{"id":"60e57898-2f8a-4334-8c41-906deb187416","arxiv_id":"2601.07506","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"LLM judges frequently mark candidates wrong when the gold reference contradicts the model's own knowledge, even if the candidate exactly matches the provided reference.","lead":"This paper finds that LLM judges used to grade QA answers often ignore the given reference answer when it clashes with what the model already knows, and then mark even matching answers wrong. It introduces a swapped-reference test and shows the problem persists across thirteen models and four datasets, and that simple prompting fixes do not solve it.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ground-truth labels rely on unvalidated GPT-4o candidate generation; without artifact-rate analysis, the reported ACC_s drops may not measure reference adherence.","rationale":"The strongest claim requires two things: (1) the ground-truth labels are correct (i.e., c_s actually asserts r_s), and (2) the normative premise that reference-adherence is the correct standard. The paper partially defends (2) with the Direct prompt, which explicitly forbids outside knowledge but still leaves a large RPAG—so the phenomenon is not purely a value choice. The unguarded link is (1): candidates are generated by GPT-4o, a model with the same kind of parametric knowledge that allegedly causes judges to override references, and no quantitative validation of the generated candidates is provided. The claimed human review is asserted but not documented (no counts, no agreement). If a nontrivial fraction of (r_s,c_s) candidates do not actually contain r_s, then the low ACC_s values in Tables 1/7 would be at least partly expected, and the 'judges reject a matching candidate' claim collapses. The same concern applies to the NOT_ATTEMPTED mapping, which is not specified in §3.3. I therefore regard the dataset-validity check as the most load-bearing open question. The paper is otherwise well-designed: 13 models, 4 datasets, controlled swap types, and converging correlational evidence (popularity/freshness) support the knowledge-conflict mechanism. If the proposed candidate-fidelity check passes, the conditional acceptance is warranted; if not, the central claim would need qualification.","tokens_in":26264,"tokens_out":9393,"duration_ms":101805,"concrete_test":"Obtain from the authors (or the released artifacts) a random sample of at least 200 instances per dataset×swap-type, and independently annotate each (r_s,c_s) triplet for whether the candidate semantically entails the swapped reference entity and does not contradict it. Compute inter-annotator agreement. If >5% of candidates are misaligned, the accuracy drops are confounded. Separately, recompute ACC_s after partitioning judged outputs into CORRECT/INCORRECT/NOT_ATTEMPTED; if NOT_ATTEMPTED accounts for a large share of errors in (r_s,c_s), revise the 'reject' interpretation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—that judges 'reject a candidate answer even when it matches the provided reference' (§4.1)—depends on the ground-truth labels assigned in §3.1. For the (r_s,c_s) condition, c_s is generated by GPT-4o with a prompt instructing it to 'treat the substituted answer as ground truth' (Figure 8). If GPT-4o sometimes fails to do so—producing a candidate that omits r_s or reverts to the original answer—then a judge's Incorrect verdict is actually correct, and the large ACC_s drops in Tables 1 and 7 are artifacts. The authors state that human annotators reviewed all generations but report no artifact rate, no inter-annotator agreement, and no release of the dataset (TBA). A second measurement issue: §3.3's accuracy metric never specifies how NOT_ATTEMPTED verdicts are coded; if they are simply counted as errors, part of the accuracy drop may reflect judges declining to grade rather than actively rejecting reference-matching candidates, weakening the stated mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates whether LLM judges in reference-based QA evaluation actually condition on the provided gold reference, focusing on occasions when that reference conflicts with the judge's parametric knowledge. The authors construct a 'swapped-reference' framework: for each question, the original reference is replaced by an incorrect entity (type-preserving, type-changing, popularity-based, or evaluator-knowledge swaps), and long-form candidate sentences are generated to align with each reference. Judges grade four triplets (original/swapped reference × aligned/misaligned candidate), and the authors measure accuracy under original vs. swapped references, defining the Reference-Polarity Accuracy Gap (RPAG). Across thirteen LLMs and four QA datasets, they report large accuracy drops under swapped references, tracing the effect to judges' over-reliance on parametric knowledge: the drop vanishes when the swapped reference aligns with the judge's own prediction, increases with entity popularity, and decreases for fast-changing facts. Prompt-based mitigations (Direct instructions, CoT, self-consistency) do not eliminate the vulnerability. The paper concludes that LLM judges do not reliably follow