{"id":"f1eef8b3-7b3a-45cc-b964-d675fd687dbf","arxiv_id":"2605.28910","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Hallucination detector-guided iterative revision plus preference optimization reduces hallucinations by 24-48% in Llama-3.1-8B clinical summarization on MIMIC-IV notes while preserving fluency and coherence.","lead":"The paper introduces an inference-time method that uses hallucination detectors to iteratively revise LLM-generated clinical summaries and converts those trajectories into preference pairs for model fine-tuning. A smart generalist might read it because factual errors in medical note summaries limit safe use of AI in healthcare settings.","discovery_kind":"unclear","skeptic_critique":{"model":"grok-4.3","headline":"Hallucination reductions rest on unvalidated detector accuracy; human eval covers only fluency/coherence/relevance","rationale":"The reader's weakest_assumption directly identifies the detector-accuracy risk; the abstract provides no counter-evidence (human hallucination labels or detector validation metrics) that would neutralize it. Full-text methods would need to show either (a) detector F1 > 0.85 against human labels or (b) separate human hallucination eval to move beyond CONDITIONAL. This matches the reader's assessment without introducing new concerns.","tokens_in":1725,"tokens_out":318,"duration_ms":15790,"concrete_test":"On a 100-note held-out MIMIC-IV subset, obtain independent human expert annotations of hallucination counts (with inter-annotator agreement) for the three conditions (base, iTermodel, model); recompute reduction percentages against these labels. If human-annotated reductions fall below 10% or lose significance, the detector-driven claim does not hold.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim quantifies hallucination reduction (24% for iTermodel, 48% for iTermodel) via detector-guided trajectories on MIMIC-IV notes. The abstract states human/LLM-Jury eval only for fluency, coherence, and relevance; no independent human hallucination annotation is described. If the detector systematically undercounts hallucinations in revised outputs or overcounts in baselines, the percentage reductions are artifacts. This is the exact condition the methods require to be true.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper introduces iTermodel, an inference-time method that uses hallucination detectors to iteratively revise clinical summaries, and iTermodel for Preference Learning (iTermodel), which converts detector-guided refinement trajectories into preference pairs for LLM fine-tuning. On summarizing real-world clinical notes from MIMIC-IV, the methods are reported to reduce hallucinations by 24% (iTermodel) and 48% (iTermodel) for Llama-3.1-8B-Instruct (with similar gains for Gemma models), while preserving fluency, coherence, and relevance per human expert and LLM-Jury evaluations.","tokens_in":1817,"tokens_out":563,"duration_ms":35075,"significance":"If the hallucination reductions hold under independent validation, the work provides a practical, automated pipeline for improving factual faithfulness in clinical summarization without degrading other summary qualities. The detector-guided preference optimization approach could be impactful for safety-critical domains where hallucinations limit LLM deployment.","major_comments":[{"comment":"Abstract and §4 (Experiments): The central quantitative claims of 24% and 48% hallucination reduction are derived solely from the hallucination detector outputs; the manuscript provides no independent human annotation of hallucination presence/absence in the revised vs. baseline summaries to confirm these percentages reflect true factual improvements rather than detector bias or error.","section":"Abstract and §4"},{"comment":"§3.1 (iTermodel) and §3.2 (Preference Pair Construction): The methods depend on the detector being sufficiently accurate to both guide revisions and label preference pairs, yet no precision/recall figures, error analysis, or ablation on detector quality are reported for the MIMIC-IV summarization domain; this is load-bearing because systematic under-detection in revised outputs would artifactually inflate the reported gains.","section":"§3.1 and §3.2"},{"comment":"§4.3 (Human and LLM-Jury Evaluation): Evaluations are restricted to fluency, coherence, and relevance; the absence of a direct hallucination-focused human study means there is no cross-check against the detector-based metric that underpins the primary results.","section":"§4.3"}],"minor_comments":[{"comment":"The abstract introduces the method names without immediate expansion; ensure the first use in the introduction spells out the full names for clarity.","section":"Abstract"},{"comment":"Table captions and experimental setup descriptions should explicitly list all baselines, detector variants, and statistical tests used for the percentage reductions.","section":"§4"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We appreciate the referee's thorough review and the opportunity to address these important points regarding our evaluation of hallucination reductions. We provide detailed responses to each