{"id":"04b67007-409b-46d4-b6a2-b9c639f18207","arxiv_id":"2412.18947","paper_version":4,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"The paper proposes MedHallBench and ACHMI for medical hallucination measurement, but provides no dataset or code, and ACHMI is an uncredited replication of CHAIR.","lead":"This paper describes MedHallBench, a benchmark for measuring when medical AI models invent or distort findings, and proposes a scoring metric called ACHMI. If validated and released, it could help developers and hospitals compare how often different AI systems fabricate medical information.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ACHMI is load-bearing but not well-defined: it counts 'medical components' without a component taxonomy, extraction protocol, or reported inter-rater agreement, so the Table 1 scores cannot be interpreted or reproduced.","rationale":"The reader's weakest assumption identifies the same load-bearing issue: ACHMI presupposes that medical components in chest X-ray captions can be unambiguously identified, enumerated, and expert-validated, but the paper provides no component taxonomy, annotation protocol, or inter-rater agreement values. I agree with that assessment. The paper does mention Cohen's Kappa and a three-type hallucination classification, but it never operationalizes 'medical components,' so Equations (1) and (2) cannot be computed independently. The CHAIR provenance is also relevant: the text explicitly says ACHMI is an advanced variant of CHAIR, but the reference list omits CHAIR, and the medical adaptation is exactly where the missing specification matters. The reader's additional concerns about unavailable data and code are secondary but reinforce the same conclusion: the benchmark and metric cannot be externally checked. I am not claiming fraud or misconduct; the paper may reflect a private annotation process that simply is not described. But as submitted, the central claim that ACHMI provides a more nuanced understanding of medical hallucination is unsupported, and the REJECT verdict remains appropriate. A concrete re-annotation study with released data would settle whether the concern lands; absent release, the concern stands.","tokens_in":10945,"tokens_out":3966,"duration_ms":37700,"concrete_test":"Release MedHallBench and the exact model-generated captions underlying Table 1, together with the annotation guidelines. Have two board-certified radiologists independently annotate medical components in a random sample of 100 captions following the guidelines; compute Cohen's Kappa and recompute ACHMI_I and ACHMI_S for each annotator's component sets. If Kappa is below 0.7 or the two ACHMI score sets differ by more than a few percentage points, the metric's core assumption fails. If the data cannot be released, the authors should at least publish the component vocabulary and a fully worked example; without either, the central claim remains unverifiable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim is that ACHMI 'provides a more nuanced understanding of the effects of hallucinations compared to traditional metrics' (abstract). Equations (1) and (2) define ACHMI_I and ACHMI_S as ratios of hallucinated medical components to all medical components, and of captions containing hallucinated components to all captions. The metric is meaningful only if the set of 'medical components' in a caption and in the reference is well-defined. The paper says ACHMI is an 'advanced variant of the CHAIR metric' (Experiment, Evaluation Metrics) but does not provide the medical-component ontology, normalization or synonym rules, or extraction method needed to instantiate CHAIR for radiology. The Methodology section describes a 'hierarchical classification system' and says inter-rater reliability was measured with Cohen's Kappa, but no taxonomy, no annotation protocol, and no Kappa values are reported; the 'Hallucination Annotation' section is programmatic rather than operational. Without a reproducible definition of the components counted in the numerator and denominator, the ACHMI numbers in Table 1 are uninterpretable, and the claimed advantage over BLEU and ROUGE cannot be evaluated. This is a correctness risk in the central argument, not merely a scope limitation, because the benchmark's primary novel contribution is that metric.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MedHallBench, a benchmark for evaluating hallucinations in medical vision-language models, together with a new metric, ACHMI (Automatic Caption Hallucination Measurement in Medical Imaging), defined over 'medical components' in generated captions. The authors describe expert and lay evaluation frameworks, an automatic annotation pipeline using active learning and RLHF, and they report an experimental comparison of ten existing models on VQA and image-report-generation tasks using ACHMI, BERTScore, METEOR, ROUGE-1/2/L, and BLEU. The central claim is that ACHMI provides a more nuanced assessment of hallucination than traditional text-matching metrics.","tokens_in":11168,"tokens_out":6845,"duration_ms":62362,"significance":"If the operational details and validation were supplied, this work would address a real gap: hallucination evaluation in medical image captioning. The idea of adapting CHAIR-style object-level hallucination metrics to medical components is promising, and the comparison of ten models is in principle useful. The paper should also be credited for tackling a clinically important problem and for attempting to combine automated scoring with human evaluation. However, the