{"id":"707b2240-a9e1-4717-8910-3c48b1963340","arxiv_id":"2504.20898","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A demo system combines concept bottleneck models with multi-agent retrieval-augmented generation for chest X-ray report generation, but it provides no quantitative evidence for its claims.","lead":"The paper describes a prototype that turns chest X-rays into readable reports by first identifying visible medical concepts and then using a team of AI agents to write the report. A live demo and code are shared, but no measurements show that the system is accurate or truly reduces false information.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cross-modal cosine similarity between CheXagent image embeddings and Mistral text embeddings is not established as meaningful; since concept scores, heatmaps, and contribution scores all derive from this similarity matrix, the paper's interpretability claim lacks foundation.","rationale":"The central claim is that CBM-RAG delivers interpretable predictions, mitigates hallucinations, and generates high-quality tailored reports. The most load-bearing condition is that the concept bottleneck actually measures clinical concepts. Section 2 establishes this through cosine similarity between CheXagent image embeddings and Mistral text embeddings, yet no argument or evidence shows these embeddings share a semantic space. CheXagent is a chest X-ray VLM for generation and interpretation, not a contrastive image-text encoder; Mistral embed is an unrelated text embedding model. Cosine similarity across such spaces is not a proxy for semantic similarity. Because the concept vector, contribution scores, and saliency heatmaps are all derived from this matrix, the interpretability mechanism fails if the assumption fails. This is a correctness risk independent of the missing quantitative evaluation; even a demo with user studies would inherit the flaw. The reader's verdict of REJECT is therefore appropriate, and I do not propose changing it. I also note the paper's own conclusion admits formal usability studies are lacking, and hallucination mitigation is asserted rather than measured, but those are secondary. The concrete test I propose would settle whether the similarity matrix carries signal; if it does, the concern is removed. My agreement with the reader's weakest_assumption is complete: it is the same embedding-alignment issue.","tokens_in":4777,"tokens_out":4055,"duration_ms":43446,"concrete_test":"Independent reimplementation test: on a labeled subset of COVID-QU (or a public CXR dataset such as CheXpert), compute the cosine similarity between each CheXagent image embedding and the Mistral text embeddings for all concepts. For each concept, measure whether similarity ranks images with the corresponding finding above images without it (e.g., per-concept AUC). If the per-concept AUC is at chance level (≈0.5) while a same-space baseline using, say, CheXagent's own text encoder recovers signal, the cross-model similarity matrix is not semantically grounded. This single check would invalidate or support the entire CBM concept score and heatmap pipeline.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2 computes a similarity matrix as cosine similarity between image embeddings produced by CheXagent [2] and text embeddings produced by the Mistral embed model [5]. These are independent models with no described shared embedding space, no projection, and no alignment step. Cosine similarity is only a semantic measure within a single embedding space; across arbitrarily trained encoders, the relative ordering of similarities is not meaningful. The concept vector is formed by max-pooling this matrix, the normalized vector feeds the fully connected classifier, contribution scores are computed as the cross-product of the classifier weight matrix and this vector, and saliency heatmaps are derived from the same matrix. Consequently, every interpretable output the paper advertises—editable concept scores, contribution scores, heatmaps, and the 'transparent disease classification' in the conclusion—inherits this unvalidated assumption. If the assumption fails, the CBM component does not measure the intended clinical concepts, and the system's interpretability claim collapses. The paper provides no calibration check, no qualitative examples comparing concept scores against known labels, and no classification accuracy, so the reader cannot infer validity from reported results. This is the most load-bearing concern because it attacks the mechanism itself, not just the absence of evaluation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents CBM-RAG, a framework that couples a Concept Bottleneck Model with a multi-agent Retrieval-Augmented Generation pipeline for chest X-ray report generation. The CBM component computes cosine similarity between CheXagent image embeddings and Mistral text embeddings for an LLM-generated set of clinical concepts, max-pools the resulting matrix into a normalized concept vector, and feeds it to a fully connected classifier trained on the COVID-QU dataset. Contribution scores and saliency heatmaps are derived from the classifier weights and the similarity matrix. The multi-agent RAG component uses five agents (three disease-specific ReAct agents, a Radiologist Agent, and a Report Writer Agent) with NIH database retrieval to produce final reports. The authors describe an interactive UI with editable concept scores, heatmap toggles, and a chat interface. The abstract and conclusion claim interpretable predictions, mitigation of hallucinations, and generation of high-quality tailored reports; the conclusion admits that formal usability studies have not yet been conducted.","tokens_in":5076,"tokens_out":3259,"duration_ms":36134,"significance":"If the claims were substantiated, integrating concept-bottleneck interpretability with multi-agent RAG would be a useful direction for clinical decision support. The paper explicitly ships a public code repository and an online demo, which is a strength for reproducibility and