual; health vs. economics). The claim that health is riskier than economics is asserted rather than demonstrated.

Measurement validity, statistics and reproducibility are weak.

  • Judge. Every rate is an LLM-judge label, but the judge model is never named. There is no human validation or agreement statistic for six subtle binary labels, such as "central vs. peripheral claim" or "beyond mere hedging."
  • Other unnamed models. The screener, generator, formatter and stakes-writer models are not named, so a model may be evaluated on text produced by its own family.
  • Missing parameters. The number of samples per cell, the number of accuracy-check samples, temperatures, reasoning budgets, the per-domain split of the 86 questions, and the post-filter counts for the cross-dataset study are all unreported.
  • No inference. There are no significance tests or effect sizes anywhere. Fig. 5's error bars are standard errors across five models, which is not sampling uncertainty.

The factorial design calls for an item-level mixed-effects model.

  • Circular plausibility analysis (Fig. 6). The judges are the evaluated models themselves, and a score of 3 means "already true." A model that judges a false document true will, by construction, not flag it. That some conflict documents score 3 at all also undercuts the claim that models "provably hold" the knowledge. ≥\geq 80% accuracy at choosing the right letter on the bare question does not show that a model rejects the written-up false claim.
  • Baseline false alarms ignored. The baseline condition has a non-trivial false-detection rate (Fig. 12), and it is never subtracted from the conflict-condition rates.
  • Partly definitional finding. In the multiple-choice setting, "detection predicts resistance" is partly true by construction: justifying the correct answer against a passage almost requires saying the passage is wrong.
  • Missing related work. ClashEval (Wu et al., 2024) already links context adoption to how far the evidence departs from the model's prior, which anticipates Fig. 6. ConflictBank and the knowledge-conflict survey (Xu et al., EMNLP 2024) are also not cited.
Questions for authors
  1. Which model is the transparency judge, and which models do the screening, generation, formatting and stakes writing? What is judge/human agreement for each dimension on a stratified sample?
  2. Exactly how are the six Fig. 1B comparisons computed (which slice of the data, which levels)? Do they survive confidence intervals or a mixed-effects test? How does Fig. 1B change if detection is measured unconditionally, as Table 2 defines it, and if uncritical compliance replaces raw compliance?
  3. For the closed models, are the "reasoning traces" raw chain-of-thought or summaries written by the provider? Sonnet's $~$8% detection in reasoning vs. $~$45% in the (Fig. 4) could reflect summarization rather than reasoning.
  4. What are the per-cell sample sizes, temperatures and reasoning budgets?
7 comments
AuthorsOpenReview ↗

Response (part1)

We thank the reviewer for finding our work “legitimately interesting and timely”, and for the detailed and constructive comments. The revision they prompted has substantially strengthened the paper, and we hope it addresses their concerns in full. We respond to each point below.

Critical Weaknesses:

  1. [Six comparisons undefined, missing uncertainty measure, inconsistent result description]

We thank the reviewer for pointing this out.

We outline the six comparisons here (and have added this to the paper; Section B.4): To assess whether models adapt their epistemic caution to contextual risk, we evaluate directional alignment across three controlled pairwise contrasts under knowledge conflict, each varying a single factor while holding the others fixed at the high-stakes, medium-format health condition: stakes (high vs. low stakes in health), register (medium vs. casual format in high-stakes health), and domain (health vs. economics at high stakes and medium format). For each model and condition, we aggregate response-level judgements by first computing topic-level means and then averaging equally across topics on two outcomes: overall compliance, and conditional conflict detection among compliant responses. A model is scored as directionally aligned on a given contrast if moving to the higher-caution pole strictly decreases overall compliance or strictly increases conditional conflict detection (yielding six binary alignment indicators per model across the three contrasts and two outcomes).

To quantify uncertainty, we have now computed 95% percentile confidence intervals and two-sided p-values via a topic-cluster bootstrap, resampling topics in pairs for within-domain contrasts (stakes and register) and independently across domains for the health-versus-economics contrast. We discuss additional results in our response to your question on statistical tests for Fig 1B comparisons below (Question #2).

Indeed, there is inconsistency in our descriptions of results, which we have addressed and highlighted the changes in the paper (lines 117-119).

  1. [Inconsistency in compliance definition, Sonnet's special pattern]

Table 2 describes how the judge scores each dimension: Detection is scored independently of Compliance on every response for implementation convenience. However, our figures report detection among compliant responses only because this is the outcome of interest in the current context. We now state this explicitly in Table 2 caption. The harm-relevant failure is silent compliance: the model does the task and does not flag the conflict. We consider a refusal that flags the conflict to be a safer outcome, and don’t measure it as a failure mode in this work.

We computed detection given refusal (see the table below).