current manuscript does not deliver the central artifact: there is no concrete definition of 'medical components', no annotation protocol with reported reliability, no expert-evaluation results, no dataset statistics or release, and no evidence connecting ACHMI to human judgments. These omissions are load-bearing because ACHMI and the claimed expert validation are the paper's main contributions.","major_comments":[{"comment":"The ACHMI metric is not operationalized. Equations (1) and (2) define ACHMI_I and ACHMI_S as ratios of hallucinated medical components to all medical components and of captions containing hallucinated components to all captions, but the paper does not define what counts as a medical component, how components are extracted from a generated caption, how synonyms and normalization are handled, or how the set of ground-truth components is obtained. The 'Hallucination Annotation' subsection describes a hierarchical classification system and says Cohen's Kappa was measured, but no taxonomy, annotation template, normalization rules, or Kappa values are reported. As a result, the ACHMI values in Table 1 cannot be reproduced or interpreted, and the abstract's claim that ACHMI is more nuanced than BLEU/ROUGE cannot be assessed.","section":"Experiment, Evaluation Metrics, Eqs. (1)-(2)"},{"comment":"The expert and lay evaluation components are described programmatically but never reported. The text states that experts use 5-point Likert scales for clinical accuracy, harm severity, and hallucination confidence, that expert and lay evaluations are weighted 0.7 and 0.3, and that inter-rater reliability is measured with Cohen's Kappa, but the paper gives no expert evaluation results, no Kappa values, no annotator counts, no composite scores, and no comparison between expert scoring and automatic ACHMI scoring. The claim that MedHallBench is 'expert-validated' is therefore unsupported by the presented evidence.","section":"Methodology, Hallucination Annotation and Framework for Hallucination Evaluation"},{"comment":"The RLHF pipeline is disconnected from the experiments. The 'Modeling' section describes a PPO-based RLHF training procedure, but Table 1 evaluates existing models, not a model trained by that pipeline, and no results show that the RLHF-based or active-learning annotation methods produce valid labels. Since these annotation methods are listed among the paper's main contributions, the paper needs either to report results validating the automatic annotation or to remove the claim that the benchmark uses RLHF-based automatic annotation.","section":"Modeling and Experiment"},{"comment":"The conclusion that ACHMI is more effective than traditional metrics rests on qualitative inspection of a single table. There are no error bars, no significance tests, no correlation with human judgments, and no case analysis connecting ACHMI to the three announced hallucination types (anatomical, pathological, measurement). To support the claim of a 'more nuanced understanding', the authors should report a comparison of ACHMI with expert ratings and analyze cases where ACHMI and BLEU/ROUGE diverge.","section":"Experiment, Experimental Results Analysis, Table 1"}],"minor_comments":[{"comment":"The acronym MLLMs is defined as 'Medical Large Language Models', but MLLM is widely used for 'Multimodal Large Language Models'; please disambiguate the terminology.","section":"Abstract and throughout"},{"comment":"Model names such as 'LLaV A1.5-7b' and 'LLaV A-Med' contain a spurious space; they should be 'LLaVA1.5-7b' and 'LLaVA-Med'.","section":"Table 1 and Figure 1"},{"comment":"The sentence 'All the above models were fine-tuned on the Slake (SF) dataset' is confusing, since Table 1 includes models like BLIP2 and MiniGPT4; clarify what 'SF' means for each model and what fine-tuning procedure was used.","section":"Experiment, Models"},{"comment":"The text says quality control includes 'statistical monitoring of inter-annotator agreement', but no such statistics appear anywhere in the paper; at least the Cohen's Kappa values should be reported.","section":"Hallucination Annotation"},{"comment":"Equation (2) writes '|{all caption}|'; this should be '|{all captions}|'.","section":"Evaluation Metrics"},{"comment":"The contributions mention an active-learning annotation method, but the Methodology section does not describe it; add a description or remove the claim.","section":"Introduction, Contributions"},{"comment":"Several references have formatting errors, including 'MIMIC-Ext-MIMIC-CXR-VQA' in the Bae et al. entry; the reference list should be carefully cleaned.","section":"References"},{"comment":"The paper never reports basic dataset statistics, such as the number of images, cases, QA pairs, or generated captions, nor does it state where the dataset will be released; a benchmark paper should provide this information.","section":"Conceptualization of MedHallBench"}],"recommendation":"major_revision","confidential_remarks":"The core idea is viable, but the current submission is closer to a proposal than a completed benchmark: ACHMI lacks an operational definition, the human evaluation is described but not delivered, and the RLHF annotation pipeline has no experimental support. I would be willing to reconsider a substantially revised version that supplies the missing taxonomy, annotation protocol, reliability statistics, dataset statistics, and validation analyses. As it stands, I cannot recommend acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, quick take on MedHallBench. The headline is that the claimed novel metric, ACHMI, is a relabeled CHAIR metric, and the benchmark that supposedly carries the paper is neither released nor described with enough specificity to reproduce. The paper has a real topic and a sensible three-way hallucination taxonomy for radiology, but the central claims don't survive contact with the evidence.