{"id":"e109461b-e4a8-4b3b-b6d8-302eeced3933","arxiv_id":"2508.08285","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"ROUGE-based evaluation inflates the performance of LLM hallucination detectors; with human-aligned metrics, many detectors drop sharply and length-based heuristics match them.","lead":"This study tests whether hallucination detectors in large language models are actually good, using human judgments instead of the usual ROUGE measure. It finds many detectors are much worse than reported, and some simple heuristics like response length perform just as well.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central numerical claims depend on unvalidated LLM-as-Judge; without evidence of its alignment to human judgments, the reported 45.9% drop may be a judge artifact rather than real evaluation failure.","rationale":"The reader's weakest assumption exactly matches the most load-bearing concern: LLM-as-Judge is treated as 'human-aligned' without evidence in the abstract. This is not an internal inconsistency but a missing validation step. The paper's other claims (ROUGE's low precision, length heuristics) are downstream of this assumption. If the judge is biased, every reported number is suspect; if the judge is validated against humans, the central claim gains strong support. The abstract alone cannot settle this, so the verdict remains UNVERDICTED, consistent with the reader's low-confidence assessment. The proposed concrete test would resolve the concern by checking whether the judge actually agrees with humans better than ROUGE does, and whether response length confounds the judge's decisions.","tokens_in":701,"tokens_out":2977,"duration_ms":37034,"concrete_test":"Extract from the paper's dataset a random sample of detector outputs (say 200 items covering all detectors). Obtain human annotations for each item, plus LLM-as-Judge scores (using the exact judge prompt). Compute pairwise agreement (e.g., Cohen's kappa and accuracy) between LLM-as-Judge and human labels, and between ROUGE and human labels. If LLM-as-Judge does not exceed ROUGE in human agreement, or if its disagreement correlates with response length, the 45.9% drop and the length-heuristic result are not interpretable as real performance overestimation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central quantitative claim — that established detectors overperform by up to 45.9% — rests on using 'LLM-as-Judge' as the ground truth. The abstract mentions comprehensive human studies, but only to show ROUGE's low precision; it does not state that the LLM judge was calibrated against human annotations on the same detection instances. If the judge is systematically biased (e.g., by response length, style, or position), then both the magnitude of the drop and the apparent success of length-based heuristics could be artifacts of the judge rather than properties of the detectors. Specifically, a length-based heuristic could score well if the judge equates verbosity with correctness. The paper's conclusion that 'simple heuristics can rival complex detection techniques' would then merely reflect the judge's biases. Without a per-item comparison of LLM-as-Judge labels to human labels, the central claim is not yet supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper re-evaluates hallucination detection evaluation practices in LLMs. It claims that ROUGE-based evaluation, relying on lexical overlap, misaligns with human judgments and systematically overestimates the performance of established hallucination detectors. The abstract reports comprehensive human studies showing ROUGE has high recall but extremely low precision, and that several detection methods drop by up to 45.9% when assessed with 'human-aligned metrics like LLM-as-Judge'. Additionally, the paper claims that simple heuristics based on response length can rival complex detection techniques, suggesting a fundamental flaw in current evaluation practices. The authors conclude that semantically aware and robust evaluation frameworks are needed.","tokens_in":889,"tokens_out":2936,"duration_ms":35470,"significance":"If the central claim holds, the paper would have substantial impact on how hallucination detection systems are evaluated and compared. It poses a falsifiable hypothesis: that ROUGE-based metrics inflate detector performance, and that simple baselines may be much stronger than previously thought. The focus on re-evaluating established methods is timely and important for a community that increasingly relies on detectors for safe LLM deployment. The paper's contribution is potentially high: it challenges an evaluation orthodoxy and offers a concrete replacement direction. However, because I had access only to the abstract, I cannot assess the experimental rigor, dataset sizes, or statistical soundness. The significance of the result is therefore conditional on the full-text validation of the claims, particularly the human alignment of the LLM-as-Judge metric.","major_comments":[{"comment":"The central quantitative claim — that established detectors overestimate performance by up to 45.9% — is calibrated against 'human-aligned metrics like LLM-as-Judge', yet the abstract provides no evidence that this judge agrees with human raters on the detection instances used. Without a per-item comparison (e.g., agreement rates or correlation with human labels) and an analysis of judge biases (e.g., length or verbosity preferences), the reported drop and the success of the length heuristic could reflect properties of the judge rather than the detectors. Please add a validation of the judge against human annotations, including agreement statistics and an error analysis, or restrict the conclusions to what is directly demonstrated.","section":"Abstract, LLM-as-Judge validation"},{"comment":"The abstract states that comprehensive human studies show ROUGE has high recall but extremely low precision, which leads to 'misleading performance estimates'. However, no dataset sizes, ROUGE