{"id":"a9e8f43e-77de-4540-ad5b-4e1b1dce89f2","arxiv_id":"2511.12110","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MediRound introduces a multi-round, entity-level medical segmentation task, a 177K-dialogue dataset built from SA-Med2D-20M with GPT-5, and a LLaVA-Med/MedSAM baseline whose inference-time judgment-and-correction module reduces error propagation across rounds.","lead":"Researchers built a 177,000-conversation dataset and a model, MediRound, that segments structures in medical images while remembering and referring to masks from earlier turns in the same dialogue. The key open question is whether the model's success on its own template-generated benchmark will hold up on real, free-form medical conversations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Potential train/test image overlap in MR-MedSeg could inflate reported multi-round segmentation scores.","rationale":"The reader's weakest assumption was that MR-MedSeg test conversations are representative of real multi-round dialogues, a legitimate external-validity concern about template memorization. My concern is different and more load-bearing: the train/test split may not be image-disjoint, which threatens internal validity of the reported benchmark numbers. If test images were seen during training, the central claim that MediRound solves MEMR-Seg is not supported even on the authors' own benchmark. This is objectively checkable once the dataset is released, and the paper currently provides no evidence of image-level separation. A condition on verifying image-disjointness is therefore necessary. Since the reader already issued CONDITIONAL, my additional concern does not change the verdict—it adds a specific, potentially fatal condition that must be checked before acceptance.","tokens_in":15296,"tokens_out":4825,"duration_ms":46075,"concrete_test":"Obtain the MR-MedSeg dataset (from the released repository or authors) and check whether any image ID (or image hash) appears in both the training split and the test split. If overlap exists, re-run MediRound evaluation on the subset of test conversations whose images are absent from training. If Dice/gIoU/cIoU drop substantially (e.g., >5 points cIoU), the reported multi-round improvements are likely inflated by image leakage. If no overlap exists, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that MediRound 'effectively addresses the MEMR-Seg task'—rests on the integrity of the MR-MedSeg test split. The paper reports 177K conversations over 118K images, i.e., ~1.5 conversations per image on average, and specifies only a conversation-level split (174,934/1,270/1,273, §3 Data Statistics). It never states that the split is image-disjoint. If the same image appears in both training and test (likely under a random conversation-level split), MediRound may have been trained on the exact test images and their entity masks. In that case, the high Dice/gIoU/cIoU numbers, especially the multi-round gains (Table 2), could reflect image/entity memorization rather than generalization of cross-round reasoning. This is a direct internal-validity threat: the benchmark would not be a clean held-out evaluation, undermining the paper's primary empirical support. The missing appendix and unreleased dataset make the split design unverifiable, so the concern cannot be dismissed from the manuscript alone.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces MEMR-Seg, a new task for multi-round, entity-level reasoning segmentation in medical images, and constructs MR-MedSeg, a large dataset of 177K conversations derived from SA-Med2D-20M with GPT-5-assisted template-based generation. The authors propose MediRound, an MLLM-based architecture (LLaVA-Med + MedSAM) that conditions on dialogue history, a referred-round crop, and bounding-box information, plus a Judgment & Correction Mechanism (JCM) that refines low-quality [SEG] features at inference. Experiments on MR-MedSeg report large gains over prior single-round methods and hybrid MLLM+segmentation baselines, and a single-round external evaluation on SA-Med2D-20M shows competitive performance. The central claim is that end-to-end multi-round conditioning and JCM mitigate error accumulation and enable accurate cross-round entity segmentation.","tokens_in":15397,"tokens_out":2606,"duration_ms":29101,"significance":"If the MR-MedSeg evaluation is valid, the paper makes a useful contribution by defining a new task, releasing a substantial dataset, and providing a reproducible baseline with a clearly specified architecture and inference procedure. The strengths are the concrete equations (Eqs. 1–2), Algorithm 1, detailed hyperparameters, and the external anchor on SA-Med2D-20M (Table 3), where MediRound reaches 62.1 Dice versus 61.2 for MediSee. However, the significance depends heavily on the integrity of the MR-MedSeg split and the representativeness of the template-generated test set. The current manuscript does not establish