\n\nWhat's genuinely useful here: the paper surveys existing medical QA benchmarks and flags the gap in hallucination evaluation for multimodal medical models. The taxonomy separating anatomical, pathological, and measurement hallucinations is reasonable and worth stealing. The authors also ran ten models on both VQA and report generation tasks, which is more than many benchmark papers do – though without error bars that comparison is only suggestive.\n\nThe soft spots are large. Equations (1) and (2) are exactly CHAIR_I and CHAIR_S with 'medical components' swapped in for 'objects'. The paper even calls ACHMI 'an advanced variant of the CHAIR metric' but doesn't cite Rohrbach et al. That is a citation problem, and it undermines the novelty claim. More importantly, the metric is not operationalized: there is no definition of what counts as a medical component, no normalization or synonym rules, no extraction protocol, and the promised Cohen's Kappa numbers never appear. Without that, the ACHMI values in Table 1 are uninterpretable, and the claim that ACHMI provides a more nuanced understanding than BLEU or ROUGE is simply asserted. The RLHF-based automatic annotation pipeline is described at length but never implemented or evaluated. The dataset itself – the core promised deliverable – is not released, and its composition is only described as a combination of MIMIC-CXR, MedQA, and 'custom case scenarios'.\n\nThe citation pattern is otherwise fine; the self-citations are to their own prior work and not problematic.\n\nBottom line: this is a paper with a useful framing and a potentially usable taxonomy, but the load-bearing parts are either unoriginal or unverified. I would not send it to a serious peer review yet. If the authors release the dataset, report Kappa and error bars, operationalize the component taxonomy, and cite CHAIR, there could be a decent workshop or short paper here. Until then, the cautious move is to treat it as a secondary source for the taxonomy only.\n\nReading group: no, except maybe as a cautionary example.","headline":"ACHMI is a relabeled CHAIR metric, the benchmark is unreleased and unspecified, and the evaluation lacks the reliability statistics that would make the metric meaningful; the taxonomy is useful but the paper is not ready for review.","tokens_in":11733,"tokens_out":5088,"would_cite":false,"duration_ms":40841,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that medical hallucinations are best measured by counting hallucinated components in captions rather than by text-similarity scores.","keywords":["medical large language models","hallucination evaluation","benchmark","ACHMI","chest X-ray captioning","RLHF","vision-language models","automatic annotation"],"falsifier":"Have two independent clinical teams enumerate hallucinated medical components in the same set of model captions using a fixed written taxonomy, then compare their ACHMII and ACHMIS values; low inter-rater agreement or score divergence beyond a pre-set clinical threshold would refute the claim that ACHMI is a stable measurement.","tokens_in":10701,"feed_emoji":"🩻","tokens_out":7507,"duration_ms":63898,"temperature":0.7,"pith_summary":"This paper claims that medical hallucinations in vision-language models are better measured by counting hallucinated medical components in generated captions than by traditional text-similarity metrics. It introduces MedHallBench, a benchmark built from expert-validated chest X-ray cases and existing medical question-answer data, plus an automatic annotation pipeline. The central result is ACHMI, an adaptation of the CHAIR metric that computes the fraction of hallucinated medical components (ACHMII) and the fraction of captions containing any hallucination (ACHMIS). Applied to ten models, ACHMI ranks models differently from BLEU, ROUGE, METEOR, and BERTScore, and the paper argues this gives a more nuanced view of clinical hallucination. If right, this gives researchers a scalable way to spot and reduce unsafe model outputs before deployment.","feed_headline":"Counting hallucinated findings spots errors BLEU misses","feed_subtitle":"The ACHMI score counts hallucinated components in chest X-ray captions, finding failures BLEU and ROUGE miss.","key_machinery":"The load-bearing object is ACHMI (Automatic Caption Hallucination Measurement in Medical Imaging), a medical adaptation of the CHAIR metric. Given an expert-validated list of ground-truth medical findings in a chest X-ray, ACHMI counts every medical component in the model's caption, marks those that do not correspond to ground truth as hallucinations, and forms two ratios: ACHMII (hallucinated components over all generated medical components) and ACHMIS (captions containing at least one hallucinated component over all captions). The metric's work is to separate 'says something wrong' from 'says something off-topic or poorly worded', which traditional n-gram and embedding similarity scores conflate.","core_discovery":"On its own terms, the paper establishes that hallucination assessment in medical image captioning should be component-level rather than surface-text-level. ACHMI, the proposed metric, asks whether each medical component in a generated caption (an anatomical