thresholds, or variability measures are reported. The magnitude of the drop (up to 45.9%) could be a small-sample artifact or sensitive to the choice of datasets and ROUGE versions. The authors should report the number of instances, the exact precision/recall values observed under the human-aligned metric, and confidence intervals or variance estimates for the performance drops across datasets and random seeds.","section":"Abstract, human studies and ROUGE precision"},{"comment":"The claim that 'simple heuristics based on response length can rival complex detection techniques' is stated without defining the competing detectors, the heuristic, or the statistical test used to establish 'rival'. If the length heuristic performs within statistical noise of the detectors, the conclusion is a null result rather than a fundamental flaw. The authors should specify which detectors were compared, the exact heuristic (e.g., response length threshold), and report significance tests with effect sizes to support the claim that the heuristic is not merely comparable in an uninformative way.","section":"Abstract, length-heuristic comparison"}],"minor_comments":[{"comment":"The phrase 'human-aligned metrics like LLM-as-Judge' suggests there are multiple such metrics; please name them in the abstract or main text so that readers know exactly which alternatives to ROUGE are being considered.","section":"Abstract, terminology"},{"comment":"The abstract states 'up to 45.9%' without specifying which detection method and dataset produce this maximum; a breakdown in the main text would improve transparency and help readers assess the variability of the result.","section":"Abstract, reporting of maximum drop"},{"comment":"The phrase 'comprehensive human studies' is not quantified; adding a sentence with participant counts, number of annotated instances, and inter-annotator agreement would make the abstract self-contained and more credible.","section":"Abstract, human studies quantification"}],"recommendation":"uncertain","confidential_remarks":"I was able to review only the abstract, which limits my ability to judge the technical soundness of the paper. The claims are important and clearly stated, and the direction of the research is valuable. However, the central quantitative result depends on the validity of the LLM-as-Judge as a human-aligned metric, and the abstract does not provide the evidence needed to assess that. I recommend sending the full manuscript to reviewers with expertise in evaluation methodology, with a specific request to scrutinize the judge-human agreement analysis and the statistical comparisons. My recommendation is 'uncertain' because the abstract alone does not allow a fully informed verdict."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The part of this that lands immediately is the length-heuristic finding. If simple response-length baselines genuinely rival state-of-the-art hallucination detectors under a semantic or human-aligned metric, that is a real challenge to how the field compares detectors. It would tell us that a lot of reported performance may be picking up surface correlates rather than actual hallucination detection. That is worth taking seriously even before the full numbers are on the table.\n\nWhat the paper does well, on the face of the abstract, is put the ROUGE critique on an empirical footing. The ROUGE-has-high-recall-but-low-precision point is not new in NLG evaluation, but the demonstration that established detectors drop by up to 45.9% under a different metric is a concrete, falsifiable claim. The human studies mentioned in the abstract also give the ROUGE critique some independent weight, assuming those studies actually measured the precision-recall gap on the detection instances.\n\nThe soft spot is exactly where the stress-test note points: the 45.9% number and the success of the length heuristic both depend on the LLM-as-Judge being a faithful proxy for human judgment on the same items. The abstract calls it “human-aligned” but does not say the judge was calibrated against human labels on the detection instances. If the judge has a verbosity bias, the length heuristics could look good for the wrong reason, and the drop could be an artifact of the judge rather than a property of the detectors. This is a legitimate concern, not a fatal one. The abstract does report comprehensive human studies, so it is possible the full paper contains the needed per-item comparison. I cannot settle it from the abstract.\n\nI also want to flag, proportionately, that this is an abstract-only read. I have not seen the experimental details, dataset sizes, statistical tests, or whether they control for judge variance. So my assessment is provisional. But the central claim is important enough, and the design is plausible enough, that a serious referee should see the full text. The right review would push hard on the judge calibration and on whether the length heuristics are truly simple or accidentally encode something semantic.\n\nWho is this for? Anyone working on hallucination detection or on evaluation metrics for LLM outputs. It will be more useful to method developers than to application builders. I would bring it to a reading group, and I would probably cite it once the full version is vetted, but I would not cite the 45.9% number until I have seen the appendix.