{"id":"7176f77e-dfe4-4f13-a5e6-2b4873833ef8","arxiv_id":"2511.16625","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A retraining-free uncertainty layer is claimed to reduce overconfident clinical QA errors, but the key derivation is invalid and the abstract and full text report different results.","lead":"This preprint proposes MedBayes-Lite, a no-retraining layer that combines Monte Carlo dropout, confidence-based abstention, and an uncertainty-weighted attention penalty for clinical question answering. The paper's headline claims of large calibration improvements are undermined by an invalid theoretical centerpiece, missing code and data, and an abstract that describes different datasets and results than the full text.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5 is not a valid variance decomposition: the layer-wise sum double-counts downstream variance, so the paper's central theoretical contribution collapses.","rationale":"The reader's weakest-assumption analysis and my own agree: Theorem 5 is load-bearing and does not hold as stated. The 'first closed-form hierarchical variance decomposition' is the paper's stated novelty (§1, §3.3.2) and is the only thing distinguishing MedBayes-Lite from MC dropout + attention reweighting + entropy abstention. A direct application of the law of total variance along the Markov chain h(0)→...→h(L)→ŷ gives a decomposition involving E[Var(ŷ|h(L))] and E[Var(E[ŷ|h(l)]|h(l−1))]; Theorem 5 instead uses Var_{e(l)}(ŷ|h(l−1)), which includes downstream variance. The simple additive two-layer counterexample shows the sum can be twice the true variance. The Gaussian perturbation assumption also contradicts the Bernoulli dropout implementation, and Algorithm 1's third term E[Var(ŷ|α)] is not in the theorem. Since no formal verification or code is provided, there is no independent support for the identity. This is an internal inconsistency, not merely a disagreement with the current consensus. The empirical sections cannot rescue the core claim; additionally, Tables 6–9 report CUS/ZTI, which are defined in terms of confidence thresholds and are tuned on evaluation data, so the safety claims are partly circular. I therefore see no reason to change the reader's REJECT verdict.","tokens_in":19428,"tokens_out":6408,"duration_ms":52267,"concrete_test":"Independently re-derive Theorem 5 via iterative conditioning on h(1),...,h(L) and verify numerically on a two-layer stochastic network with additive Gaussian or dropout perturbations (e.g., h1 = W1 x + e1; y = W2 h1 + e2). Estimate true Var[y] by Monte Carlo over (e1,e2); compute the theorem's RHS by Monte Carlo conditioned on h0 and h1. If Theorem 5 is valid, RHS equals true variance within sampling error; I predict RHS ≈ 2·Var[y], exposing the double-counting. The same test on the corrected recursive formula E[Var(y|h(2))] + E[Var(E[y|h(2)]|h(1))] + Var(E[y|h(1)]|h(0)) should match.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central contribution is Theorem 5 in §3.3.2, called the 'theoretical backbone.' The claimed identity Var[ŷ] = Σ_l E_{h(l−1)}[Var_{e(l)}(ŷ|h(l−1))] + Var_{h(l−1)}[E_{e(l)}(ŷ|h(l−1))] does not follow from iterating the law of total variance. Correct recursive conditioning on the transformer states gives Var[ŷ] = E[Var(ŷ|h(L))] + Σ_l E[Var(E[ŷ|h(l)] | h(l−1))]. In the theorem, Var_{e(l)}(ŷ|h(l−1)) is the full variance of the final prediction with respect to perturbation at layer l, including all downstream layers h(l+1),...,h(L); summing this over l counts the same downstream variance multiple times. And Var_{h(l−1)}[E_{e(l)}(ŷ|h(l−1))] is not the layer contribution Var(E[ŷ|h(l)]|h(l−1)) unless downstream layers are deterministic. A minimal counterexample: two independent additive Gaussian layers, h1=e1, y=h1+e2; the true variance is σ1²+σ2², while Theorem 5 gives 2(σ1²+σ2²). Moreover, the Gaussian perturbation e(l)∼N(0,σ_l² I) is inconsistent with the actual MC dropout (Bernoulli masks), and Algorithm 1 adds a third term E[Var(ŷ|α)] that the theorem does not contain. If Theorem 5 fails, the claimed 'first layer-wise Bayesian variance decomposition' is unsupported; the framework reduces to standard MC dropout plus a heuristic attention penalty and entropy abstention.