{"id":"ae102dbd-371d-4719-9083-0b1041a8b1e6","arxiv_id":"2608.12592","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ReCoGen outperforms six baselines in downstream utility on all sixteen settings across three clinical datasets by decoupling condition representation (per-modality masked autoencoders) from generation (flow matching).","lead":"A new two-step AI system, ReCoGen, first learns compact, missing-tolerant summaries of each available health signal, then generates a target signal such as blood pressure from those summaries. Across three clinical datasets it produces signals that let a downstream predictor perform as well as or better than it does on real signals, though the authors note this may partly reflect re-using the input information.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ReCoGen's 'informative surrogate' claim is not settled by the reported protocol: a real-trained probe can reward condition re-expression, and the missing train-on-synthetic/test-on-real direction leaves the headline untested.","rationale":"The paper is honest and well-constructed: it isolates the conditioning path with ablations, removes label leakage by de-leaking, uses external ICD labels that cannot be copied from a conditioning variable, and concedes the fidelity/re-expression conflation in Section 3.2. Those are real strengths. Nevertheless the central headline, 'informative surrogates,' is a claim about what the generated signal carries, and the reported protocol only measures how a real-trained probe reads the generated signal. The missing reverse direction, train on synthetic and evaluate on real, is the load-bearing gap. Table 2's static-only column (0.656 on MIMIC-IV sepsis) shows the conditioning set carries much of the label signal, so re-expression is a live alternative to target fidelity. The undefined 'CGM-enhanced' static features in Appendix A are a possible specific leak on AI-READI. The paper's own limitation statement does not resolve either point, and the 'no clean target' justification is not fully convincing because observed target steps exist in held-out windows. The reader's CONDITIONAL verdict is the right one: accept only if the transfer test is run or the abstract's surrogate claim is softened. My stress-test therefore leaves the verdict unchanged.","tokens_in":17198,"tokens_out":10309,"duration_ms":106359,"concrete_test":"Retrain the identical 1D-CNN probe on ReCoGen-generated training-set signals (real labels, de-leaked conditions) and evaluate on held-out real signals; do the same for the strongest baseline in each setting. If ReCoGen's synthetic-trained AUROC approaches the real-trained reference within roughly 0.03 on the ABP tasks, the surrogate claim survives; if it drops much more while the train-on-real/test-on-synthetic score stays high, the headline result is best explained by condition re-expression or probe artifact, and the abstract's 'informative surrogates' claim should be withdrawn.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the inference from Table 1's train-on-real, test-on-synthetic scores to the abstract's 'informative surrogates' for invasive targets. Section 2.1 defines the augmentation gap using h* trained on real signals; a small gap only says a real-trained probe reads the synthetic signal well. It does not say a predictor trained on synthetic signals would work on real signals, which is what 'surrogate' and the data-augmentation framing require. A generator can make the reported gap small by re-expressing the conditioning set inside the waveform: Table 2 shows the static conditions alone already yield AUROC 0.656 on MIMIC-IV sepsis, so encoding the 27-d static vector into the generated ABP could explain much of ReCoGen's margin without the waveform carrying real ABP physiology. On AI-READI the risk is sharper: Appendix A lists six numeric 'CGM-enhanced' static features but never defines them; if any are CGM-derived summaries (mean glucose, time-in-range, etc.), the target's own information is available as a condition and the CGM-only results are largely re-expression. The paper acknowledges this conflation in Section 3.2 but does not resolve it; the claim that no clean target exists is also incomplete, since held-out windows retain observed target steps on which paired fidelity could be measured. The direction that would test surrogate utility, train on synthetic and evaluate on real, is absent.