the provided reference under knowledge conflict, motivating stronger reference-adherence protocols.","tokens_in":26498,"tokens_out":4543,"duration_ms":53220,"significance":"If the empirical claims hold, the paper identifies an important and under-studied failure mode in LLM-as-a-judge QA evaluation. The controlled swapped-reference design, the breadth of models (13) and datasets (4), and the converging mechanism analyses (§5.1–5.3) are noteworthy strengths. The findings have practical implications for evaluation pipelines where the gold reference is authoritative and for counterfactual or time-sensitive QA settings. The paper also provides a useful diagnostic methodology. However, the central quantitative claims rest on the quality of the automatically generated candidate sentences and on the treatment of NOT_ATTEMPTED verdicts, and all reported numbers are point estimates without uncertainty quantification. Reproducibility is currently limited because the dataset/code are 'TBA'. The paper is likely correct in essence, but the missing validation details must be supplied before the magnitude and interpretation of the drops can be fully trusted.","major_comments":[{"comment":"The ground-truth labels for the (r_s, c_s) condition depend entirely on GPT-4o obeying the 'treat the substituted answer as ground truth' instruction when generating c_s. If GPT-4o sometimes produces a candidate that omits the swapped entity or reverts to the original answer, an 'Incorrect' verdict from the judge would be objectively correct, and the ACC_s drops in Tables 1 and 7 would be overestimated. The authors state that human annotators reviewed all generations, but they report no artifact rate, no inter-annotator agreement, and the dataset is not yet released. Because the headline mechanism—'judges reject a candidate even when it matches the provided reference' (§4.1)—rests specifically on the (r_s, c_s) row of Table 7, this validation gap is load-bearing. Please quantify the artifact rate (e.g., percentage of c_s candidates that mention the intended swapped entity and not the ori","section":"§3.1, Figure 8"},{"comment":"The accuracy metric defines ground truth as Correct whenever a=b and Incorrect whenever a≠b, but it never specifies how a judge's NOT_ATTEMPTED verdict is coded. Since the evaluation prompt explicitly permits NOT_ATTEMPTED, counting it as an error would make judges that abstain under conflict appear to 'reject' reference-matching candidates—a different behavior from actively grading them Incorrect. The paper's central description in §4.1 ('reject a candidate answer') and the mechanism analysis in §5 depend on this distinction. Please report the distribution of NOT_ATTEMPTED verdicts by condition, and either exclude them from the accuracy computation or analyze them separately.","section":"§3.3 (evaluation metric)"},{"comment":"All reported accuracy values and RPAG scores are point estimates without confidence intervals or significance tests. Some drops are small in absolute terms (e.g., Table 1: GPT-5 on NQ-Open TP, -2.8 pp; on PopQA TP, -0.4 pp), and with thirteen models and four datasets some differences could be noise. This matters for three claims: the existence of a consistent 'substantial accuracy drop' across models, the near-zero RPAG under Evaluator-Knowledge swaps (§5.1), and the claim that CoT makes things worse (§6). Please provide bootstrap confidence intervals or pairwise significance tests for the main RPAG values, and at least confidence intervals for the 'nearly zero' claim.","section":"Tables 1, 6, and 7; Figure 3"}],"minor_comments":[{"comment":"The paper labels reference-following as the 'ideal' judge behavior and deviations as 'failures.' This is a defensible normative premise for reference-conditioned evaluation, but it is a value choice; a judge that refuses to endorse an implausible swapped reference could be interpreted as cautious rather than faulty. The Limitations section partially addresses this, but the abstract and conclusion state 'failure mode' unconditionally. Consider framing the contribution as measuring reference-conditioning fidelity rather than 'failure' without qualification.","section":"§2.2 and Limitations"},{"comment":"Minor presentation issues: 'Sciq' is inconsistently capitalized; the caption of Table 7 refers to 'TP; Corpus' and 'TC; Type_swap,' which is confusing; the dataset/code link is 'TBA'—this should be resolved before publication; and in §5.2 there is a typo 'levell.'","section":"Appendix A and Table 3"},{"comment":"The manual inspection of CoT reasoning paths covers only 50 examples per swap type from GPT-4o. This is acknowledged as qualitative evidence, but since it is used to support a strong claim about CoT amplifying parametric reliance, a more systematic coding (e.g., with a second annotator) would strengthen the analysis.","section":"§6 and Appendix C.