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MedBayes-Lite, a retraining-free inference-time uncertainty layer for transformer-based clinical language models that combines MC dropout, uncertainty-weighted attention (softmax reweighting by exp(-λU(x_j))), and entropy-based confidence abstention. It claims a 'first closed-form hierarchical variance decomposition for transformer architectures' (Theorem 5), two new clinically motivated metrics (CUS, ZTI), and large empirical gains: 32–48% overconfidence reduction, ECE reductions of 0.23–0.33, and prevention of up to 41% of diagnostic errors on biomedical QA and clinical prediction benchmarks (MedQA, PubMedQA, MIMIC-III in the body; MedMCQA and MedQA-USMLE in the abstract). The framework is evaluated against post-hoc calibration, deep ensembles, and SWAG, with additional cross-dataset and prompt-sensitivity experiments.","tokens_in":19920,"tokens_out":2547,"duration_ms":24263,"significance":"If the central claims were correct, a lightweight, retraining-free uncertainty layer for clinical LLMs would be practically valuable, and the layer-wise variance decomposition would be a meaningful theoretical contribution. The paper also has strengths in scope: it targets an important safety problem, proposes explicit abstention behavior, and includes a computational profiling table. However, the theoretical backbone (Theorem 5) is not a valid variance decomposition, and the empirical reporting is internally inconsistent across the abstract, the main text, and the tables. The new CUS/ZTI metrics are defined in the paper but are not validated against any external ground truth, and their interpretation shifts between sections. These issues are load-bearing: the headline contribution and headline numbers are not established.","major_comments":[{"comment":"Theorem 5, described as the 'theoretical backbone,' is not a valid layer-wise decomposition of predictive variance. Iterating the law of total variance requires conditioning on the full history of latent states. The expression Σ_l E[Var_e(l)(ŷ|h^(l-1))] + Var[E_e(l)(ŷ|h^(l-1))] is not the recursive decomposition Var[ŷ] = E[Var(ŷ|h^(L))] + Σ_l E[Var(E[ŷ|h^(l)] | h^(l-1))]. As written, each term Var_e(l)(ŷ|h^(l-1)) is the full variance of the final prediction with respect to perturbation at layer l, including all downstream stochastic layers, so the sum double-counts downstream variance. A minimal counterexample with two independent additive Gaussian layers gives the true variance σ1²+σ2² while the theorem yields 2(σ1²+σ2²). Additionally, e(l)~N(0, σ_l² I) is inconsistent with the Bernoulli MC dropout masks used in Algorithm 1, and Algorithm 1 introduces a third term E[Var(ŷ|α)] that is ab","section":"§3.3.2, Theorem 5 and Algorithm 1"},{"comment":"The abstract reports evaluation on MedMCQA and MedQA-USMLE with ECE reductions of 0.23–0.33 and a reduction of confident high-severity errors from ~21% to near zero. The full text evaluates PubMedQA, MedQA, and MIMIC-III, and no table reports the headline 'confident high-severity errors' or the specific ECE reductions in the abstract. The CUS/ZTI tables report different quantities, and the ECE improvements shown (e.g., Table 6: 0.182→0.113; Table 8: 0.0698→0.0456) are far smaller than 0.23–0.33. This discrepancy prevents verification of the paper's central empirical claims.","section":"Abstract vs. §4–§5 and Tables 6–11"},{"comment":"Section 3.4 states that with M=10 samples MedBayes-Lite incurs 'under a 10% increase in inference latency,' but Table 11 reports baseline ClinicalBERT latency 32.04 ms and MedBayes-Lite (MC=10) latency 60.13 ms, an ~88% increase. Even MC=5 gives 32.61 ms, only 1.8% overhead, while the later ablation recommends M=10–20. The stated '<10% increase' claim is contradicted by the paper's own efficiency table, and this affects the practical feasibility argument for clinical deployment.","section":"§3.4 vs. Table 11"},{"comment":"The two headline safety metrics CUS and ZTI are introduced in this paper and are not validated against any external or independently established measure of clinical harm. The severity weights w_clinical(b) in CUS are never specified, and the abstract's claim that CUS 'strongly correlates with harmful overconfidence (r≈0.88)' is not substantiated by any table or analysis in the main text. Moreover, the ablation study in §5.3 describes increasing CUS (with dropout) as 'intentional conservatism,' but CUS is defined as a lower-is-better risk-aware calibration error; the interpretation is thus inconsistent. These issues undermine the safety-oriented conclusions that depend on CUS and ZTI.","section":"§4.2.2, §5.3, and abstract"}],"minor_comments":[{"comment":"There are multiple typographical errors and inconsistent labels: 'SW AG' appears in §4.4 and Table 11, 'V AE' in §3.2.1, and 'MMedBayes-Lite' in §5.4.2. These should be corrected.","section":"Throughout"},{"comment":"For generative models, temperature is defined as