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ReCoGen, a two-stage conditional physiological time-series generator. Stage I trains one masked autoencoder per conditioning modality to distill irregularly sampled, partially observed streams into missingness-tolerant token sequences; Stage II freezes these encoders and trains a flow-matching DiT-style generator that mixes the token sequences with static clinical features via cross-attention and AdaLN. The target signals are ABP on MIMIC-III/IV and CGM on AI-READI. Evaluation is by downstream utility: a 1D-CNN classifier trained on real signals is applied to each method's generated signals, with AUROC/AUPRC reported for sepsis, heart failure, mortality, and study_group classification. ReCoGen reports the best score on all sixteen (dataset, task, metric) settings versus six conditional baselines, and matches or exceeds the real-signal reference on thirteen of them. The paper includes ablations of the conditioning injection path, static-feature encoding, and Stage-I mask ratio, and a statistical significance appendix with Holm-corrected t-tests and a Wilcoxon signed-rank test.","tokens_in":17461,"tokens_out":5580,"duration_ms":54544,"significance":"If the central claim is taken as 'ReCoGen produces signals that carry label-relevant information recoverable by a real-trained probe,' the result is well supported and the design contribution is solid: the decoupling of condition representation from generation is clearly motivated and the ablations show it is the source of the gains. The paper is also exemplary in transparency, explicitly acknowledging the fidelity/re-expression conflation and the anchor-not-ceiling interpretation, and in shipping statistical rigor (Holm corrections, Wilcoxon tests, five-seed means). The weakness is that the abstract and conclusion state a stronger claim—'informative surrogates'—which implies the generated signal can replace the real one in downstream use. The reported protocol does not test that implication, and the paper's own caveats admit that condition re-expression can explain the scores. The missing train-on-synthetic/test-on-real direction is directly testable, so the gap is fixable within the manuscript's scope.","major_comments":[{"comment":"The evaluation protocol, a real-trained probe scored on synthetic signals, supports the claim that generated signals contain label-relevant information readable by a real-trained model, but it does not support the abstract's and conclusion's 'informative surrogates' claim. A surrogate for augmentation requires the converse direction: a predictor trained on synthetic data should transfer to real data, or at least a train-on-synthetic-plus-real/test-on-real experiment. The paper itself acknowledges in Section 3.2 that a high score can arise from re-expressing the conditions, and Table 2 shows that the static conditions alone already reach AUROC 0.656 on MIMIC-IV sepsis versus ReCoGen's 0.692, leaving room for most of the reported margin to be re-expression. As written, the headline claim is not settled by the evidence; either add the train-on-synthetic/test-on-real evaluation or restrict the claims of Section 5 and the abstract to what the protocol actually measures.","section":"Section 3.2 and Conclusion/Abstract"},{"comment":"The static descriptor on AI-READI is described as containing '6 numeric CGM-enhanced features,' but these features are never defined. If any of them are summaries derived from the CGM trace itself (mean glucose, time-in-range, glycemic variability, etc.), then the CGM-only rows of Table 1 are substantially explained by the generator re-expressing the target's own information through the static condition, and the study_group AUROC of 0.771 could be achievable without generating a novel CGM waveform. Please define these features explicitly in Appendix A, and if any are CGM-derived, remove them from the conditioning set or quantify how much of the AI-READI result is attributable to re-expression rather than to the synthesized signal.","section":"Appendix A (AI-READI static features)"},{"comment":"The paper justifies omitting all pointwise fidelity metrics by saying 'the setting offers no clean target,' but this is incomplete. The held-out test subjects do have observed target steps; these steps are simply not used as ground truth in the main protocol. On the union of observed target steps, paired fidelity measures (e.g., correlation, absolute error, or a distributional distance between generated and real windows conditioned on the same inputs) can be computed and would partially separate target fidelity from condition re-expression. Without any such metric, the 'necessary but not sufficient' caveat of Section 2.1 remains unquantified, and the thirteen-of-sixteen 'reaches or exceeds the real signal' result is left open to the re-expression interpretation that the authors themselves flag.","section":"Section 3.2 ('no clean target')"}],"minor_comments":[{"comment":"The statement 'every method faces the same probe' is contradicted by the supplementary protocol (Appendix F, item iv), which reports that on AI-READI ReCoGen was scored by its own five-seed probe instance rather than the baseline probe instance; the two probe draws agree within 0.005 AUROC, but the main-text sentence should be qualified to reflect that AI-READI comparisons are across probe draws.","section":"Section 3.1 (Fairness of the comparison)"},{"comment":"Bridge is described as drawing a 'same-label real target window as its prototype.' This means Bridge receives label information through the exemplar, which is a form of label leakage even though the categorical label is disabled elsewhere. The paper acknowledges