{"id":"951ec95f-9603-446b-9765-bb212d68f991","arxiv_id":"2608.12904","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A single 3B transformer jointly trained on CT understanding and CT completion stays competitive with specialized models on radiology QA while also performing denoising, virtual contrast enhancement, and text-and-mask-to-CT generation.","lead":"HounsWorld is a 3B model that treats CT scans and clinical language as observations of one hidden patient state, and HounsBench is a corresponding benchmark for readout, reconstruction, and simulation. A smart generalist should care because it tests whether a single medical AI system can both understand and generate CT data, which could reduce the need for separate specialized radiology models.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The completion-to-readout transfer in Table 4 lacks a correspondence-shuffle control; without it, the +2.16 average gain could be generic multi-task regularization or replay of overlapping CT volumes, not evidence of a shared patient state.","rationale":"The reader's weakest assumption correctly points at the reliability of Eq. (4) pairings, but the sharper vulnerability is causal: even if every low-dose, contrast, text, and mask pair were perfectly patient-aligned, the paper provides no control that removes the alignment while keeping the loss family fixed. Such a control is required to attribute the Table 4 gains to a shared latent patient state rather than to generic multi-task regularization or replay of overlapping CT-RATE volumes. The absence of error bars on the ablation makes the +2.16 average gain (and especially the +0.94 difference between UND+VE and UND+CSPC) impossible to assess for significance. This does not overturn the paper's useful engineering contribution: the model does demonstrate competitive unified performance on a broad benchmark, and the authors are appropriately cautious in Section 5 about not claiming diagnostic interchangeability or long-term forecasting. But the central conceptual claim, that completion supervision improves readout because all tasks observe the same hidden patient state, is exactly the part that the current experiments do not isolate. A correspondence-shuffle control would settle whether the paired structure matters or whether any dense auxiliary loss would produce the same readout improvement. I therefore keep the reader's CONDITIONAL verdict: the paper should be accepted only with this control (or a released audit of pair correspondence and volume overlap) before the world-model interpretation is treated as established.","tokens_in":22820,"tokens_out":10760,"duration_ms":105414,"concrete_test":"Rerun the Table 4 CSPC ablation with a correspondence-shuffle control: train the identical UND+CSPC objective, 60/10/10/20 sampling weights, step count, and hyperparameters, but permute text+mask conditions across target CTs and swap low-dose/contrast pairs across patients so Eq. (4) no longer pairs observations of the same latent state. Evaluate on the fixed 34K HounsBench-Readout holdout with the same metrics. If the readout average stays within about 1 point of the true-CSPC row (62.02), the improvement is generic multi-task regularization or data replay, and the shared-patient-state interpretation is unsupported; if it drops materially, patient-aligned correspondence is causally load-bearing. Run at least three seeds for UND and both CSPC variants and report mean plus standard deviation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central evidence that CSPC learns a shared patient state is the completion-to-readout transfer in Table 4 and the partial-observation robustness in Fig. 3(a). The design compares UND against UND+DEN/VE/T2C/CSPC, but it never breaks the paired-condition assumption of Eq. (4) while keeping the loss family and data scale fixed. If the same text+mask-to-CT or denoising losses are trained with randomly permuted conditions (text/mask from one patient, target CT from another), a generic multi-task model can still improve readout through regularization, memorization of dataset statistics, or better low-level feature extraction. Section C.3 concedes that TotalSegmentator masks are automatic pseudo-labels and that text conditions are produced by an instruction model, so the authenticity of the 'same patient state' pairing is not independently auditable; the private cohort is unreleased. Moreover, HounsBench counts are materialized rows, not unique patients, so the 188K text+mask rows may be generated from the same CT-RATE volumes used in readout training; the +2.16 average readout gain could then reflect replay of the same volumes under an additional loss rather than cross-observation state sharing. Table 4 also reports no error bars or seeds; the +0.94 difference between UND+VE and UND+CSPC is within plausible run-to-run variance for a 3B model. Thus the paper's central causal claim, that clinically structured completion improves CT understanding through a shared patient state, is not yet separated from generic auxiliary-task benefit or data overlap.