structure, a pathological finding, a measurement) matches expert-validated findings in the image; it then reports the proportion of hallucinated components and the proportion of captions that contain them. The paper's empirical comparison across ten vision-language models shows that models with respectable lexical-overlap scores can still carry high ACHMI values, and models fine-tuned on the Slake dataset score markedly lower on ACHMI. From this the authors conclude that ACHMI captures hallucination effects that traditional metrics miss and should guide training and fine-tuning of medical LLMs.","pith_inferences":["The paper names CHAIR as the basis for ACHMI but does not cite it, and it does not publish the component taxonomy, annotation protocol, or inter-rater agreement values; a reader who wants to reuse ACHMI would first need to fix and validate that taxonomy.","ACHMI as defined weights every hallucinated component equally, so a missed life-threatening finding and an extra harmless anatomical detail would affect the score the same way; a clinically weighted variant would be safer for deployment decisions.","The benchmark mixes international sources such as MIMIC-CXR and MedQA with an stated alignment to mainland China's medical conditions, so users should check whether the component definitions transfer across healthcare systems before generalizing conclusions.","A direct testable extension would be to compare ACHMI against expert harm ratings on the same captions: if ACHMI values do not track clinical severity, it should be paired with a severity-weighted measure rather than used alone."],"forward_implications":["Adopting ACHMI as a reporting standard would let medical imaging benchmarks compare models by their rate of false findings rather than by text quality alone.","Fine-tuning and RLHF pipelines can use ACHMI as a reward signal or early-stopping criterion to actively reduce hallucinated findings.","Because ACHMI needs only ground-truth component lists, applying the same component-level logic to CT, MRI, or pathology captions is a direct extension of the paper's approach.","The published baseline for ten mainstream models gives future work a reference point for measuring hallucination improvement.","Using ACHMII and ACHMIS together separates whether a model hallucinates often from whether it hallucinates in every caption, which is information BLEU and ROUGE do not provide."],"supporting_citations":[{"why":"Supplies the MIMIC-CXR visual question answering dataset that anchors the chest X-ray component of MedHallBench.","marker":"Bae et al. 2024"},{"why":"Supplies the MedQA (USMLE) question-answer corpus used to build the medical case scenarios.","marker":"Shi et al. 2023"},{"why":"Contributes MultiMedQA question-answer pairs that are expert-annotated and added to the benchmark.","marker":"Qian et al. 2024"},{"why":"Defines hallucination in chest radiograph interpretation and provides XrayGPT, a baseline model in the comparison.","marker":"Thawkar et al. 2023"},{"why":"Provides InstructBLIP-7b/13b, baseline models whose outputs are scored with ACHMI and traditional metrics.","marker":"Dai et al. 2023"},{"why":"Provides LLaVA-7b/13b baselines and the visual instruction tuning recipe used by several compared models.","marker":"Liu et al. 2023a"},{"why":"Supplies the factually augmented RLHF approach that the paper's automatic annotation pipeline draws on.","marker":"Sun et al. 2023"},{"why":"Provides BLIP-2, a baseline model in the MedHallBench comparison.","marker":"Li et al. 2023c"},{"why":"Med-HALT is the prior medical hallucination benchmark whose scope limitations the paper argues MedHallBench addresses.","marker":"Pal, Umapathi, and Sankarasubbu 2023"}],"fun_headline_variants":["MedHallBench: New yardstick for medical AI hallucinations","Counts fake findings in X-ray captions that BLEU overlooks","Benchmark reveals hallucination blind spots in medical LLMs","Component-level scoring catches errors BLEU and ROUGE miss","New medical benchmark exposes untruthful AI captions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Every ACHMI score depends on a written list of 'medical components' that can be counted in a caption and checked against expert ground truth; if two clinical teams cannot reliably produce the same list, the metric's numbers are not stable.","fun_headline_variants_meta":{"raw":{"variants":["MedHallBench: New yardstick for medical AI hallucinations","Counts fake findings in X-ray captions that BLEU overlooks","Benchmark reveals hallucination blind spots in medical LLMs","Component-level scoring catches errors BLEU and ROUGE miss","New medical benchmark exposes untruthful AI captions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000283,"raw_usage":{"total_tokens":1664,"prompt_tokens":928,"completion_tokens":736,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":649}},"tokens_in":544,"tokens_out":736,"duration_ms":6903,"temperature":1.0,"reasoning_tokens":649,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:18:21.022387+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Have two independent clinical teams enumerate hallucinated medical components in the same set of model captions using a fixed written taxonomy, then compare their ACHMII and ACHMIS values; low inter-rater agreement or score divergence beyond a pre-set clinical threshold would refute the claim that ACHMI is a stable measurement.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes MultiMedQA question-answer pairs that are expert-annotated and added to the benchmark."}],"review_version":1}