np.random.uniform(0.7, 1.2), which is a random draw rather than a fixed configuration. This makes the experimental setup ambiguous and hampers reproducibility.","section":"§4.3, Table 5"},{"comment":"The appendix reports total runtime as 'approximately XX GPU hours,' which is a placeholder. This should either be filled with the actual number or removed.","section":"Appendix A.1"},{"comment":"References [14] and [31] are duplicated (both are Guo et al., 'On calibration of modern neural networks'). The reference list should be deduplicated.","section":"References"},{"comment":"The discussion of 'reject-option classifiers and selective prediction frameworks' cites Geifman and El-Yaniv [22] and Jardine et al. [23] but does not connect them to the specific implementation of entropy-based abstention, which would help position the contribution.","section":"§2.3"}],"recommendation":"reject","confidential_remarks":"The central theoretical result is invalid, and the abstract and body report inconsistent datasets, metrics, and efficiency numbers. These are not local presentation issues; they affect the paper's main claims. My recommendation is reject. I would note that the paper's topic is timely and the proposed combination of MC dropout, attention reweighting, and abstention is not without practical interest, but as submitted the manuscript does not provide a sound basis for those claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper's central theoretical contribution doesn't hold up, and the experimental reporting is messy enough that I wouldn't trust the headline numbers. You'd be right to reject it as is.\n\nWhat's actually here: the problem is real—clinical LLMs are overconfident, and a lightweight inference-time layer combining MC dropout, uncertainty-weighted attention, and entropy-based abstention is a sensible engineering idea. The paper is also right that post-hoc calibration alone doesn't fix overconfidence under shift. The CUS and ZTI metrics are a reasonable attempt to make calibration numbers clinically meaningful, even if they're new and self-defined.\n\nThe soft spots are large. Theorem 5, presented as the 'theoretical backbone,' is not a valid variance decomposition. Applying the law of total variance conditionally on each layer's input without accounting for downstream layers double-counts the same variance. The stress-test counterexample is correct: for two independent additive Gaussian layers, the theorem gives twice the true variance. The proof also assumes Gaussian perturbations e(l), which contradicts the actual MC dropout implementation using Bernoulli masks. And Algorithm 1 adds a third term for attention-level uncertainty that the theorem doesn't include. So the 'first layer-wise Bayesian variance decomposition' collapses.\n\nThe empirical side has its own problems. The abstract reports numbers on MedMCQA and MedQA-USMLE; the full text uses MedQA, PubMedQA, and MIMIC-III. ECE reductions of 0.23–0.33 in the abstract are not directly tied to the tables I see. Table 11 shows MedBayes-Lite at MC=10 taking 60.13 ms vs 32.04 ms baseline—that's an 87% latency increase, not 'under 10%.' There are no error bars anywhere, and key hyperparameters (λ, τ, dropout p, MC count) appear to be tuned on the same data used for evaluation. The CUS/ZTI safety claims are validated only against the authors' own severity weights, so the 'near-zero harmful overconfidence' language is partly circular.\n\nWhat's worth keeping: the framing of clinical risk-aware UQ is good, and the three-component pipeline is a reasonable baseline for future work. But the current evidence doesn't support the headline claims.\n\nWho should read it: someone working on clinical UQ might skim it for the metric ideas, but the theory isn't usable and the numbers aren't reproducible. I wouldn't cite it in my own work. If I were an editor, I'd desk reject this version—the internal contradictions are serious enough that referee time would be spent documenting flaws rather than evaluating a contribution.","headline":"The paper's central theoretical claim collapses under the law of total variance, and the experimental reporting has enough internal contradictions that the headline safety numbers can't be trusted.","tokens_in":20400,"tokens_out":3079,"would_cite":false,"duration_ms":29159,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MedBayes-Lite claims a retraining-free uncertainty layer that cuts clinical LLM overconfidence by 32–48% and flags uncertain cases for human review.","keywords":["clinical uncertainty quantification","Monte Carlo dropout","medical question