\n\nRecommendation: send it to peer review. It is exactly the kind of paper that should be refereed rather than desk-rejected, even though my own verdict right now is “unverified, plausible, with a load-bearing assumption that needs scrutiny.”","headline":"The length-heuristic result is a sharp, plausible challenge to current evaluation practice, but the 45.9% drop stands or falls on the LLM-as-Judge calibration, which the abstract alone does not establish.","tokens_in":1393,"tokens_out":1297,"would_cite":false,"duration_ms":17781,"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":"This paper claims that ROUGE-based evaluation systematically overestimates hallucination detector performance, with established methods dropping up to 45.9% under human-aligned LLM-as-Judge scoring, and that length-based heuristics match…","keywords":["hallucination detection","ROUGE","LLM-as-Judge","evaluation metrics","length heuristics","human alignment","semantic evaluation","LLM hallucinations"],"falsifier":"An independent benchmark that replaces LLM-as-Judge with a large set of human labels for the same detector outputs would settle the claim: the reported orderings and the up-to-45.9% drop should reproduce under human ground truth.","tokens_in":543,"feed_emoji":"📉","tokens_out":4214,"duration_ms":44376,"temperature":0.7,"pith_summary":"This paper claims that the standard practice of evaluating hallucination detectors with ROUGE, a lexical-overlap metric, overstates their true quality. Using human studies and an LLM-as-Judge metric, the authors report that several established methods drop by as much as 45.9 percentage points in measured performance. They also find that heuristics based only on response length can perform as well as complex detection methods, which suggests the field's progress measures are partially an illusion. The paper concludes that semantically aware evaluation frameworks are needed before reported detector quality can be trusted.","feed_headline":"ROUGE inflates hallucination-detector performance by up to 45.9%","feed_subtitle":"Human studies and LLM-as-Judge scores reveal that simple length heuristics match complex detectors.","key_machinery":"The argument is carried by the contrast between two scoring instruments: ROUGE, a metric that measures lexical n-gram overlap with a reference text, and LLM-as-Judge, a language-model rating used here as a human-aligned semantic evaluation. The paper uses human studies to anchor this alignment and uses response-length heuristics as a control baseline. ROUGE's high recall but very low precision relative to human judgment is the mechanism that produces inflated performance estimates, because a detector can match many superficial word sequences without detecting actual hallucination.","core_discovery":"The central finding is that ROUGE-based evaluation inflates hallucination detection results because ROUGE favors high recall but suffers from very low precision relative to human judgments. When the same detectors are scored with LLM-as-Judge, a human-aligned metric in the paper's framing, several established methods suffer performance drops up to 45.9%. In addition, the authors report that response-length heuristics rival complex detection methods, implying that current evaluation scores do not measure semantic detection skill. The paper argues this constitutes a fundamental flaw in current evaluation practice and that the community should move to semantically aware, robust evaluation frameworks.","pith_inferences":["If length heuristics are so competitive, a cheap and interpretable sanity check for any hallucination detector would be to compare its output with a length-based baseline on the same test set.","The precise 45.9% figure is judge-dependent, since different LLM-as-Judge prompts or model choices could widen or shrink the gap, so the qualitative ordering of detectors is probably more robust than the exact number.","The findings suggest a broader concern: benchmarks that rely on any lexical-overlap metric, not just ROUGE, may overstate progress in other generation-quality tasks beyond hallucination detection.","A direct follow-up would be to test whether detector rankings change when ground-truth labels come from human raters rather than an LLM judge in a multi-domain benchmark."],"forward_implications":["Established hallucination detectors may be substantially weaker than their published ROUGE-based scores suggest, with measured performance drops up to 45.9 percentage points.","Simple response-length heuristics setting a strong baseline implies that many complex detection methods add little beyond superficial cues.","Future evaluations need semantically aware metrics, and results reported under ROUGE should be re-examined before deployment decisions are made.","Comparisons between detectors should be based on human-aligned metrics to avoid ordering methods incorrectly."],"supporting_citations":[],"fun_headline_variants":["ROUGE inflates hallucination detector performance up to 45.9%","Length heuristics rival complex hallucination detectors","Human-aligned metrics cut hallucination detector scores up to 45.9%","ROUGE's low precision hides true hallucination detection performance","Hallucination detector scores inflated up to 45.9% by ROUGE"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim depends on LLM-as-Judge being a trustworthy stand-in for human judgment; if that judge is itself biased or miscalibrated, the reported performance drops do not reflect real detector quality.","fun_headline_variants_meta":{"raw":{"variants":["ROUGE inflates hallucination detector performance up to 45.9%","Length heuristics rival complex hallucination detectors","Human-aligned metrics cut hallucination detector scores up to 45.9%","ROUGE's low precision hides true hallucination detection performance","Hallucination detector scores inflated up to 45.9% by ROUGE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00104,"raw_usage":{"total_tokens":4313,"prompt_tokens":818,"completion_tokens":3495,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":434,"completion_tokens_details":{"reasoning_tokens":3400}},"tokens_in":434,"tokens_out":3495,"duration_ms":28189,"temperature":1.0,"reasoning_tokens":3400,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:50:05.332540+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An independent benchmark that replaces LLM-as-Judge with a large set of human labels for the same detector outputs would settle the claim: the reported orderings and the up-to-45.9% drop should reproduce under human ground truth.","supporting_citations":[],"review_version":1}