Bridge 'receives strictly more,' but it should also state explicitly that this prototype gives Bridge access to the downstream label at generation time, and discuss the implication for the fairness of the comparison.","section":"Appendix C (Bridge prototype)"},{"comment":"The limitations in Appendix F—AUPRC not tested, probe seeds shared yet unpaired tests used, rounded statistics, AI-READI probe mismatch, and the untested Real-Valid* anchor—are commendably transparent, but several of them (especially the first two) should be summarized in the main text near Table 1 so that readers of the main paper are not misled about the strength of the significance claims.","section":"Appendix F (limitations list)"},{"comment":"The caption notes the ρ=0.3 point need not reproduce the main table exactly because the sweep is a single end-to-end re-run; it would help to state whether the error bars reflect probe-seed variation only or also training-seed variation, since the latter is not otherwise reported.","section":"Figure 5 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is thorough and unusually honest about its evaluation's limitations. The main risk is that the abstract and conclusion overstate the 'informative surrogate' claim relative to the train-on-real/test-on-synthetic protocol; the missing train-on-synthetic/test-on-real direction is the obvious fix. I also urge the editor to require clarification of the 'CGM-enhanced features' in Appendix A, as this bears directly on whether the AI-READI results are a genuine generation result or largely a re-expression artifact. The technical contribution and the benchmark ranking are credible if the claims are appropriately scoped."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the method is genuinely new: decoupling condition representation (per-modality masked autoencoders) from generation (flow matching) is a sensible idea, and the systematic comparison of conditioning injection paths is a real contribution. Second, the headline claim — that ReCoGen turns routine signals into 'informative surrogates' for invasive ones — is not settled by the experiments, and the paper half-admits this in Section 3.2 while the abstract does not.\n\nWhat the paper does well: the framework is clearly described, the ablations are unusually thorough, and the authors are candid about the evaluation's central limitation. Downstream utility conflates target fidelity with condition re-expression, and they say so. Table 2 even shows static conditions alone support much of the MIMIC-IV sepsis score, which is the right kind of honesty. The significance testing appendix is also careful, noting the AUPRC rows are untested and that the AI-READI probe comparisons cross probe draws.\n\nWhere it falls short: the evaluation protocol is train-on-real, test-on-synthetic. That shows a real-trained probe reads the synthetic signal well. It does not show that a predictor trained on synthetic signals would work on real ones, which is what 'surrogate' and the data-augmentation framing require. The missing direction — train on synthetic, test on real — is absent, and that is the load-bearing experiment. The stress-test note is right about Appendix A too: the six 'CGM-enhanced' static features are never defined. If they include CGM summaries (mean glucose, time-in-range), the CGM-only results are partly re-expression of the target itself. That needs to be fixed before the abstract can make its current claim. The 'no clean target' argument is also too quick: held-out windows retain observed target steps on which paired fidelity could be measured.\n\nThe AUPRC rows carry no significance tests, and ReCoGen was scored by its own probe instance on AI-READI; both are minor but should be cleaned up. No code or data released, so independent verification is limited.\n\nOverall, the methodological core is solid and the evaluation is honest about its limitations. The main problem is framing: the abstract overreaches relative to what the protocol establishes. This is a paper for researchers working on conditional time-series generation and clinical ML evaluation; the discussion of the fidelity/re-expression confound is worth reading even for people who don't use the method. I'd send it to a serious referee — it deserves careful review, and the missing train-on-synthetic experiment plus the CGM-feature clarification are addressable. A revision that fixes those would make the paper's claims match its evidence.","headline":"A genuinely new two-stage generator with an honest evaluation section, but the 'informative surrogates' claim outruns the protocol's train-on-real, test-on-synthetic design.","tokens_in":17994,"tokens_out":3804,"would_cite":true,"duration_ms":33065,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ReCoGen decouples condition encoding from generation to make synthetic physiological signals outrank six generators on all sixteen benchmark settings.","keywords":["conditional time-series generation","multimodal conditioning","irregular missingness","masked autoencoders","flow