image-disjointness between training and test conversations, and the test distribution is generated by the same pipeline as the training distribution. These issues directly threaten the paper's main empirical claims, so the contribution is currently promising but not yet established.","major_comments":[{"comment":"The paper states the split only at conversation level: 174,934/1,270/1,273 conversations, with 177K conversations over 118K images (~1.5 conversations per image). It does not state that the split is image-disjoint. Under a random conversation-level split, the same image and its entity masks will very likely appear in both training and test. Since MediRound receives the full input image and can memorize image–mask associations, the reported test Dice/gIoU/cIoU (Table 1) and the per-round gains (Table 2) may reflect image memorization rather than multi-round reasoning generalization. Please (a) clarify whether any image appears in both training and test, and (b) if not already done, re-run the evaluation with an image-disjoint split or report the overlap statistics and results on a disjoint subset.","section":"§3 Data Statistics"},{"comment":"The test conversations are produced by the same semi-automatic pipeline as the training conversations: 50–80 GPT-5-generated semantically equivalent templates per entity relationship, with entities manually inserted. The test set is therefore drawn from the same template manifold as the training set. The paper's practical-utility claims (§1: medical education, continuous clinical interaction) require generalization to free-form user queries and unseen phrasings. A model could achieve high scores by pattern-matching on template structure rather than by genuinely resolving cross-round entity references. Please provide an out-of-distribution evaluation (e.g., held-out template families, paraphrases written by humans, or real clinical queries) or analyze the diversity of the test set to show that template leakage is not driving the results.","section":"§3 Step 3 / Data Statistics"},{"comment":"The 'Human-Thinking + X' protocol is not specified. No details are given for how human experts phrase multi-round queries, what conversation history is provided, how mistakes or corrections are handled, how many human annotators participated, or how the MLLM-hybrid baselines are prompted. Since these baselines are a central part of the 'outperforms conventional medical referring segmentation methods' claim, the comparison cannot be assessed for fairness. Please report the complete protocol, including prompts, error handling, and annotator expertise, or explicitly reposition these numbers as upper-bound anchors rather than direct comparisons.","section":"§5.2 Human-Thinking baselines"},{"comment":"The Quality Judgment and Correction Modules are described as lightweight MLPs, but the training objective, loss function, and data labeling for q (the quality score) are not given in the main text and are deferred to an appendix that is not included in the submission. Without this information the JCM ablation (Figure 7) and the final test improvement (58.4 vs 55.8 Dice) cannot be fully interpreted. Please include the JCM training details in the main paper or a clearly accessible appendix, and specify whether β is selected on the validation set and then fixed for the reported test results.","section":"§4 Algorithm 1 / JCM training"}],"minor_comments":[{"comment":"In Eq. (1), G_v^enc(x_crop) and G_b^enc(x_box) are written as inputs to G_i, but Figure 4 suggests these are concatenated with text/image embeddings. Clarify the concatenation/embedding order in the equation or caption.","section":"§4 Eq. (1)"},{"comment":"The β sweep is reported on the validation set without error bars or multiple seeds. State whether β=0.6 was chosen on validation and then frozen for the test numbers in Table 1; otherwise the reported test improvements could include selection bias.","section":"§5.5 Figure 7"},{"comment":"The symbol 'T' in the table header is not defined in the main text. Add a footnote explaining what 'independent model' means and why SegLLM and MediRound are marked as such.","section":"Table 1"},{"comment":"The paper cites GPT-4's technical report for GPT-5-based generation. If a specific GPT-5 model or version was used, cite the actual model or API documentation; otherwise clarify the exact version and access date.","section":"References"},{"comment":"Several details, including scenario definitions, data examples, and JCM training, are deferred to 'the Appendix,' but no appendix is present in the submitted version. Please include the appendix in the revised submission so that the claims can be verified.