- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
The paper investigates whether LLMs reduce their reliance on a document's target claim when the document contains intrinsic credibility flaws. It introduces CREDO, with 1,292 retrieval-dependent yes/no questions and 7,752 document variants spanning opposing claim directions and one to three flaws. A controlled search tool returns fixed, inconclusive background material. Across 22 models, the study measures claim adoption (ADOPT), flaw identification (IDENTIFY), search (SEARCH), and overclaiming (OVERCLAIM). It reports that most models readily adopt the supplied claim, that supplying the same document as a tool response increases adoption, and that evidence-reflection SFT improves these metrics.
Critical strengths
-
Separating whether a model adopts a claim, identifies a flaw, or expresses uncertainty is a useful idea.
-
The comparison between supplying the same document in a user message and in a tool response, and the observation that Llama models search by restating the question, are of some interest.
Critical weaknesses
- The core construct is self-contradictory, and the paper's claims do not hold under either reading.
The paper gives two incompatible descriptions of where the flaws are placed:
- Main text (Abstract; Section 1, p. 2; Section 2.2, p. 3; Section 5, p. 9): documents "retain the target evidence", flaws are located "outside the evidence relevant to the target claim", and none "directly contradict" that claim.
- Appendix A.7.3 (pp. 23–24): the synthesis prompt requires a document "whose critical support chain contains one subtle gap"; the generator must "choose one conclusion-critical bridge", "put the assigned and only gap in that bridge", and copy the cue span "from the weak premise before the final claim". The review prompt (p. 24) further requires that, once the cue-bearing material is removed, the remaining body must not independently entail the conclusion.
The appendix's own examples match the latter description. In Figure 9 (p. 21), the inconsistent Slovenian yields and the conflicting fulfillment totals sit directly on the data needed to answer the question. In Figure 1, the conflicting trial-operation dates also bear directly on the opening date.
Reading A (main text): the flaws do not affect the target claim. In this case ADOPT↓ lacks a normative basis. For a yes/no question with no other evidence, a source that is not fully reliable, but not deliberately misleading, still raises the probability that its claim is true, so answering in the document's direction remains the best single guess. The paper itself states that the flaws "warrant caution in relying on its target evidence, even if the target claim is true" (p. 1); caution calls for expressing reservations, not for changing the answer. Because ADOPT also counts abstention as non-adoption, ADOPT↓ effectively rewards abstention. Gemma's adoption dropping to 0.2–0.7% after SFT (Table 14) illustrates this problem. In addition, the paper lacks an ablation with zero-flaw or repaired-flaw documents that retain the same target evidence, so it cannot distinguish a model that is sensitive to flaws from one that distrusts everything. Under this reading, neither the value of the dataset nor the meaning of the main metric is established.
Reading B (appendix): the flaws lie in the support chain of the conclusion. In this case the document genuinely fails to support its conclusion, non-adoption is the correct behavior, and ADOPT becomes meaningful. However:
- The task then reduces to detecting whether the context suffices to support the answer, i.e., the sufficiency dimension that the paper explicitly claims to go beyond (Sections 1 and 5). It overlaps substantially with RGB, FaithEval (in Related Work Section 5), and Sufficient Context (Joren et al., ICLR 2025), and the claimed novelty no longer holds.
- The title, abstract, introduction, Figures 1 and 2, and the positioning in the related-work section do not describe the actual data.
Therefore, under Reading A the dataset has little value and the main metric is normatively unfounded; under Reading B the novelty claim fails and the paper misdescribes its own data. This cannot be explained as an outdated template or unclear documentation. Whichever reading the authors confirm, the paper requires a substantial rewrite.
- The metric definitions cannot be reconstructed and do not align with the validated fields.
- In Section 3.2, IDENTIFY only requires a specific description of a flaw, whereas the judge additionally requires explaining how the flaw weakens support for the conclusion (p. 25). This merges recognition with assessing its implications, which the paper sets out to distinguish.