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is timely and likely correct in its central phenomenon, but the missing artifact-rate analysis and the unstated NOT_ATTEMPTED coding are exactly the kind of details that could shift the measured effect sizes. I would encourage the editor to require the authors to provide the artifact analysis and error bars before publication, but I do not see a fundamental flaw that would warrant rejection. The work fits the journal/venue well and makes a useful methodological contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuine new finding — LLM judges will override the provided gold reference in favor of their parametric knowledge when the two conflict, and the effect is large and consistent across 13 models and 4 datasets. The swapped-reference framework and the RPAG metric are the real contributions; the evaluator-knowledge control, popularity, and freshness analyses give converging evidence for the mechanism. The paper deserves peer review.\n\nWhat it does well: the design is clean. Swapping the reference while keeping the candidate aligned or unaligned gives four triplets with well-defined ground truth. The fact that (r_s, c_s) accuracy collapses while (r_o, c_o) stays high is a striking asymmetry. The control where the swap matches the judge's own QA prediction (Evaluator-Knowledge Swap) drives RPAG to zero, which is the strongest evidence that the failure is belief-driven rather than format-driven. Showing CoT makes it worse under conflict is a useful, non-obvious result. The related work is appropriately positioned against knowledge-conflict in QA models.\n\nSoft spots: the measurements are all point estimates, with no confidence intervals or significance tests. For a paper that makes a fairly universal claim, I would want to see error bars. More important: the ground-truth labels depend on GPT-4o-generated candidates faithfully realizing the swapped reference. The authors say human annotators reviewed everything but give no artifact rate or inter-annotator agreement, and the data is TBA. If some fraction of c_s candidates do not actually mention r_s, the ACC_s drops are partly an artifact of bad candidates rather than judge behavior. The paper's own reasoning trace examples (Table 5) show judges explicitly saying \"the reference entity is incorrect\" and ruling Incorrect, so the core phenomenon is real, but the size of the drop may be overstated. Also, the accuracy metric counts NOT_ATTEMPTED as an error without reporting how often it occurs; refusing to grade a bizarre (r_s, c_s) pair is a different failure from active rejection.\n\nThe normative assumption — that a judge should grade against a factually wrong reference — is a protocol choice, not an empirical fact. It is stated clearly in §2.2 and defended in the Limitations, so I do not treat it as a flaw, but reviewers should be aware the practical relevance depends on deployments where the reference is authoritative (e.g., counterfactual or time-shifted benchmarks).\n\nWho this is for: anyone building or using LLM-as-judge systems for QA, and benchmark designers. It is a solid meta-evaluation, not a method paper.\n\nRecommendation: send to peer review. It needs a revision that releases artifacts, reports candidate artifact rates, and adds at least basic uncertainty quantification. But it is a real result, and the field should engage with it.","headline":"This paper shows a real and large failure mode in LLM-as-a-judge QA evaluation — judges override the provided reference under knowledge conflict — but the effect size needs more measurement transparency before I'd trust it fully.","tokens_in":26972,"tokens_out":3288,"would_cite":true,"duration_ms":36447,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"LLM judges flunk when the gold answer contradicts what they know","keywords":["LLM-as-a-judge","QA evaluation","reference adherence","parametric knowledge conflict","swapped-reference evaluation","evaluation fidelity","knowledge conflict","prompt mitigation"],"falsifier":"Run the same swapped-reference protocol with human raters independently checking every generated candidate for fluency and semantic faithfulness to its intended reference. If a substantial share of swapped candidates are awkward or accidentally echo the original answer, the accuracy drop could be attributed to candidate quality instead of knowledge conflict. Alternatively, if any judge model shows identical verdicts for original and swapped references on a large sample, its RPAG would be near zero and would directly contradict the paper's central claim.","tokens_in":26178,"feed_emoji":"⚖️","tokens_out":2815,"duration_ms":28683,"temperature":0.7,"pith_summary":"This paper tries to establish that LLM judges, when grading question-answering outputs against a provided gold reference, do not reliably follow that reference if it conflicts with what the model already knows. The evidence is a controlled swapped-reference framework: when the reference is replaced with a plausible but incorrect entity, judges frequently mark a candidate as Incorrect even when the candidate and reference give the same (swapped) answer. This failure appears across thirteen judge models and four QA datasets, and it does not disappear with larger models or with common prompting strategies such as chain-of-thought and self-consistency. The authors argue that judges over-rely on parametric knowledge and that this is a fundamental limitation of LLM-as-a-judge evaluation.","feed_headline":"LLM