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes HounsWorld, a 3B multimodal world model for CT-centered clinical intelligence. It formulates readout, reconstruction, and simulation as prediction problems conditioned on an implicit hidden patient state, and trains a shared transformer with a joint objective combining language cross-entropy (Eq. 9) and flow-matching CT completion losses (Eq. 10) for low-dose denoising, virtual contrast enhancement, and text-plus-mask-to-CT generation. The authors also construct HounsBench, a CT benchmark with patient-disjoint splits across readout, reconstruction, and simulation, and report experiments showing competitive readout (second overall behind OmniCT in Table 3), strong generation metrics (Tables 1 and 2), and ablations (Table 4) suggesting that clinically structured completion improves readout. The paper includes detailed dataset construction, evaluation protocols, and explicit limitations in the appendix.","tokens_in":23215,"tokens_out":7018,"duration_ms":67709,"significance":"If the empirical claims hold, HounsWorld is a meaningful step toward unified patient-state modeling in medical imaging, demonstrating that a single 3B model can handle both understanding and generation across three task families. The completion-to-readout transfer result, if valid, would provide a practical training recipe for medical multimodal models, and the reward-model transfer to an independent Qwen3-VL policy is an interesting additional contribution. The paper is unusually transparent about data construction, automatic label provenance, and metric limitations, and it releases a project page. The main concern is that the central causal claim about the shared patient state rests on an ablation that needs a correspondence-shuffle control and seed-level statistics; these are fixable within a revision.","major_comments":[{"comment":"The completion-to-readout transfer experiment does not include a correspondence-shuffle control. The design holds the loss family and data scale fixed while adding DEN, VE, T2C, and CSPC, but it never breaks the patient-matched pairing in Eq. (4), for example by training the same completion losses with text or mask conditions randomly permuted across patients. Without such a control, the +2.16 average readout gain in Table 4 could be explained by generic multi-task regularization, memorization of dataset statistics, or replay of overlapping CT volumes rather than by learning an implicit shared patient state. Because the abstract and Section 5 present this transfer as evidence for the world-model formulation, this control is load-bearing and should be added.","section":"Section 4, 'Ablation Study and Model Analysis' (Table 4)"},{"comment":"No error bars, confidence intervals, or number of training seeds are reported for the readout benchmarks or the ablation. For a 3B model, run-to-run variance is likely nontrivial, and several reported differences are small: UND+VE (61.49) versus UND+CSPC (61.84), and the further +CO increment to 62.02. The claim that completion supervision 'consistently improves' CT understanding and that CSPC outperforms individual completions requires at least three seeds with standard deviations, or a paired significance test, for the metrics in Table 4 and the headline scores in Table 3.","section":"Tables 3-4 and Section 4"},{"comment":"The patient-level integrity of HounsBench is not auditable from the reported materialized counts. Appendix A states that counts are 'materialized records/targets, not unique patients,' and Figure 7 shows that text+mask-to-CT rows are generated from CT-RATE reports, while Table 5 shows CT-RATE is the dominant readout supervision source. It must be demonstrated that the same patient volumes do not appear in both the readout and completion training pools and that train/eval splits are patient-disjoint across all three families; otherwise the completion-to-readout transfer in Table 4 may reflect data overlap rather than a learned patient-state representation. Please provide patient-level counts and an explicit overlap audit.","section":"Appendix A, Figure 5, and Section C.3"}],"minor_comments":[{"comment":"Retention in the partial-observation experiment is normalized by each model's own full-observation score; if full-observation scores differ materially, the retention percentages are not directly comparable across models. Please report the absolute partial- and full-observation scores alongside the retention values.","section":"Figure 3(a)"},{"comment":"The GRPO+HW transfer experiment lacks details on the number of seeds, the evaluation split construction, and how the HounsWorld reward is computed for open-ended answers when the reward model is described as producing closed-form-only rewards. These details are necessary to interpret the open-ended improvement claim.","section":"Figure 3(b) and 'Transferable Reward Supervision'"},{"comment":"The open-ended composite Sopen uses fixed weights (0.1 BLEU, 0.1 ROUGE, 0.4 RadGraph-XL, 0.4 BioBERTScore); the paper should state explicitly how the score is defined when one component is unavailable, as in the organ-level analysis that uses only (BLEU+ROUGE)/2, and confirm that the weights are not renormalized in that case.","section":"Section D.1, Eq. (14)"}],"recommendation":"major_revision","confidential_remarks":"The paper is well structured and unusually detailed in its data audit, and I would be willing to review a revised version. My main reservations are methodological: the central causal claim needs a shuffling control and repeated-seed statistics, and the benchmark's patient-level disjointness needs a formal audit. I also note that the base model (Lance), the data protocol (OmniCT), and the benchmark construction come from overlapping groups, so external validation would considerably strengthen the impact of the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one. HounsWorld is a serious attempt to unify CT readout, reconstruction, and simulation in a single 3B transformer, and it mostly delivers as an engineering and systems contribution. The joint understanding-generation objective over 3D CT—flow-matching for generation, language CE for readout—is a genuine combination not present in CheXWorld or OmniCT. The HU-window-conditioned observation operator and branch-specific adapters are sensible. On their own benchmark it is competitive with specialized readout models (about a point behind OmniCT) and strong on generation metrics. The authors also write honest limitations: masks are TotalSegmentator pseudo-labels, generated CT is not presented as diagnostic, and radiologist studies are still needed. That puts it above the usual conference filler.