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The main unresolved issue is internal validity: the conversation-level split may allow train/test image overlap, and the synthetic template pipeline likely creates a distribution where pattern matching suffices. These are not fatal if the authors can provide an image-disjoint split and an out-of-distribution test; both are within the scope of a revision. I would also flag that the dataset is not released and the appendix is absent, which makes independent verification difficult. If the split issue cannot be fixed, the paper would need substantial re-scoping of its claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper introduces a genuinely new task (multi-round entity-level reasoning segmentation) and a large synthetic dialogue dataset, plus a baseline model with an inference-time correction module. That part is worth taking seriously. The architecture is assembled from prior pieces — LLaVA-Med, MedSAM, SegLLM-style [SEG] conditioning — but the medical instantiation and the JCM are new, and the single-round anchor on SA-Med2D-20M (62.1 Dice) shows the backbone is competitive with existing medical referring segmenters.\n\nThe problem is the evaluation. The paper reports only a conversation-level split of MR-MedSeg (174,934/1,270/1,273) and never states that the split is image-disjoint. With 177K conversations over 118K images, that is about 1.5 conversations per image. A random conversation split will very likely put the same image — and the same entity masks — in both training and test. If so, the multi-round gains in Tables 1 and 2 partly reflect the model memorizing the exact test images, not generalizing the cross-round reasoning. The stress-test note is right: this is a load-bearing validity threat and it cannot be dismissed from the manuscript.\n\nA second, related concern is that the test set is generated through the same GPT-5 template pipeline used for training. The model could be pattern-matching on template distribution rather than doing transferable reasoning. The human-guided baselines are also under-specified — prompt details, error handling, and the expertise of the human are not described. And no error bars or significance tests are reported. The missing appendix and unreleased code/data make none of this independently checkable.\n\nFor all that, this is not a sloppy paper. The method is clearly specified, the ablations are sensible, and the authors honestly flag the teacher-forcing discrepancy and the error-prone human baselines. The central problem is evaluation validity, not internal incoherence.\n\nThe paper deserves peer review, but a referee should ask for an image-disjoint split, free-form or externally collected test dialogues, and release of the dataset. Right now I would not cite the reported numbers or use the benchmark as-is. I would bring it to a reading group to talk about what counts as a held-out evaluation for template-generated datasets.","headline":"Useful new task and dataset, but the multi-round evaluation may be compromised by a non-image-disjoint train/test split.","tokens_in":16048,"tokens_out":2389,"would_cite":false,"duration_ms":23029,"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":"MediRound demonstrates that medical image segmentation can be driven by multi-round queries that refer back to previously segmented entities, with an inference-time correction mechanism preventing errors from compounding.","keywords":["multi-round segmentation","medical image segmentation","entity-level reasoning","referring segmentation","dialogue history","reasoning segmentation","error correction","multimodal LLM"],"falsifier":"Collect a small set of naturally occurring multi-round clinical or educational Q&A sequences (e.g., from medical teaching files or clinician notes) and run MediRound on them, comparing against the template-based test set; if the gap over single-round baselines shrinks sharply on such queries, the template distribution explains the result.","tokens_in":15019,"feed_emoji":"🩻","tokens_out":3288,"duration_ms":30065,"temperature":0.7,"pith_summary":"The paper introduces a new task, Multi-Round Entity-Level Medical Reasoning Segmentation, where each query refers back to an entity segmented in a previous round. To study it, the authors built a 177K-conversation dataset of medical image dialogues with five types of cross-round entity relationships. They propose MediRound, which feeds conversation history, a crop of the previously referred mask, and the bounding box into a multimodal language model that outputs a [SEG] token whose hidden feature is decoded into a mask. The claimed result is that MediRound outperforms prior single-round medical segmentation methods and a general multi-round reasoning segmenter, with the largest gains in later rounds. The practical importance is that interactive, education-style medical image questioning becomes possible if these results hold in real use.","feed_headline":"Multi-round medical segmentation holds up via history plus correction","feed_subtitle":"A new model and 177K-dialogue benchmark let users refer back to earlier masks in medical conversations.","key_machinery":"The load-bearing machinery is