{"id":"0e58d06e-ee4d-48e6-80c9-f29d44d117e9","arxiv_id":"2505.07001","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A new hallucination-annotated GI image-text dataset and a detect-then-correct finetuning strategy improve VLM report scores over standard finetuning on Kvasir-v2 images.","lead":"This paper introduces Gut-VLM, a gastrointestinal endoscopy image dataset with expert-corrected reports and sentence-level hallucination labels, and uses it to train vision-language models to detect and correct their own errors. The authors report that hallucination-aware finetuning beats standard finetuning on report generation, though the comparison is not fully controlled.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The H-model's test-time input is unspecified; if it receives a draft report while standard finetuning generates from the image alone, the reported QAAS gains are confounded.","rationale":"The reader's weakest assumption identifies exactly the same concern: the hallucination-aware finetuned model is evaluated under unspecified and potentially asymmetric inference conditions. My read of the paper confirms that the test-time input is never described in Section 4, and Section 3.3's language suggests the model is trained to correct existing VLM responses, which at test time would require a draft to correct. This is the single most load-bearing issue for the central claim, since it directly threatens the internal validity of the finetunedH versus finetuned comparison. The paper otherwise has real strengths: a novel dataset with expert-verified hallucination tags, sentence-level corrections, a benchmark across four VLMs, and a clear expert evaluation protocol with consistency checking. These support conditional acceptance rather than rejection. I therefore keep the reader's conditional verdict unchanged: the central claim should be accepted only after the authors specify and match the inference protocol and ideally add a controlled ablation and significance testing.","tokens_in":10098,"tokens_out":3007,"duration_ms":28636,"concrete_test":"Require the authors to report the exact test-time prompts for finetuned and finetunedH, and run a controlled ablation on the same 366-image test set: (a) finetunedH with an image-only prompt, (b) standard finetuned with an image-plus-draft correction prompt, and (c) finetunedH with a draft but no image. If condition (a) matches standard finetuning rather than the 90.89 QAAS, the advertised gain comes from the draft input. Additionally, recompute QAAS on the same responses using a non-ChatGPT extractor or human-verified Q&A; if the H-vs-standard gap shrinks, the result is partly an artifact of using ChatGPT to evaluate outputs that inherit ChatGPT-generated structure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—hallucination-aware finetuning outperforms standard finetuning (LLaVA QAAS 90.89 vs 83.07, Table 1)—rests on an unstated inference protocol. Section 3.3 says the H-model is finetuned to 'identify hallucinated sentences and correct the response' from 'existing VLM responses,' and Fig. 2 contrasts standard with hallucination-aware finetuning, but Section 4 never states what input finetunedH models receive at test time. If finetunedH is prompted with the pretrained model's draft report (or ChatGPT's draft) alongside the image and asked to correct it, while standard finetuned models are prompted with the image only, the comparison conflates the training objective with access to a strong textual prior. The gap of 7.8 QAAS points is plausible for a correction task, since the model can preserve the draft's correct structure and clinically relevant wording rather than generating a report from scratch. This concern is amplified by the evaluation: QAAS converts the model's output to Q&A using ChatGPT, and if the H-model's output is a lightly corrected version of the ChatGPT-generated draft, the same LLM extractor may favor its own original phrasing, giving H an additional advantage unrelated to hallucination awareness. No error bars or significance tests are reported, so the observed gap could also reflect run-to-run variation in LoRA finetuning. The limitations section acknowledges single-VLM bias but does not address this test-time asymmetry, which is the load-bearing weakness for the paper's central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Gut-VLM, a multimodal dataset for gastrointestinal (GI) image analysis built from Kvasir-v2 images. Draft diagnostic reports are generated with ChatGPT-4 Omni and then reviewed and corrected by expert gastroenterologists, yielding sentence-level hallucination tags and corrected reports. The authors propose 'hallucination-aware finetuning,' in which a VLM is finetuned to detect and correct hallucinated sentences rather than to generate reports directly, and they report that this strategy outperforms standard finetuning on four VLMs (LLaVA-1.6-7B, DeepSeek-7B-VL, Qwen2-7B, mPLUG-Owl-2B). Evaluation uses classical metrics (ROUGE-L, BLEU, METEOR), two proposed LLM-based metrics (R-Sim and QAAS), and a small expert evaluation. The headline result is Table 1: LLaVA-1.6-7B finetunedH reaches QAAS 90.89% versus 83.07% for standard finetuning and 50.89% for the pretrained model.","tokens_in":10400,"tokens_out":4371,"duration_ms":42736,"significance":"The Gut-VLM dataset is a potentially useful resource for studying hallucination in medical VLMs, as it is, to my knowledge, the first GI-specific dataset with hallucination tags and expert corrections. The proposed hallucination-aware