- The formal OVERCLAIM metric requires an uncertainty variable U, and per-type recall requires a matched flaw set F, but neither field appears in the judge's output schema. Appendix A.7.4 explicitly states that the avoided_overclaim field differs from the main-text OVERCLAIM metric, yet gives no mapping between them.
- Human validation covers the raw judge fields, not the final metrics. Expert-consensus agreement is lowest for avoided_overclaim (85.23%). The valid_evidential_inference field is validated but not used by any metric.
- The construction and quality-control specifications directly conflict with the main text in several places.
- Number of flaws. The review prompt states that "each document has one primary flaw" and includes a MULTIPLE_DEFECTS rejection code (pp. 24–25), whereas the main text states that each document receives three flaws of distinct types (Section 2.2). Under the appendix's review criteria, the three-flaw documents described in the main text should all have been rejected.
- Undisclosed data language. The synthesis prompt requires documents of "650–1,050 Chinese characters" (p. 23), and all appendix prompts are presented as English translations, indicating that the data are in Chinese. The main text, Figure 1, and all examples are presented in English, and this is never stated. Were the target models evaluated in Chinese or in English?
- Inconsistent query-selection criteria. The main text states that long-tail or time-sensitive queries are retained, but acceptance requires scores of at least 4 on long-tail-ness, temporal variability, and specificity alike (Section 2.2; Appendix A.7.2).
- Source of the target evidence. The synthesis prompt allows synthetic records to carry the desired conclusion, attributed to neutral repositories or project logs (p. 23); that is, the support for the target claim is itself fabricated. Figure 2 and Section 2.2 emphasize that background material comes from real retrieval with no added factual content, which may give readers an impression of realism, while the synthetic nature of the target evidence is not stated.
Note on the scope of the response: Weakness 1 affects the paper's core conclusions, and I ask the authors to focus on it. For the items in Weaknesses 2 and 3, as well as other inconsistencies in the paper, the authors may respond selectively; a point-by-point reply is not required. These points are raised to help the authors clarify the actual experimental setup and strengthen the soundness of the paper.
- Lack of reproducibility.
Given the issues above, the only statement on availability is the sentence "Code and data will be released" at the end of the abstract; no anonymous code repository or supplementary material is provided. All appendix examples are marked as having altered "names, dates, values, and wording" (pp. 19–20), so the paper contains not a single unaltered benchmark instance, and the prompts are provided only as English translations. Under these conditions, none of the points above can be independently verified, which further weakens the soundness of the paper.
Questions for authors
-
Does the actual data correspond to Reading A or Reading B? If A, please explain why flaws that do not affect the target claim should change the model's answer, i.e., the normative basis for ADOPT↓. If B, please explain how this work differs from evidence-sufficiency detection, and revise the title, abstract, and introduction accordingly.
-
How are "each document has one primary flaw" and the MULTIPLE_DEFECTS rejection code in the review prompt reconciled with three flaws per document? In what language are the data constructed and the models evaluated?
-
How are A, U, and F obtained from the actual judge outputs? Please provide the exact mapping and human agreement on the final metrics.
-
Please report supported-answer, abstention, and false-flaw-identification rates for the base and SFT models on zero-flaw/repaired-flaw documents and on the conclusive-evidence control.
-
The ICLR 2027 Author Guidelines allow the paper's content and supplementary materials to be revised during the discussion period. Please provide anonymized code, the complete data (or at least a representative subset), and unaltered original examples by updating the supplementary materials.
3 comments
We hope you are doing well. We sincerely appreciate the thoughtful feedback and constructive suggestions you have provided.
With the discussion deadline approaching, we would be grateful if you could review our responses when you have a chance and let us know whether any concerns remain. Your feedback would allow us sufficient time to address additional questions and provide any further information that would support the discussion.
If our rebuttal and manuscript revisions have resolved your concerns, we would sincerely appreciate your consideration of an increase in your rating.
Thank you again for your time and consideration. We look forward to continuing the discussion with you.