the [SEG] hidden feature from the multimodal LLM, combined with a reference-aware input. For the current round, MediRound concatenates the text of the current query and prior dialogue, the full image, and the cropped region from the mask cited in the query, together with its bounding-box coordinates. That joint embedding goes into the LLM; the [SEG] token's last-layer hidden feature is decoded into the mask by a medical segmentation head. The Judgment & Correction Mechanism then acts as an inference-time gate: an MLP scores the feature, and if the score falls below a threshold, a second MLP rewrites the feature before decoding, so later rounds receive a cleaner","core_discovery":"The central discovery is that error accumulation in chain-like multi-round segmentation can be mitigated at inference time without retraining the end-to-end model. The paper shows that a lightweight two-MLP Judgment & Correction Mechanism, applied after training, judges the quality of the [SEG] hidden feature each round and replaces low-quality features with corrected ones before they are used as reference for the next round. With the mechanism, the model's cIoU on the MR-MedSeg validation set rises from 55.8 to 58.9 overall, and from 46.1 to 54.8 at round 8, indicating that gains concentrate precisely where error propagation should be worst. This is presented as evidence that end-to-end con","pith_inferences":["The test set is generated by the same template-based pipeline as the training set, so the reported scores likely upper-bound performance on unscripted queries; a natural-language paraphrase challenge would be a more stringent test.","The five relationship types form a taxonomy of cross-round references that could transfer to other medical imaging settings, such as pathology or 3D volumes, where anatomical constraints differ.","Because the correction module is threshold-gated, logging quality scores per round and correlating them with human-perceived query difficulty would be a direct test of whether the gate measures true uncertainty.","The 'Human-Thinking + X' baseline protocol lacks a published prompt and human-expertise specification; documenting it would clarify what the 15-point performance gap actually measures."],"forward_implications":["If correct, interactive medical education tools could let learners ask follow-up questions that refer to previous segmentations and get coherent masks.","The inference-time correction suggests error propagation in any chain-like referring segmentation can be partly offset without training a new model.","The dataset provides a fixed benchmark for comparing multi-round medical reasoners across relationship types such as organ–lesion, anatomical hierarchy, spatial, attribute, and strong inference.","MediRound also remains competitive on single-round referring segmentation, so the multi-round capability does not sacrifice standard utility.","Per-round results show that with the correction mechanism the model holds up past round six, the regime where naïve pipelines typically collapse."],"fun_headline_variants":["Fix errors on the fly to boost multi-round medical segmentation","Self-correcting pipeline keeps medical segmentation on track","Medical segmentation that improves with each round","Error correction at inference time enhances multi-round segmentation","Inference-time fix stops error cascades in medical segmentation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the MR-MedSeg test conversations — generated from 50–80 paraphrased templates per entity relationship on one underlying medical image collection — represent the distribution of real multi-round medical reasoning dialogues; if real users phrase queries outside that template manifold, the reported gains may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Fix errors on the fly to boost multi-round medical segmentation","Self-correcting pipeline keeps medical segmentation on track","Medical segmentation that improves with each round","Error correction at inference time enhances multi-round segmentation","Inference-time fix stops error cascades in medical segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000967,"raw_usage":{"total_tokens":3934,"prompt_tokens":708,"completion_tokens":3226,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":452,"completion_tokens_details":{"reasoning_tokens":3166}},"tokens_in":452,"tokens_out":3226,"duration_ms":21189,"temperature":1.0,"reasoning_tokens":3166,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T22:05:27.314803+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect a small set of naturally occurring multi-round clinical or educational Q&A sequences (e.g., from medical teaching files or clinician notes) and run MediRound on them, comparing against the template-based test set; if the gap over single-round baselines shrinks sharply on such queries, the template distribution explains the result.","supporting_citations":[],"review_version":1}