finetuning idea is interesting and, if validated, could provide a practical alternative to standard finetuning for medical report generation. The benchmark covers four modern VLMs and includes both automatic and expert evaluation, which is a credible effort. However, the paper's central claim rests on a comparison whose test-time protocol is unspecified, and the proposed metrics are computed with the same model family used to create the ground truth, so the results as presented are not yet fully convincing.","major_comments":[{"comment":"The test-time inference protocol for the hallucination-aware (finetunedH) models is never stated. Section 4 says only that the model is finetuned 'to learn to detect hallucinated sentences and then correct them,' but it does not say whether, at generation time, the model receives a draft report (e.g., the pretrained model's own output or ChatGPT's output) to correct, or whether it must produce a corrected report from the image alone. If the H-model receives a draft while the standard finetuned model is prompted with the image only, the 7.8-point QAAS gap in Table 1 is confounded with access to a strong textual prior, and the comparison is not on equal footing. The manuscript must specify the test-time input and, if drafts are used, match the inference protocols or include an ablation in which standard finetuning also receives the same draft.","section":"Section 4 / Section 3.3"},{"comment":"The two proposed metrics depend entirely on ChatGPT-4, the same model family that produced the draft reports from which the ground truth was derived. In particular, QAAS first converts each model output into Q&A format using ChatGPT-4; a finetunedH model trained to correct ChatGPT-generated drafts may produce outputs that are lexically and structurally closer to the ChatGPT-generated ground truth, so the ChatGPT-based extractor could favor H-models for reasons unrelated to clinical accuracy. The paper acknowledges single-VLM bias in Section 7, but it does not quantify how much of the reported H-model gain is an artifact of the shared model family. I request an additional analysis that decouples the extractor from the generation family, for example by computing QAAS with a different LLM or by reporting expert evaluation on all models rather than only LLaVA.","section":"Section 5 (QAAS and R-Sim)"},{"comment":"All quantitative results are single point estimates with no error bars, confidence intervals, or significance tests. The headline comparison of QAAS 90.89 versus 83.07 could plausibly lie within run-to-run variability of five-epoch LoRA finetuning, especially given the small rank and the modest test set of 366 images. At minimum, the authors should report standard deviations over multiple seeds or a paired statistical test on the test-set predictions, for both the automatic metrics and the expert ratings on the 30-response sample. The paper itself notes the need for significance testing in Section 7, so this is a known gap rather than an acceptable omission.","section":"Table 1 / Section 6"},{"comment":"Beyond the test-time protocol, the training data format for hallucination-aware finetuning is underspecified. The paper says the model is finetuned to 'identify hallucinated sentences and correct the response' using the VLM-generated response, tags, and corrected response, but it never shows the concrete input/output template (e.g., how the hallucinated sentences are marked, what the instruction text is, and whether the model is trained to output both tags and corrections). Without this information, the method cannot be reproduced, and the reader cannot tell whether the reported gains come from the correction objective or from a different prompt formatting.","section":"Section 3.3"}],"minor_comments":[{"comment":"References [25] and [26] are duplicates: both cite Liu et al., 'Mitigating hallucination in large multi-modal models via robust instruction tuning.' The duplicate should be removed and the numbering adjusted.","section":"References"},{"comment":"The phrase 'Descriptive F ormat Diagnostic Report' contains an erroneous space in 'F ormat'; this should be corrected.","section":"Section 6"},{"comment":"Model names are written inconsistently: 'LLaVa-1.6-7b' in Table 1 and 'LLaVA-1.6-7B' in the text and elsewhere. Please standardize, and use the correct official casing for each model.","section":"Throughout"},{"comment":"The statistic 'only 30.39% of the VLM-generated responses are fully correct' (Section 3.2) is reported without defining 'fully correct.' A sentence describing how a response is classified as fully correct versus mixed or fully hallucinated would improve clarity.","section":"Figure 1"},{"comment":"The expert evaluation is reported only for LLaVA-1.6-7B and only on 30 responses per model variant; the paper should state explicitly that expert ratings are not available for the other three models and should provide the sampling procedure and any inter-rater or intra-rater measures beyond the coefficient of variation.","section":"Section 6, Expert Evaluation"}],"recommendation":"major_revision","confidential_remarks":"The central ambiguity about the test-time input is fixable, and the dataset contribution is real, so I do not recommend rejection. However, the paper as submitted overclaims the method result: there is a plausible confound (draft access), the metrics share a model family with the ground-truth generation, and there are no significance tests. If the authors clarify the protocol and add a matching ablation, the paper may become acceptable. I would also encourage the editor to check whether the duplicate reference [25]/[26] indicates a more general citation-checking issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper introduces Gut-VLM, a GI-specific dataset built on Kvasir-v2 with ChatGPT-generated reports, sentence-level hallucination tags, and expert corrections. That part is genuinely useful: it's the first GI multimodal hallucination dataset I know of, and the annotation pipeline is cost-sensitive and thoughtful. The benchmark on four VLMs with the proposed R-Sim and QAAS metrics is also a reasonable starting point, even if the metrics lean on ChatGPT.