matching","physiological signals","continuous glucose monitoring","arterial blood pressure"],"falsifier":"Take a held-out cohort where the real target waveform was recorded but withheld from the generator; compare generated and real waveforms pointwise (for example by correlation or mean absolute error) while also running the downstream probe. If downstream utility remains near the reported levels while pointwise fidelity is near chance, or if a downstream probe built from the conditioning covariates alone already matches ReCoGen's scores, then the utility numbers come from condition re-expression rather than target fidelity, and the surrogate-signal claim would collapse.","tokens_in":16977,"feed_emoji":"🩺","tokens_out":10522,"duration_ms":86799,"temperature":0.7,"pith_summary":"The paper sets out to show that conditional generation of a hard-to-measure physiological signal from co-recorded, irregularly missing multimodal signals fails when raw streams are fed straight into the generator, and that the fix is to give condition representation its own trained stage. ReCoGen does this by training one masked autoencoder per time-variant modality, forcing each to impute from context and output a compact token sequence, then freezing those encoders and training a flow-matching generator on top. On three clinical benchmarks the method beats six existing conditional generators on all sixteen dataset-task-metric settings, and on thirteen of those settings a classifier trained on real signals reads the generated signal as carrying at least as much label-relevant information as the real signal. The paper is explicit that this measures information transfer, not pointwise waveform fidelity, and treats the real-signal value as an approximate anchor rather than a ceiling. If the claim holds, routinely collected signals can stand in for invasive or unavailable ones without losing clinically relevant structure.","feed_headline":"Two-stage generator wins all 16 clinical signal benchmarks","feed_subtitle":"Synthetic ABP and CGM traces carry label-relevant information at or above real-signal levels in most settings.","key_machinery":"The load-bearing mechanism is the conditioning path rather than the generative backbone. Each Stage-I encoder turns an irregular, partially observed stream into a latent sequence; a bank of learnable per-modality query tokens then cross-attends over that latent to produce a compact set of in-context condition tokens, and when an entire modality is missing a learnable missing-token sequence stands in. Static clinical descriptors are encoded into one or two tokens and injected twice, once as an in-context token and once as adaptive layer-norm (AdaLN) modulation, which scales and shifts normalized activations per condition, an arrangement the ablations show to be complementary. The target series is handled through an invertible delay embedding that maps a one-dimensional waveform to a square image, so the flow-matching denoiser, a vision-transformer backbone, can generate the signal as an image and read it back as a waveform; the training loss is applied only on valid non-padded pixels, and inference integrates the flow ODE from noise to data. This pipeline is what allows the generator to consume robust condition tokens instead of learning to encode raw streams and generate at the same time.","core_discovery":"ReCoGen's central claim is that decoupling multimodal condition representation from target generation solves the problem of conditional physiological time-series generation under irregular missingness. In Stage I, each time-variant conditioning modality is distilled by its own masked autoencoder, trained to reconstruct held-out observed steps from context, so the resulting token sequences tolerate sparse and irregular sampling. In Stage II, a flow-matching vision transformer generates the target series from frozen per-modality tokens plus static descriptors, with learnable cross-attention queries summarizing each modality and static features injected both as in-context tokens and as adaptive layer-norm modulation. The paper reports that this design achieves the best downstream utility on all sixteen evaluated settings across AI-READI, MIMIC-III and MIMIC-IV, beating six representative conditional generators and reaching or exceeding the utility measured on real signals in thirteen of those settings.","pith_inferences":["A direct stress test the paper does not run: encode the conditioning covariates directly into the probe input (no generated waveform) and measure the same downstream AUROC. If that already recovers most of ReCoGen's reported utility, the scores are condition re-expression rather than target fidelity, the ambiguity the paper itself flags.","The decoupling recipe is domain-agnostic: any setting with heterogeneous, irregularly missing co-observed context and a target that is expensive to measure, such as environmental sensor networks or financial tick streams, could use the same represent-then-generate split.","Because Stage-I encoders are frozen and per-modality, a