major comment and indicate the revisions we will make to strengthen the manuscript.","responses":[{"response":"We acknowledge that the 24% and 48% reductions are measured via the hallucination detector, which enables scalable, consistent quantification across the MIMIC-IV test set. The human expert and LLM-Jury evaluations in §4.3 were intentionally focused on fluency, coherence, and relevance to verify that detector-guided revisions do not degrade other summary properties. Because the iterative process explicitly targets and corrects detector-flagged spans, the reduction in detected hallucinations is a direct outcome of the method. We will add a limitations paragraph clarifying the reliance on the detector and noting that future work could include targeted human hallucination annotation; no such annotation is added in this revision due to resource constraints.","revision_made":"partial","referee_comment":"Abstract and §4: The central quantitative claims of 24% and 48% hallucination reduction are derived solely from the hallucination detector outputs; the manuscript provides no independent human annotation of hallucination presence/absence in the revised vs. baseline summaries to confirm these percentages reflect true factual improvements rather than detector bias or error."},{"response":"The detector is drawn from prior published work whose general-domain performance is documented in its source paper. We did not report MIMIC-IV-specific precision/recall or conduct an ablation in the submitted manuscript. In the revision we will insert a short error-analysis subsection that (a) cites the detector’s original metrics, (b) discusses the risk of domain shift, and (c) notes that any systematic under-detection would affect both baseline and revised outputs equally, thereby preserving the relative reduction. If space permits, we will also report a small manual spot-check on a subset of samples.","revision_made":"partial","referee_comment":"§3.1 and §3.2: The methods depend on the detector being sufficiently accurate to both guide revisions and label preference pairs, yet no precision/recall figures, error analysis, or ablation on detector quality are reported for the MIMIC-IV summarization domain; this is load-bearing because systematic under-detection in revised outputs would artifactually inflate the reported gains."},{"response":"We agree that a dedicated human hallucination annotation study would constitute an independent cross-check. Our current protocol instead uses expert review of overall summary quality together with an LLM-Jury to confirm that revisions preserve (or improve) readability and clinical utility. We will revise §4.3 and the discussion section to explicitly state the absence of hallucination-specific human labels and to frame this as a natural direction for follow-up work. The combination of detector-driven quantitative gains and preserved qualitative scores remains the core evidence presented.","revision_made":"partial","referee_comment":"§4.3: Evaluations are restricted to fluency, coherence, and relevance; the absence of a direct hallucination-focused human study means there is no cross-check against the detector-based metric that underpins the primary results."}],"tokens_in":1423,"tokens_out":670,"duration_ms":33624,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper's main contribution is an inference-time method that uses hallucination detectors to iteratively revise clinical summaries, followed by a preference learning version that turns those revision trajectories into training pairs. On Llama-3.1-8B-Instruct summarizing MIMIC-IV notes, it claims 24% reduction from the inference method and 48% from the fine-tuned one, while human experts and an LLM jury see no drop in fluency, coherence, or relevance.\n\nWhat is new is the end-to-end use of detector outputs to both correct summaries on the fly and generate preference data for the clinical domain. It takes existing detection tools and alignment techniques and applies them in a loop tailored to medical note summarization.\n\nThis has value because it offers an automated path to better factual faithfulness without extra human labels. The choice of real-world clinical notes from MIMIC-IV grounds the experiments in a high-stakes setting where hallucinations matter.\n\nThe soft spot is the lack of independent verification for the hallucination reductions. Since the same detectors guide the revisions and score the outputs, any systematic error in the detector could make the improvements look larger than they are. The evaluations from humans and the jury focus only on style and relevance, not on whether the content is more accurate or less hallucinated. Without details on detector performance or separate factuality checks, the percentage claims are difficult to trust fully.\n\nThis work is for researchers focused on making LLMs safer for healthcare applications. A reader interested in practical ways to combine detection and preference tuning would find the pipeline worth considering.