\n\nThe main claim is that hallucination-aware finetuning beats standard finetuning for report generation. I'm not convinced yet, because the paper never says what input the finetunedH model receives at test time. Section 3.3 frames training as 'identify hallucinated sentences and correct the response,' but the evaluation section is silent on whether at inference the model is given a draft report (its own pretrained output, or ChatGPT's) to correct, or generates from the image alone. If it gets a draft, the 7.8-point QAAS gap is easily explained by access to a strong textual prior, and the comparison is unfair. This is not a nitpick; it's the load-bearing comparison of the paper. Also, DeepSeek finetunedH actually has lower R-Sim than standard finetuning (3.63 vs 3.76), so 'outperforms across all metrics' is an overstatement. No error bars or significance tests are reported, so run-to-run variance is unknown.\n\nThere's also circularity: ground truth reports are corrected ChatGPT outputs, and QAAS extracts Q&A from model outputs via ChatGPT. The authors acknowledge ChatGPT bias in descriptive scores, and the same extractor is applied to all models, which helps, but if the H-model's output is a lightly edited ChatGPT draft, the extractor will favor its own phrasing. That compounds the confound.\n\nWhat the paper does well is the dataset construction and the explicit sentence-level annotation of hallucinations, plus an honest limitations section about single-VLM bias. The idea of finetuning on detection and correction is reasonable and worth pursuing.\n\nMy recommendation: this deserves a serious referee, but the central claim needs to be re-evaluated after the authors specify the exact inference protocol and match inputs across conditions. Also add error bars or significance tests. If those are fixed, the dataset alone makes the paper worth publishing.","headline":"A useful new GI hallucination dataset, but the paper's central finetuning claim is confounded by an unspecified, likely asymmetric test-time protocol.","tokens_in":10977,"tokens_out":2849,"would_cite":true,"duration_ms":27690,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Finetuning a vision-language model to detect and correct hallucinated sentences produces more accurate gastrointestinal diagnostic reports than finetuning on corrected text alone.","keywords":["multimodal medical dataset","gastrointestinal image analysis","vision-language model","hallucination","hallucination-aware finetuning","medical report generation","Kvasir-v2","visual question answering"],"falsifier":"Run the released evaluation code and record the exact inputs given to the finetunedH models at test time. If those inputs include a draft report from ChatGPT or the model's own pretrained output, rerun standard finetuning with the same draft input; if the QAAS gap between finetunedH and finetuned disappears, the method's advantage was caused by input asymmetry rather than by hallucination-aware training.","tokens_in":9895,"feed_emoji":"🩺","tokens_out":12063,"duration_ms":112282,"temperature":0.7,"pith_summary":"Hallucination—model output that does not match the image—plagues vision-language models, and the paper argues that the standard fix, finetuning on corrected reports, leaves accuracy on the table. The paper instead trains models to detect and correct hallucinated sentences, an approach it calls hallucination-aware finetuning. The supporting resource is Gut-VLM, a set of 1,816 gastrointestinal endoscopy images paired with machine-generated draft reports, sentence-level hallucination labels, and expert corrections. Across four open VLMs, hallucination-aware finetuning consistently scores higher than standard report finetuning on the paper's 12-question accuracy metric; for LLaVA-1.6-7B it reaches 90.89 percent versus 83.07 percent. If correct, this offers a cheap route to more reliable medical VLMs: let a general model draft, have experts mark its errors, and train a smaller model to repair them.","feed_headline":"Fix-the-hallucination training beats standard report finetuning","feed_subtitle":"A 7B vision-language model that learns to repair errors answers 90.9% of diagnostic questions, versus 83.1% for standard finetuning.","key_machinery":"The carrying mechanism is hallucination-aware finetuning: with low-rank adaptation, the model is trained on image-plus-draft inputs where part of the text is marked hallucinated and the target is the expert-corrected report, so the model must first identify the erroneous sentences and then regenerate a corrected version. This contrasts with standard finetuning, where the target is simply the corrected ground-truth report. The dataset supplies the essential supervision: ChatGPT-generated drafts over Kvasir-v2 images, sentence-level hallucination tags, expert corrections, and a diagnostic Q&A format derived from 12 predefined questions, which also