deployment-time extension would be to add a new condition modality by training only its masked autoencoder and attaching a query-token bank, leaving the generator untouched; the paper does not test incremental-modality addition.","The thirteen-of-sixteen at-or-above-real results suggest generated signals could serve as training augmentation for downstream classifiers, but only for labels that are not leaked into the conditioning set; the paper already removes the outcome label from all generators, so this is a natural continuation."],"forward_implications":["Synthetic ABP waveforms generated from heart rate, respiratory rate, and SpO2 plus static clinical data can be read by a real-trained classifier as carrying sepsis, heart-failure, and mortality-relevant information at or above the level of real ABP on most tested settings.","Generated CGM traces from wearable modalities and static features preserve enough glycemic structure to reproduce the ordinal study-group separation seen in real signals, supporting cohort analysis when CGM was not collected.","The ablations attribute most of the gain to the conditioning path, so the design principle of training missingness-aware encoders separately and then freezing them is what future conditional generators would need to retain to match this result.","Because the pipeline emits a regular 24-hour, 5-minute target from irregularly sampled conditions, it directly addresses the sampling patterns of ICU charting and wearable sensors rather than assuming densely aligned covariates."],"supporting_citations":[{"why":"Provides the masked-autoencoding reconstruction objective that defines Stage I condition encoders.","marker":"He et al., 2022"},{"why":"Supplies the invertible delay embedding that turns 1-D time series into images, and is one of the strongest baselines and in-painting alternatives.","marker":"Naiman et al., 2024"},{"why":"Gives the flow-matching objective used to train the Stage II generator.","marker":"Lipman et al., 2023"},{"why":"Supplies the vision-transformer denoiser backbone the generator is built on.","marker":"Peebles & Xie, 2023"},{"why":"Provides the MIMIC-III dataset and waveform-matched numerics used for ABP generation.","marker":"Johnson et al., 2016"},{"why":"Provides the MIMIC-IV cohort and bedside charted vitals for the second ABP benchmark.","marker":"Johnson et al., 2023"},{"why":"Provides the AI-READI type-2-diabetes cohort used for CGM generation and study-group probing.","marker":"AI-READI Consortium, 2024"},{"why":"Defines the Diffusion-TS baseline, which also represents the in-painting conditioning alternative ReCoGen is compared against.","marker":"Yuan & Qiao, 2024"},{"why":"Defines the TimeWeaver baseline that admits heterogeneous covariates but assumes dense alignment without modality missingness.","marker":"Narasimhan et al., 2024"}],"fun_headline_variants":["ReCoGen beats 6 generators on all 16 clinical benchmarks","Two-stage conditioning yields real-signal utility in 13/16 settings","Decoupled token-based conditioning: synthetic signals rival real ones","From multimodal tokens to target series: ReCoGen wins 16/16"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a probe trained on real signals and scored on generated signals measures whether the synthetic waveform preserves label-relevant physiology; the paper itself notes that a high score could instead mean the generator re-expresses conditioning covariates in a form the probe reads more easily than the real waveform.","fun_headline_variants_meta":{"raw":{"variants":["ReCoGen beats 6 generators on all 16 clinical benchmarks","Two-stage conditioning yields real-signal utility in 13/16 settings","Decoupled token-based conditioning: synthetic signals rival real ones","From multimodal tokens to target series: ReCoGen wins 16/16"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000333,"raw_usage":{"total_tokens":1887,"prompt_tokens":1016,"completion_tokens":871,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":632,"completion_tokens_details":{"reasoning_tokens":795}},"tokens_in":632,"tokens_out":871,"duration_ms":7619,"temperature":1.0,"reasoning_tokens":795,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:04:23.901251+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out cohort where the real target waveform was recorded but withheld from the generator; compare generated and real waveforms pointwise (for example by correlation or mean absolute error) while also running the downstream probe. If downstream utility remains near the reported levels while pointwise fidelity is near chance, or if a downstream probe built from the conditioning covariates alone already matches ReCoGen's scores, then the utility numbers come from condition re-expression rather than target fidelity, and the surrogate-signal claim would collapse.","supporting_citations":[{"cited_title":"Nature Metabolism , volume=","cited_arxiv_id":null,"evidence_quote":"Provides the AI-READI type-2-diabetes cohort used for CGM generation and study-group probing."}],"review_version":1}