community uptake. However, the contribution as written is a system demonstration: no classification accuracy, no report quality metric, no hallucination measurement, no concept-validity check, and no clinician or user study appear anywhere in the manuscript. The interpretability mechanism also rests on an unvalidated cross-modal embedding similarity assumption. The significance of the work therefore currently lies in the proposed architecture and interactive interface rather than in established empirical evidence.","major_comments":[{"comment":"The similarity matrix is computed as cosine similarity between image embeddings from CheXagent and text embeddings from the Mistral embed model. These are independent models trained separately with no described shared embedding space, projection, or alignment step. Because the concept vector, the normalized input to the fully connected classifier, the contribution scores, and the saliency heatmaps all derive from this matrix, the interpretability claim rests entirely on the unvalidated assumption that cross-model cosine similarity is semantically meaningful. The paper should either replace this mechanism with a jointly trained multimodal encoder (e.g., BiomedCLIP, which is already cited) or provide a calibration experiment showing that similarity scores correlate with ground-truth concept presence on a held-out set. Without such evidence, the concept scores and heatmaps may not measure the intended clinical concepts.","section":"Section 2, similarity matrix computation"},{"comment":"The abstract states that the framework 'deliver[s] interpretable predictions, mitigate[s] hallucinations, and generate[s] high-quality, tailored reports,' and the conclusion repeats that the framework 'mitigat[es] hallucination and opacity issues.' However, the manuscript reports no classification accuracy, no report quality metric (e.g., BLEU, ROUGE, or a clinical correctness metric), no hallucination measurement, and no radiologist or user study. The Conclusion itself concedes that 'formal usability studies in real clinical settings are yet to be conducted.' These claims therefore exceed the evidence presented. The authors should either add a quantitative evaluation of report quality and hallucination rate for the actual system or substantially temper the abstract and conclusion to describe intended capabilities rather than demonstrated ones.","section":"Abstract and Conclusion, evaluation claims"},{"comment":"The fully connected layer trained on COVID-QU is described but no training details, hyperparameters, or evaluation results are provided. The paper never reports classification accuracy, precision, recall, F1, or a confusion matrix, and it does not compare against a standard CXR classifier or an uninterpretable baseline. Without this information, the statement that the system provides 'transparent disease classification' is unsupported. At minimum, the authors should report classification performance on the COVID-QU test split and examine whether the concept bottleneck substantially degrades accuracy relative to a non-interpretable classifier, since that trade-off is central to the CBM promise.","section":"Section 2, disease classification component"},{"comment":"The concept set is generated automatically by an LLM, but the paper gives no details about the concept vocabulary: the number of concepts, the prompt used, whether concepts were deduplicated or filtered, or whether the set was validated against a medical ontology or clinician feedback. If the concept set is incomplete, contains spurious terms, or uses inconsistent phrasing, the contribution scores and heatmaps will be misleading even if the embedding alignment issue is resolved. The authors should describe how the concept set was generated, show the actual concepts used, and provide some evidence of medical completeness or correctness.","section":"Section 2, LLM-generated concept set"}],"minor_comments":[{"comment":"The paper refers to the 'Mistral embed model' but cites Mistral 7B, which is a language model and not specifically an embedding model. Clarify the exact embedding model name and version used for the text embeddings to support reproducibility.","section":"References [5] and Section 2"},{"comment":"The interface description would benefit from explicit mention of whether the editable concept scores are constrained to the [0,1] range and how the system reacts when a user edits a concept score that conflicts with the original similarity-based value.","section":"Section 3, user interface"},{"comment":"The multi-agent RAG description is brief; providing the prompts used for the ReAct agents and the retrieval top-k or reranking strategy would help readers understand how the report content is grounded and how hallucination is supposedly reduced.","section":"Section 2, RAG system"},{"comment":"Reference [3] is the COVID-QU dataset paper and is appropriate, but the paper should also state which train/test split was used and whether the class distribution was balanced, since this affects any future comparisons.","section":"References"}],"recommendation":"reject","confidential_remarks":"This is a demonstration paper with no quantitative evaluation and a fundamentally unvalidated relationship between the two embedding spaces. The load-bearing issues are fixable in principle (switch to a joint embedding model, add classification and report-quality experiments, temper the claims), but as it stands the manuscript does not meet the evidence bar for a journal publication. The authors should be encouraged to submit to a workshop or demo track after adding a concept-validity experiment and a small user study. The novelty relative to the authors' own ECIR paper should also be clarified explicitly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a three-page demo companion, not a scientific paper. The architecture is from the same authors' ECIR 2025 paper; what's new here is the chat interface and the editable concept scores. The abstract claims the system 'delivers interpretable predictions, mitigates hallucinations, and generates high-quality, tailored reports.' That is more than the demonstration supports.