Thank you for the response. However, it seems the author has not adequately addressed my concerns regarding the internal contradictions in the core argument (W1, Q1). Consequently, I intend to maintain my current score, and I would also like to draw the Area Chair’s attention to the fact that the author has failed to resolve these issues.
Thank you for the detailed review. We address the central distinction and implementation questions below.
Q1. Credibility concerns do not logically refute the target claim
A flaw not logically refuting a claim does not mean it has no bearing on trust in the document. CREDO requires that flaws must not establish the opposite answer. Appendix A.7.3 explicitly states: "The gap may weaken the current claim but must not prove the opposite answer." Detecting a flaw therefore does not establish that the target claim is false or necessarily require reversal or abstention.
When a model cannot answer from parametric knowledge and receives an answer-bearing document of unknown provenance, its options extend beyond complete refusal and unconditional acceptance. It can attempt tool-based cross-verification. Even if the returned information remains inconclusive, it can identify contradictions or provenance concerns as leads for further investigation and communicate reservations about evidence quality.
Thus, the proposed implications for answer behavior under Readings A and B do not necessarily follow: a flaw that does not negate the claim need not be ignored, while weakened support does not make abstention the only appropriate response. We assess whether models identify credibility concerns and reflect them in their answers, not whether they automatically change answer direction.
Section 3.2 explicitly allows qualified adoption: adopting the claim with evidence-related uncertainty counts toward ADOPT but not OVERCLAIM. We will clarify flaw placement and evidential effects and avoid presenting lower ADOPT alone as improvement. We recognize that the non-refutation requirement alone does not resolve the separate question of where flaws lie relative to the support chain; the construction description must make that distinction explicit.
Q2. Flaw counts and evaluation language
Section 2.2 describes constructing three distinct flaws and removing complete flaw paragraphs to produce nested variants. The 1,292 questions have two claim directions and three flaw-count conditions, yielding 7,752 directional instances (Table 5). Table 4 distinguishes five top-level categories from finer construction labels.We have publicly released the relevant data for inspection.
Q3. Metric extraction and human agreement
Section 3.2 defines A as answer direction, F as identified annotated flaws, U as explicit evidence-related uncertainty, and S as observed search activity. The appendix stance labels can map to yes/no/abstain, and matched cue IDs in detection_evidence provide information for F.
We will release the evaluation and aggregation code to specify the exact sources and mapping of A, F, and U. In particular, detected_injected_cue requires identification and explanation of a flaw's implications; avoided_overclaim is an auxiliary confidence judgment whose complement cannot simply replace OVERCLAIM; valid_evidential_inference diagnoses understanding of evidential implications.
Table 7 reports expert-consensus agreement: Answer 100.00%, Flaw Identification 95.18%, Evidential inference 97.83%, Avoided-overclaim 85.23%, and overall 94.46%. These validate raw judge fields, not all final aggregated metrics.
Q4. Controls and post-SFT response behavior
Appendix A.2 and Figure 7 assess standard QA before and after SFT under a common retrieval setup; substring EM performance is preserved. We additionally sampled 100 questions from each of seven QA datasets, totaling 700, and manually reviewed responses from both models after evidence-reflection SFT:
These are aggregate sample statistics, not full-test-set or per-dataset estimates. Uncertainty while answering is distinct from complete refusal. This evidence concerns external QA; it does not supply the requested matched-control supported-answer and false-flaw-identification rates.
For zero-flaw documents, it is difficult to guarantee that collected examples are entirely correct and free of credibility flaws, especially for highly time-sensitive questions. Consider, for example, a hypothetical question about whether an unverified allegation concerning a public figure is true. A model may be unable to give a definite yes/no answer without refusing to engage with the question, it can discuss the available information and its limitations. Extensive online discussion does not itself establish the allegation’s truth. As LLMs are increasingly used as search engines, we believe that evaluating model behavior in such situations is valuable.
Q5. Anonymous resources
The questions, documents, and flaw annotations are available at: https://huggingface.co/datasets/anonymous-submission-01/CREDO/tree/main