powers the QAAS evaluation metric.","core_discovery":"Gut-VLM is built by asking a commercial VLM to describe Kvasir-v2 endoscopy images under 12 diagnostic questions, then having expert gastroenterologists mark every hallucinated sentence and provide its correction; only 30.39 percent of the machine-generated reports were fully correct. The paper's central claim is that using these sentence-level tags as training signal—hallucination-aware finetuning—improves test-time report generation more than finetuning on the corrected reports alone. On the Question Answering Accuracy Score (QAAS), which converts each generated response into answers to the same 12 questions and checks them against expert-corrected ground truth, the best hallucination-aware model, LLaVA-1.6-7B, scores 90.89 percent versus 83.07 percent for standard finetuning and 50.89 percent for the pretrained model. The same ordering appears in all four open models tested, and a small expert evaluation also ranks the hallucination-aware variant above standard finetuning.","pith_inferences":["A fair replication should give the hallucination-aware and standard models identical test-time inputs; the paper does not state whether the H-model receives a draft report to correct, so part of the reported gap could come from draft access rather than from the training objective.","If correction-learning transfers, the same recipe could be applied to other medical imaging domains where expert text is scarce, such as chest radiography or pathology: generate drafts, have experts tag errors, and finetune to repair them.","The sentence-level tags also invite a grounding test: aligning hallucinated sentences with image regions could show whether correction training changes what the model actually attends to, not just what it writes."],"forward_implications":["Across all four open VLMs tested, hallucination-aware finetuning outperforms standard finetuning on QAAS, so the training objective, not a particular model, appears to drive the improvement.","Gut-VLM turns hallucination detection into a directly supervised task: sentence-level tags and corrections can be used to train a detector independent of report generation.","The two-stage pipeline (machine draft, expert correction) can lower the cost of building medical image-text datasets, since experts review rather than write from scratch.","The same annotations power a second task: finetuned on extracted diagnostic Q&A, LLaVA-1.6-7B reaches 87.91 percent QAAS versus 49.26 percent pretrained."],"supporting_citations":[{"why":"Supplies the 1,816 Kvasir-v2 endoscopy images that Gut-VLM annotates.","marker":"[34]"},{"why":"Generates the draft diagnostic reports whose hallucinated sentences are later tagged and corrected.","marker":"[1]"},{"why":"Provides the sentence-level hallucination labeling convention used by the expert annotators.","marker":"[12]"},{"why":"Defines the 12 diagnostic questions that structure the reports and drive the QAAS evaluation.","marker":"[13]"},{"why":"Provides the low-rank adaptation method used in every finetuning experiment.","marker":"[14]"},{"why":"The open vision-language model that achieves the highest hallucination-aware QAAS and is also evaluated on VQA and by an expert.","marker":"[28]"}],"fun_headline_variants":["Hallucination-aware finetuning lifts GI report accuracy","Training on tagged errors beats standard report finetuning","Gut-VLM: error-tag training outperforms standard FT","Detect-and-correct finetuning wins for GI VLMs","Hallucination tags in finetuning yield 90.9% QAAS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes both finetuning strategies are evaluated with the same inputs, but the paper never states what the hallucination-aware model receives at test time; if it gets a draft report while the standard model gets only the image, the reported advantage may not be from the training objective.","fun_headline_variants_meta":{"raw":{"variants":["Hallucination-aware finetuning lifts GI report accuracy","Training on tagged errors beats standard report finetuning","Gut-VLM: error-tag training outperforms standard FT","Detect-and-correct finetuning wins for GI VLMs","Hallucination tags in finetuning yield 90.9% QAAS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000885,"raw_usage":{"total_tokens":3889,"prompt_tokens":1081,"completion_tokens":2808,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":697,"completion_tokens_details":{"reasoning_tokens":2719}},"tokens_in":697,"tokens_out":2808,"duration_ms":18285,"temperature":1.0,"reasoning_tokens":2719,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:27:06.296412+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released evaluation code and record the exact inputs given to the finetunedH models at test time. If those inputs include a draft report from ChatGPT or the model's own pretrained output, rerun standard finetuning with the same draft input; if the QAAS gap between finetunedH and finetuned disappears, the method's advantage was caused by input asymmetry rather than by hallucination-aware training.","supporting_citations":[{"cited_title":"In: Proceedings of the AAAI Conference on Artificial Intelligence (2024)","cited_arxiv_id":null,"evidence_quote":"Provides the sentence-level hallucination labeling convention used by the expert annotators."},{"cited_title":"In: Conference and Labs of the Evaluation Forum (2023)","cited_arxiv_id":null,"evidence_quote":"Defines the 12 diagnostic questions that structure the reports and drive the QAAS evaluation."}],"review_version":1}