\n\nWhat's good: the code and demo are public and the UI is genuinely interactive. You can upload a CXR, see concepts, toggle heatmaps, edit concept scores, and watch the predicted class change in response. For a demo, that is a solid artifact. The multi-agent RAG with five specialized agents is also concretely implemented, and the design choice to let clinicians adjust concept scores is a useful feature for exploring trust.\n\nNow the soft spots. The biggest problem is the complete absence of evaluation. There is no classification accuracy, no report quality metric, no hallucination measurement, and no clinician study. The conclusion itself says 'formal usability studies in real clinical settings are yet to be conducted.' So the abstract's 'mitigates hallucinations' is not backed by any evidence. That is not a minor omission; it is the main claim.\n\nThe second soft spot is exactly what the stress-test flags: the cross-modal cosine similarity between CheXagent image embeddings and Mistral text embeddings. These are independent models with no shared embedding space, no projection, and no alignment step. Cosine similarity is only semantic within one space. Without some calibration or qualitative check against known labels, the concept scores, contribution scores, and heatmaps are not established as meaningful. This is load-bearing because every interpretable output the paper advertises derives from that similarity matrix. I checked the paper for a counterexample or a sanity check—there is none.\n\nNovelty is thin relative to the authors' own ECIR paper, but for a companion demo track that is acceptable. The paper is honest about the lack of user studies, so it is not deceptive—it is just over-claimed in the abstract and underevaluated.\n\nWho is this for? Someone building demo systems for clinical AI might find the UI and the agent scaffolding a useful starting point. Someone looking for evidence that CBM plus RAG reduces hallucinations or improves interpretability will not find it here.\n\nRecommendation: I would not send this to a full peer-review venue as a research contribution. The evidence gap is too wide. For a demo track, it is fine—the artifact is inspectable and the interactive editing idea is worth showing. But the claims in the abstract need to be scaled way back unless the authors add even basic quantitative results.","headline":"A clean demo companion with a live UI, but the abstract's claims about mitigating hallucinations are unsupported and the cross-modal embedding assumption is unvalidated.","tokens_in":5543,"tokens_out":2734,"would_cite":false,"duration_ms":30384,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"AI radiology reports become traceable to editable clinical concepts.","keywords":["Interpretable radiology report generation","Disease classification","Medical imaging","Concept Bottleneck Models","Retrieval-Augmented Generation","Information Retrieval","Vision-language models","Large language models"],"falsifier":"Take a set of chest X-rays with expert-annotated findings and compute the concept scores and heatmaps: if the heatmap for 'pulmonary consolidation' does not localize to the annotated consolidation, or if the cosine-similarity ranking between the image and concept texts changes substantially when the two embeddings come from a jointly trained multimodal model, the grounding claim collapses. A second check is intervention fidelity: if editing a concept score leaves the predicted disease unchanged for most images, the editable-interpretability claim fails.","tokens_in":4536,"feed_emoji":"🩻","tokens_out":8194,"duration_ms":78789,"temperature":0.7,"pith_summary":"This paper argues that radiology report generation becomes both more interpretable and more reliable when the image is first distilled into human-understandable clinical concepts, and only then handed to a retrieval-augmented report writer. The authors build a pipeline in which a vision-language model scores concepts such as 'pulmonary consolidation' for a chest X-ray, a classifier reads only those scores to predict the disease, and five specialized agents retrieve clinical knowledge and write a tailored report. The point of the design is to give clinicians something they can inspect and edit before the text is produced: each concept has a contribution score and a saliency heatmap, and changing a score changes the downstream prediction. If the approach works, it addresses the two obstacles that keep generative models out of the reading room, namely opaque reasoning and hallucinated findings. The paper is a demonstration and notes that formal usability studies in real clinical settings are still future work.","feed_headline":"AI radiology reports become traceable to editable clinical concepts","feed_subtitle":"A concept bottleneck plus retrieval agents aims to stop hallucinations by grounding each finding in visible, changeable evidence.","key_machinery":"The central object is the concept bottleneck: a layer that forces the final prediction to depend only on a vector of human-understandable clinical concepts rather than on raw image features. In this implementation the concept set is generated by an LLM, each concept is embedded with a text encoder, the image is embedded with a chest-X-ray vision-language model, and cosine similarity between the two embedding sets forms a matrix; max pooling condenses the matrix into a concept vector, a fully connected layer classifies from it, and the weight-vector cross-product produces contribution scores while the similarity matrix doubles as the source of saliency heatmaps. This single bottleneck carries the whole interpretability argument, and the editable concept list is the mechanism by which a clinician can intervene in the prediction.","core_discovery":"On its own terms, the central discovery is a workflow: placing a concept bottleneck in front of a multi-agent retrieval-augmented generator, with an interactive editing loop joining the