\n\nThe paper shows honest engagement with the problem and has enough structure to warrant review. I would send it to peer review so that the methods and any additional analyses can be examined closely.","headline":"The paper chains hallucination detectors into iterative revision at inference time and then into preference pairs for fine-tuning on clinical notes, with reported reductions that rest on the detector's own outputs.","tokens_in":2331,"tokens_out":435,"would_cite":false,"duration_ms":37780,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Detector-guided iterative revisions and preference pairs derived from them reduce hallucinations in clinical note summaries by up to 48 percent.","keywords":["hallucination reduction","clinical summarization","preference optimization","factual faithfulness","LLM alignment","MIMIC-IV"],"falsifier":"Independent clinicians counting unsupported statements in summaries produced by the fine-tuned models on a new set of MIMIC-IV notes would show no reduction relative to the base models.","tokens_in":2629,"feed_emoji":"🩺","tokens_out":623,"duration_ms":25893,"temperature":0.7,"pith_summary":"Large language models frequently insert unsupported statements when summarizing clinical notes, limiting their use in healthcare. The paper tests an inference-time process that applies hallucination detectors to iteratively revise summaries toward factual corrections. It then converts the resulting revision trajectories into preference pairs and uses them to fine-tune the model. Experiments on real-world notes from the MIMIC-IV database show that the inference-time method cuts hallucinations by 24 percent and the preference-learning method cuts them by 48 percent on Llama-3.1-8B-Instruct, with human experts and LLM juries reporting no drop in fluency, coherence, or relevance. The same pattern holds for Gemma models.","feed_headline":"Detector-guided revisions cut clinical summary hallucinations by 48%","feed_subtitle":"Preference pairs built from those revisions let models learn factual corrections on real hospital notes without losing coherence","key_machinery":"Hallucination-detector-guided iterative revision trajectories converted into preference pairs for model fine-tuning.","core_discovery":"Detector-guided iterative refinement produces factual corrections at inference time; converting those refinement trajectories into preference pairs for direct preference optimization produces models that generate summaries with substantially fewer unsupported statements on clinical notes.","pith_inferences":["The same detector-guided trajectory approach could be tested on other medical text-generation tasks such as discharge-instruction generation or radiology report expansion.","Further gains would require detectors whose error patterns do not overlap with the errors the model is already making.","The preference pairs created this way are synthetic and could be combined with human preference data to test whether mixed training yields additional improvements."],"forward_implications":["The inference-time revision method alone reduces hallucinations by 24 percent on Llama-3.1-8B-Instruct while preserving summary quality.","The preference-learning method reduces hallucinations by 48 percent on the same model with no measured loss in fluency, coherence, or relevance.","Both methods apply across Llama and Gemma model families when summarizing real clinical notes from MIMIC-IV.","Detection-informed refinement and preference learning together provide an automated route to higher factual faithfulness in clinical summarization."],"fun_headline_variants":["Detector-guided preference optimization reduces hallucinations in clinical notes","Refinement trajectories from detectors yield factual clinical summary models","Hallucination detectors inform preference pairs for accurate clinical summarization","Detection-guided iterative revisions improve factual faithfulness in LLMs"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The hallucination detectors used to guide revisions and create preference pairs are accurate enough that they do not systematically miss errors or introduce new ones that would invalidate the measured reductions.","fun_headline_variants_meta":{"raw":{"variants":["Detector-guided preference optimization reduces hallucinations in clinical notes","Refinement trajectories from detectors yield factual clinical summary models","Hallucination detectors inform preference pairs for accurate clinical summarization","Detection-guided iterative revisions improve factual faithfulness in LLMs"]},"model":"grok-4.3","cost_usd":0.006096,"raw_usage":{"total_tokens":2837,"prompt_tokens":582,"num_sources_used":0,"completion_tokens":62,"cost_in_usd_ticks":60962000,"prompt_tokens_details":{"text_tokens":582,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2193,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":582,"tokens_out":62,"duration_ms":23380,"temperature":1.0,"reasoning_tokens":2193,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T12:26:30.123657+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Independent clinicians counting unsupported statements in summaries produced by the fine-tuned models on a new set of MIMIC-IV notes would show no reduction relative to the base models.","supporting_citations":[],"review_version":1}