\n\nNow the soft spots, and they matter. The load-bearing claim—that clinically structured completion improves readout through a shared patient state—rests on Table 4, which compares understanding-only against understanding plus denoising/contrast/text-mask objectives. There is no correspondence-shuffle control. If you train the same completion losses with conditions paired to wrong patients, you would see whether the gain is genuinely from same-patient state coherence or just from additional losses regularizing the representation. The stress-test note is right: without that control, the +2.16 average could be generic multi-task benefit, memorization of dataset statistics, or replay of the same CT-RATE volumes that already appear in readout training. The appendix makes the overlap risk concrete: 188K text+mask rows are materialized from 47.1K target CT volumes, and the readout set is built from the same source datasets using the OmniCT protocol. That doesn't make the result false, but it makes it unauditable as evidence for a world model. The private cohort adds to the problem, and no code or weights are released. The overlapping lineage with OmniCT and Lance is not disqualifying, but it raises the bar for independent audit.\n\nAlso, no error bars or repeated seeds anywhere in the key ablations. The +0.94 difference between UND+VE and UND+CSPC is within plausible run-to-run variance for a 3B model. This is a fixable weakness, not a fatal one, but it should be fixed before the claim is accepted.\n\nWho is this for? Researchers working on unified medical multimodal models and anyone thinking about benchmark design for task families. It deserves a serious referee—send it out—but the revision should require the shuffled-pairing control, repeated seeds, and either release of the benchmark or a detailed overlap audit. I'd use it as a related-work citation, not yet as evidence for patient-state modeling.","headline":"Serious unified CT modeling with a promising benchmark, but the completion-to-readout transfer needs a correspondence-shuffle control before it supports the world-model claim.","tokens_in":23710,"tokens_out":4063,"would_cite":true,"duration_ms":40074,"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":"The paper argues that CT scans and clinical text are views of one hidden patient state, and that a single 3B model should read, reconstruct, and simulate from it.","keywords":["world model","patient-state inference","computed tomography","multimodal understanding and generation","low-dose CT denoising","virtual contrast enhancement","text-to-CT generation","clinical VQA"],"falsifier":"Shuffle the patient correspondence in the paired CT items — replace each low-dose CT's routine-dose target with another patient's scan, and each non-contrast CT's arterial target with another patient's phase — and retrain; if the completion-to-readout gain over understanding-only training persists, the transfer is not evidence for a shared patient state.","tokens_in":22651,"feed_emoji":"🩻","tokens_out":7188,"duration_ms":62435,"temperature":0.7,"pith_summary":"The paper argues that CT-centered clinical intelligence should be posed as inference over a shared hidden patient state, so that reading out an answer, reconstructing a report or a cleaned scan, and simulating a future acquisition are all the same state-dependent prediction operation. If that is right, one model can replace task-specific systems for VQA, report generation, low-dose denoising, virtual contrast enhancement, and text-to-CT generation, and completion objectives become a legitimate training signal for medical understanding. To make the claim testable, the paper builds HounsBench, a patient-disjoint benchmark that organizes these tasks into readout, reconstruction, and simulation families, and HounsWorld, a 3-billion-parameter multimodal model trained with joint understanding-generation learning. The reported evidence is that HounsWorld stays competitive on readout, leads generation baselines on CT completion, and that clinically structured completion supervision consistently improves readout accuracy and robustness to partial scans.","feed_headline":"One 3B model reads, denoises, and enhances CT from one patient state","feed_subtitle":"HounsWorld treats CT and clinical text as views of the same hidden state, and completion tasks sharpen its readout.","key_machinery":"The key mechanism is Clinically Structured Patient-State Completion (CSPC), a training objective whose items pair source observations with patient-matched target observations under a changed condition — different dose, different contrast phase, or a text-plus-mask description of the target anatomy. Each item forces the shared causal transformer to maintain a latent state that explains more than one view of the same patient. Supporting mechanisms include zero-initialized residual adapters that leave pretrained multimodal and generative interfaces intact while adding CT-specific corrections, condition-explicit Hounsfield-unit window sampling, pseudo-frame construction that packs adjacent axial slices into RGB channels so volumetric CT can enter pretrained vision and VAE encoders, and flow matching in VAE latent space for CT generation.","core_discovery":"The central claim is that a single shared transformer can estimate an implicit patient state from volumetric CT, language, anatomy masks, and condition descriptors, and then use that state for three output families: query-conditioned answers (readout), reports and captions (language