two. The concept bottleneck maps each image to a normalized vector of clinically meaningful concept scores, classifies the disease from that vector alone, and derives both per-concept contribution scores and saliency heatmaps from the same similarity matrix. The multi-agent stage then takes those concepts and retrieved clinical documents to produce a final report, so the report text is anchored to externally sourced evidence rather than generated from the image only. The authors claim this yields transparent disease classifications, tailored evidence-based reports, and reduced hallucination and opacity compared with end-to-end image-to-report generation.","pith_inferences":["The most important unstated risk is that the concept scores are computed by cosine similarity between image embeddings from one model and text embeddings from a different model; those embedding spaces were never aligned, so the similarity ranking may not mean what it appears to mean. A cross-encoder or jointly trained multimodal model would provide a direct comparison.","The concept set itself is assumed to be medically complete because an LLM generated it; a missing concept cannot enter the bottleneck, so the system could silently ignore a finding a radiologist would consider essential. An expert coverage audit would test this.","The hallucination-mitigation claim is asserted rather than measured. A controlled benchmark comparing this pipeline with a plain image-to-report baseline on factual-consistency metrics would turn the claim into a number.","The authors state that formal usability studies in real clinical settings remain future work, so the trust and workflow benefits are, for now, demonstrated in an interactive prototype rather than established in practice."],"forward_implications":["A radiologist can audit a predicted disease by reading the concepts that drove it and checking the heatmaps that localize those concepts in the image.","Retrieval-grounded report generation should reduce invented findings, because the writer synthesizes from retrieved clinical documents rather than free generation.","Editing a concept score provides a concrete intervention: the user can correct a mistaken or missing concept and the classification and report should respond.","The conversational chat interface turns the system into a context-aware assistant that can answer follow-up questions about the image, the report, or the clinical condition."],"supporting_citations":[{"why":"Introduces concept bottleneck models, the architecture that forces predictions through interpretable concepts.","marker":"[7]"},{"why":"Supplies CheXagent, the vision-language model used for chest X-ray image embeddings.","marker":"[2]"},{"why":"Supplies the Mistral embedding model used to embed the clinical concept set for cosine similarity.","marker":"[5]"},{"why":"Provides the COVID-QU dataset of 33,920 chest X-rays with Pneumonia, COVID-19, and Normal labels used to train the classifier.","marker":"[3]"},{"why":"Introduces retrieval-augmented generation, the mechanism for grounding reports in external clinical knowledge.","marker":"[8]"},{"why":"Defines ReAct agents, the reasoning-and-acting pattern adopted by the per-disease report agents.","marker":"[17]"},{"why":"Provides label-free concept bottleneck methods that let the concept set be acquired automatically from an LLM.","marker":"[11]"},{"why":"The authors' earlier paper proposing concept bottlenecks with a multi-agentic RAG, which this demonstration builds on.","marker":"[1]"}],"fun_headline_variants":["Radiology reports grounded in editable concepts and external evidence","Concept bottleneck plus retrieval agents for explainable radiology AI","Traceable radiology reports via concept bottlenecks and multi-agent RAG","Editable clinical concepts make AI radiology reports transparent","Multi-agent RAG meets concept bottlenecks to curb hallucinations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that cosine similarity between image embeddings from one model and text embeddings from a different, separately trained model measures genuine clinical relevance, so that the resulting concept scores, heatmaps, and classifications are meaningful rather than artifacts of mismatched embedding spaces.","fun_headline_variants_meta":{"raw":{"variants":["Radiology reports grounded in editable concepts and external evidence","Concept bottleneck plus retrieval agents for explainable radiology AI","Traceable radiology reports via concept bottlenecks and multi-agent RAG","Editable clinical concepts make AI radiology reports transparent","Multi-agent RAG meets concept bottlenecks to curb hallucinations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000339,"raw_usage":{"total_tokens":1819,"prompt_tokens":837,"completion_tokens":982,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":911}},"tokens_in":453,"tokens_out":982,"duration_ms":7683,"temperature":1.0,"reasoning_tokens":911,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:16:16.445014+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of chest X-rays with expert-annotated findings and compute the concept scores and heatmaps: if the heatmap for 'pulmonary consolidation' does not localize to the annotated consolidation, or if the cosine-similarity ranking between the image and concept texts changes substantially when the two embeddings come from a jointly trained multimodal model, the grounding claim collapses. A second check is intervention fidelity: if editing a concept score leaves the predicted disease unchanged for most images, the editable-interpretability claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces concept bottleneck models, the architecture that forces predictions through interpretable concepts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies CheXagent, the vision-language model used for chest X-ray image embeddings."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the COVID-QU dataset of 33,920 chest X-rays with Pneumonia, COVID-19, and Normal labels used to train the classifier."}],"review_version":1}