ModelDomain HealthDomain EconFormat Cas.Format Med.Stakes HighStakes LowTarget Resp.Target Reas.
GPT-5.41.0 (6/6)1.0 (1/1)1.0 (2/2)1.0 (6/6)1.0 (6/6)-1.0 (6/6)1.0 (4/4)
Gemini 3.5 Flash-------1.0 (14/14)
Claude Sonnet 50.9 (35/41)1.0 (30/30)0.8 (17/20)0.9 (35/41)0.9 (35/41)0.8 (9/10)0.9 (27/29)0.9 (19/22)
Qwen 3.6 27B1.0 (1/1)1.0 (2/2)1.0 (1/1)1.0 (1/1)1.0 (1/1)-1.0 (1/1)1.0 (1/1)
DeepSeek V4 Pro0.0 (0/1)0.7 (2/3)-0.0 (0/1)0.0 (0/1)-0.0 (0/1)-
Mean ± SEM0.7±0.20.9±0.10.9±0.10.7±0.20.7±0.20.8±0.00.7±0.21.0±0.0
Models w/ comp=0(4/5)(4/5)(3/5)(4/5)(4/5)(1/5)(4/5)(4/5)

Only Sonnet 5 has enough refusals to analyze. Sonnet flags the conflict in most refusals (e.g. 35/41 and 30/30 in the health and economics high-stakes slices). The more informative result is the following gap: Sonnet’s detection is significantly higher when it refuses than when it complies in the health and economics high-stakes slices (see the table below). So Sonnet’s lower detection among compliant responses reflects an observable behavioral difference: when it complies, it flags less often. This reinforces our focus on the compliant-but-unflagged case. We have updated the paper to reflect this point (lines 367-369).

ModelSliceP(det|c=0)P(det|c=1)Delta95% CIp (2-sided)
Claude Sonnet 5Health / High / Med. / Resp.0.856 (35/41)0.455 (14/28)+0.401[+0.093, +0.697]0.0112*
Claude Sonnet 5Econ / High / Med. / Resp.1.000 (30/30)0.633 (40/66)+0.367[+0.200, +0.545]0.0000*
Claude Sonnet 5Health / High / Cas. / Resp.0.800 (17/20)0.717 (35/49)+0.083[-0.135, +0.286]0.4170
Claude Sonnet 5Health / Low / Med. / Resp.0.833 (9/10)0.515 (27/58)+0.318[-0.014, +0.606]0.0585
Claude Sonnet 5Health / High / Med. / Reas.0.879 (19/22)0.735 (24/35)+0.143[-0.092, +0.403]0.2604
AuthorsOpenReview ↗

Response (part2)

  1. [Raw compliance, uncritical compliance]

We agree with the reviewer, and hence don’t use raw compliance as the target: we posit in this work that a desirable behavior under knowledge conflict is to either refuse the task, or to complete the task and simultaneously flag the conflict. We only consider compliance problematic when it is not complemented by conflict detection / explicit correction, as we consider both silent adoption and silent correction to be undesirable when it comes to epistemic transparency: even if the model silently corrects the wrong claim while completing the task, the user doesn’t become aware that the given evidence is problematic, and might rely on it again in the future.

  1. [Credibility axis runs backwards]

When a document is more credible-looking, by our definition, models ought to be more cautious, which means we expect them to comply less and flag more. This is because the more credible the presentation of misleading information is, the higher we hypothesise the risk is that a user may rely on this information, hence the higher the need for appropriate epistemic transparency by a model. Could the reviewer clarify their point on the credibility axis running backwards?

  1. [Levels and domains dropped]

We have revised the paper to state that the framework supports multiple stakes levels, formality registers, and domains, while our experiments instantiate a deliberately simple version of it (lines 332-340). This is in order to streamline the methodology and analysis in the present work, illustrating our argument for epistemic transparency and bearing in mind that future work should build our formats, domains, etc. For format, we compare casual vs. medium, since this contrast is enough to test whether credibility shifts behavior. For domain, we compare health vs. economics, assuming health is higher-risk because errors cause more direct, immediate harm to individuals. Results for the Psychology domain are reported in the appendix (Fig. 10) and are excluded from the directional comparison, because its risk ordering relative to health and economics is less clear.

[Missing judge, human validation]

We already name the judge in Appendix 2.4: “we use Opus 4.8 as the scoring judge”. We now mention it in the main text as well to avoid confusion. We provide additional results regarding human agreement in our response to Question #1 below.

[Unnamed models, circularity]

Claude Opus 4.8 was used for screening and stake-specific prompt writing. GPT-5.4 was used for conflict knowledge generation and formatting because Claude Opus 4.8 frequently declined these generation requests.

We recognize that GPT-5.4 was used both to construct the stimuli and as an evaluated model; using a generator outside the evaluated model set would have provided cleaner separation. However, this overlap does not directly compromise our primary alignment analyses, which estimate within-model directional contrasts across conditions rather than rank models by absolute response rates. Because both poles of each contrast used the same stimulus-generation pipeline, any stable affinity of GPT-5.4 for same-model-generated content should affect both conditions similarly and shift absolute response levels rather than the direction of the contrast.

Moreover, GPT-5.4 did not exhibit the pattern expected under preferential uncritical uptake: it detected the conflict in 75.2% of its compliant responses, and its uncritical-compliance rate was among the lowest of the evaluated models. Nevertheless, because stimulus-generator identity was not experimentally manipulated, we cannot exclude a condition-specific interaction or make a causal claim about self-preference.