judges flunk when the gold answer contradicts what they know","feed_subtitle":"Swapped-reference tests show graders reject candidates that match the provided answer, breaking reference-based QA evaluation.","key_machinery":"The central mechanism is the swapped-reference QA meta-evaluation framework. Each instance is a quintuple (question, original reference, swapped reference, candidate aligned with the original, candidate aligned with the swapped reference), producing four grading triplets whose ground-truth labels are known by construction. The key metric is the Reference-Polarity Accuracy Gap (RPAG), the difference between accuracy under the original reference and accuracy under the swapped reference; a large positive RPAG indicates that the judge degrades when the reference conflicts with its knowledge. Additional swap variants—type-preserving, type-changing, popularity-based, and evaluator-knowledge swaps—","core_discovery":"The paper's central discovery is that reference-based LLM judges fail when the provided reference answer conflicts with the judge's parametric knowledge. In the authors' controlled setting, they replace the original gold answer with a swapped entity and show that grading accuracy drops sharply: in the most striking case, a judge labels a candidate Incorrect even though both the reference and the candidate say 'Paris' for a question about the Boston Tea Party, while the same question with the original reference ('England') is graded correctly. The drop is most severe when both the reference and the candidate are swapped, and it is substantially reduced when the swapped reference is chosen to","pith_inferences":["A natural next-step test is to add an adversarial 'swap test' to LLM-judge meta-evaluation: a judge that maintains near-zero RPAG on swapped references should be more trustworthy for evolving-knowledge QA.","The finding suggests that reference adherence may need architectural enforcement—for example, grounding the judge's decision in the reference via a separate non-parametric lookup—rather than prompt-level nudges.","If the paper is correct, prior LLM-judge evaluations of QA datasets with time-sensitive facts may have silently underestimated model performance, since judges would reject candidates that contradict their stale knowledge even when the gold reference has been updated."],"forward_implications":["Reference-conditioned QA evaluation is unreliable when gold answers are stale, region-specific, or counterfactual, because judges will quietly overrule the reference.","Scaling model size does not restore reference adherence; larger models may rely even more on parametric knowledge.","Common prompting mitigations—direct instruction, chain-of-thought, self-consistency—do not eliminate the failure and can make it worse.","The severity of the failure is predictable from properties of the swapped entity: more popular entities and long-stable facts cause larger accuracy drops.","Benchmark results obtained with LLM judges must be treated with caution when the underlying ground-truth answers are not in line with the judge's world knowledge."],"fun_headline_variants":["LLM judges fail when reference answers clash with memory","Swapped gold answers expose judges' blind trust in prior knowledge","When references contradict models, judges mark correct as wrong","LLM graders disregard provided answers when they know better","Judge models buckle when reference disputes their internal facts"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper assumes that an ideal LLM judge should always defer to the provided reference, even when that reference is factually wrong or implausible—so a judge that rejects a nonsensical reference is labeled as failing rather than as sensibly refusing.","fun_headline_variants_meta":{"raw":{"variants":["LLM judges fail when reference answers clash with memory","Swapped gold answers expose judges' blind trust in prior knowledge","When references contradict models, judges mark correct as wrong","LLM graders disregard provided answers when they know better","Judge models buckle when reference disputes their internal facts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000199,"raw_usage":{"total_tokens":1189,"prompt_tokens":707,"completion_tokens":482,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":451,"completion_tokens_details":{"reasoning_tokens":415}},"tokens_in":451,"tokens_out":482,"duration_ms":5903,"temperature":1.0,"reasoning_tokens":415,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T11:03:17.704747+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same swapped-reference protocol with human raters independently checking every generated candidate for fluency and semantic faithfulness to its intended reference. If a substantial share of swapped candidates are awkward or accidentally echo the original answer, the accuracy drop could be attributed to candidate quality instead of knowledge conflict. Alternatively, if any judge model shows identical verdicts for original and swapped references on a large sample, its RPAG would be near zero and would directly contradict the paper's central claim.","supporting_citations":[],"review_version":1}