reconstruction), and condition-specific CT volumes (low-dose denoising, virtual contrast enhancement, and text-plus-mask-to-CT generation). HounsWorld, a 3B model trained with the joint understanding-generation objective, is reported to rank first among generation baselines on completion metrics, to match or approach specialized readout models within roughly one point, and to improve readout on every evaluated category when structured completion is added to understanding-only training. The paper interprets this as evidence that completion and readout are not separate capabilities but complementary observations of the same patient state.","pith_inferences":["If the shared-state claim holds, paired routine clinical acquisitions (low-dose plus routine-dose, non-contrast plus contrast phases) could serve as a large-scale self-supervision signal for medical VLMs in settings where radiologist annotations are scarce.","Because the anatomy masks are automatic pseudo-labels rather than radiologist ground truth, a direct test of the mechanism is to rebuild text+mask conditioning with verified masks and see whether the completion-to-readout gain grows or vanishes; that would separate mask fidelity from genuine state learning.","The same framework should extend to other observation channels of a patient state, such as MRI sequences, PET, or longitudinal follow-up, with the same CSPC objective operating across modalities.","A practical consequence the authors leave implicit is counterfactual interrogation: controlling dose, phase, or anatomy lets a clinician ask 'what would this scan look like under condition X?' for the same patient, which the qualitative examples illustrate but the paper does not validate diagnostically."],"forward_implications":["A single HounsWorld-style model can serve readout, reconstruction, and simulation without per-task architectures, so hospitals could run one checkpoint instead of separate VQA, report, denoising, and synthesis systems.","Structured completion objectives — denoising, phase transfer, and anatomy-conditioned generation — become a principled source of supervision for CT understanding, not just auxiliary outputs.","Readout degrades more gracefully under sparse or partial CT coverage when completion supervision is present, which matters for low-dose or limited-view acquisition protocols.","The learned patient-state representation transfers across architectures: a frozen HounsWorld can act as a reward model to improve an independent vision-language model on clinical QA.","The scope is bounded to condition-specified, short-horizon prediction (for example, non-contrast to arterial/venous phase), not unrestricted longitudinal disease forecasting."],"supporting_citations":[{"why":"CT-RATE supplies the chest CT–report data and CT-CHAT baseline that anchor HounsBench-Readout.","marker":"Hamamci et al. (2026)"},{"why":"M3D supplies whole-body CT VQA and caption records plus the M3D-LaMed baseline.","marker":"Bai et al. (2024)"},{"why":"OmniCT's instruction-construction protocol and readout scores define HounsBench's readout data recipe and the strongest specialized comparison.","marker":"Lin et al. (2026)"},{"why":"TotalSegmentator produces the 117-label automatic anatomy masks used as spatial conditions in text+mask-to-CT.","marker":"Wasserthal et al. (2023)"},{"why":"Flow matching provides the generative objective used to complete CT volumes in VAE latent space.","marker":"Lipman et al. (2023)"},{"why":"Lance checkpoint is the pretrained unified model from which HounsWorld's stage-2 training initializes.","marker":"Fu et al. (2026)"},{"why":"World models supply the latent-state abstraction that the patient-state formulation adapts to medicine.","marker":"Ha and Schmidhuber (2018)"},{"why":"Masked autoencoding is cited as evidence that reconstructive objectives learn transferable representations, motivating completion-to-readout transfer.","marker":"He et al. (2022)"}],"fun_headline_variants":["A single 3B model reads, denoises, and enhances CT from one hidden state","CT readout, reconstruction, and simulation from one shared state","HounsWorld: one model for CT answers, reports, and volume generation","One transformer turns CT scans and text into answers and images"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the paired-condition training items are genuine observations of the same hidden patient state — same-patient low-dose and routine-dose CT, same-voxel non-contrast and arterial/venous phases, and text-plus-mask pairs that faithfully describe the target CT — since the masks are automatic pseudo-labels and the private cohort is not externally audited.","fun_headline_variants_meta":{"raw":{"variants":["A single 3B model reads, denoises, and enhances CT from one hidden state","CT readout, reconstruction, and simulation from one shared state","HounsWorld: one model for CT answers, reports, and volume generation","One transformer turns CT scans and text into answers and images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000356,"raw_usage":{"total_tokens":1943,"prompt_tokens":969,"completion_tokens":974,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":895}},"tokens_in":585,"tokens_out":974,"duration_ms":8719,"temperature":1.0,"reasoning_tokens":895,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:54:24.727429+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Shuffle the patient correspondence in the paired CT items — replace each low-dose CT's routine-dose target with another patient's scan, and each non-contrast CT's arterial target with another patient's phase — and retrain; if the completion-to-readout gain over understanding-only training persists, the transfer is not evidence for a shared patient state.","supporting_citations":[],"review_version":1}