We have updated the paper with the above clarification (lines 923-935).

[Missing parameters]

Assuming by “per cell” the reviewer is referring to one unique combination of question x evidence condition x format x stakes condition x evaluated model, we collect 3 samples per cell. For accuracy check, the screening results in 142 total topics (i.e., questions) and per topic accuracy is computed by averaging across 5 samples per evaluated model. For the accuracy-check study, we set temperature=1 for all models and use the default reasoning effort. For the main knowledge conflict study, we set temperature=1 and reasoning to high (or equivalent) for evaluated models. Per-domain split of the 86 questions: 32 economics, 29 psychology, and 23 health.

For the cross-dataset study, accuracy filtering leaves us with the following number of questions per dataset: 25 ConflictQA, 24 FaithEval, 32 MMLU (economics), 27 MMLU (psychology), 22 MMLU (health). Note that 1-2 additional questions were dropped for two of the MMLU datasets compared to the main study due to different model reasoning settings used in the two studies.

AuthorsOpenReview ↗

Response (part3)

[No statistical tests]

We have provided additional statistical analysis results in our responses to your question above on statistical tests for the Fig 1B comparisons (Weakness #1 and #2).

[Circular plausibility analysis]

We evaluate the plausibility using the model(s) evaluated on purpose: our goal is to understand the impact of self-assessed plausibility on behavior. However, for completeness, we repeat the plausibility analysis using 3 held-out models (Grok 4.7, Llama 4 Maverick, and Mistral Large 4). We observe near-identical results when evaluating plausibility using an ensemble of the 5 study models, 3 held-out models, or all 8 models. See the updated Appendix B.2.1 for additional results.

[Document with score of 3 will not be flagged]

We disagree with the reviewer’s claim that “a model judges a false document true will not flag it by construction.” On the contrary, Figure 13 shows that GPT 5.4, for example, flags conflicts in approximately 50% of documents that the model itself scored as 3. We should also clarify that MCQ-style evaluation is the simplest approach to verifying model knowledge free of confounds such as format.

The existence of conflict documents scored 3 and yet flagged as conflicts does not undercut the reliability check, because conflict documents are confounded by formality whereas MCQ evaluation is not. If anything, this raises an interesting question: why does a model judge a document as true when asked to score it, yet flag a conflict when asked to perform a task with that same document? We interpret this as a calibration issue rather than a capability one, and suggest it as a direction for future research.

[Baseline ignored]

The baseline false-alarm rates are reported in Fig. 12 and are far below the conflict-condition rates. We acknowledge in the Discussion that some false alarms likely come from imprecise content introduced by the generation pipeline, which could be improved further. We report the two conditions side by side rather than subtracting, because subtraction would obscure the per-condition behavior and assume false alarms are additive.

[Partly definitional finding]

We do not understand this point, please could the reviewer clarify?

[Missing related work]

Thank you for these references, which we have now cited including the survey. Both ClashEval and ConflictBank fall into the two categories of prior work we already discuss, and our contribution moves beyond them.

ClashEval (Wu et al., 2024) measures which source a model prefers (prior vs. context) under conflicts generated by entity substitution. ConflictBank (Su et al., 2024) likewise measures source preference (memory, context, or competing contexts), with conflicts built by substituting same-type entities in Wikidata triples and having an LLM write them up as Wikipedia pages, book passages, and news articles. Neither measures whether and how the model discloses the conflict to the user or other traits of epistemic transparency, and neither varies evidence format or real-world stakes while holding the claim fixed.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

We thank the reviewers for a productive discussion. All three engaged with our rebuttal, and the paper is stronger for it. Reviewers recognized the work as timely, provides a more comprehensive evaluation and as a key shift from prior knowledge-conflict research. Existing work asks which source a model follows. We think it is equally important to ask whether, and how, models should disclose a conflict to the user, and how the ideal form of disclosure changes with circumstances such as the stakes of the task and the credibility of the evidence. Our contribution is a framework for making these questions measurable: a taxonomy for scoring epistemic transparency, and a pipeline that varies evidence credibility, real-world stakes and domain under verified model knowledge.

The two main concerns raised were the lack of statistical analysis and the lack of human validation of the taxonomy. We addressed both.

Statistical analysis: We added bootstrap significance tests for all directional comparisons and made silent compliance the primary metric, as requested. The headline claim is better supported: there is no strong evidence that models become more cautious as credibility, stakes or domain risk increase.

Human validation: Human annotators and a second judge from a different model family agree strongly with our judge on the dimensions our headline results rest on, namely compliance and conflict detection. Agreement is weaker on the more exploratory, subjective dimensions.

We acknowledge the remaining concern by one of the reviewers regarding the lower agreement on the auxiliary taxonomy dimensions. However, we respectfully disagree that this undermines the contribution. Our headline results do not depend on those dimensions, and the disagreement shows they need sharper definitions and expert validation in future work. That reinforces our central point: epistemic transparency in LLM responses is hard to measure, still unsolved, and is important to get right.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.