answering","calibration","abstention","overconfidence","layer-wise variance decomposition","transformer"],"falsifier":"Run many stochastic forward passes on a fixed medical QA input and compute the empirical total variance of the prediction. Separately, using the same passes, compute each layer's conditional variance and conditional-mean variance and sum them as in Theorem 5. If the two numbers disagree beyond sampling error, the decomposition is not exact; a similar check should replace the theorem's Gaussian noise with the actual Bernoulli dropout masks used in Algorithm 1.","tokens_in":19316,"feed_emoji":"🩺","tokens_out":6673,"duration_ms":58178,"temperature":0.7,"pith_summary":"The paper sets out to show that a clinical language model can be equipped with uncertainty awareness after training, by adding a lightweight governance layer rather than retraining. The layer runs stochastic forward passes with dropout, downweights unreliable tokens in attention, and abstains when predictive entropy is high. If the empirical claims hold, a drop-in wrapper would cut overconfident wrong answers in medical QA—especially high-severity cases—while adding under 3% parameters and less than 10% inference latency. The paper's theoretical centerpiece is a claimed layer-wise variance decomposition for transformers, which is supposed to make uncertainty propagation interpretable and principled. The authors flag limits in §6.3 (extreme distribution shift, all-noisy inputs, excessive dropout) and note in the abstract that the method does not improve risk-coverage ranking.","feed_headline":"Uncertainty layer cuts clinical AI overconfidence by up to 48%","feed_subtitle":"Medical QA wrapper defers uncertain predictions and nearly eliminates confident high-severity errors.","key_machinery":"Three coupled mechanisms carry the argument: (1) Bayesian Embedding Calibration—M MC-dropout forward passes produce an embedding mean and variance, treating dropout as a variational posterior; (2) Uncertainty-Weighted Attention—each attention weight α_ij is multiplied by exp(-λU(x_j)) and renormalized so high-variance tokens contribute less; (3) Confidence-Guided Decision Shaping—entropy-based confidence C(p)=1-H(p)/logK gates predictions, deferring anything below threshold τ. The theorem that is meant to unify them is Theorem 5's layer-wise variance decomposition, which is offered as the 'theoretical backbone' and the basis for interpretable, layer-level uncertainty attribution.","core_discovery":"MedBayes-Lite's central claim is that uncertainty can be made a first-class part of transformer inference without retraining: MC dropout converts each token embedding into a distribution, attention weights are renormalized with an uncertainty penalty exp(-λU(x_j)), and predictions are accepted only when normalized confidence C(p)=1-H(p)/logK clears a threshold τ. The paper further claims Theorem 5, the first closed-form hierarchical variance decomposition for transformer architectures, decomposing Var[ŷ] into a sum over layers of per-layer aleatoric and epistemic terms, and thus offering token- and layer-level uncertainty maps. On MedMCQA and MedQA-USMLE, the paper reports ECE reductions of","pith_inferences":["If Theorem 5's identity survives numerical checking, the same layer-wise variance tracing could be reused as a model-auditing tool independent of MedBayes-Lite—for instance, to locate which layers inject the most uncertainty in any transformer.","The most conservative reading is that the empirical gains might persist even if the theorem fails, because MC dropout plus entropy-thresholded abstention are known to improve calibration on their own; the theorem's role would then be explanatory rather than load-bearing.","A testable extension is to combine the layer with temperature scaling: the paper reports no improvement in risk-coverage ranking, so a two-stage pipeline—Bayesian abstention for safety, temperature scaling for ranking—could recover both benefits.","The reported CUS correlation of 0.88 needs validation outside QA benchmarks, e.g., on real clinical notes with severity labels, before using CUS as a safety monitor in practice."],"forward_implications":["If MedBayes-Lite works as described, deploying it requires only inference-time changes: no retraining, no new trainable parameters, under 3% parameter overhead and roughly linear MC-sample latency.","Confident high-severity errors, the most dangerous failure mode, fall from about 21% to near zero when moving from MedMCQA to MedQA-USMLE; ECE drops by 0.23–0.33 and calibration drift roughly halves.","The Clinical Uncertainty Score (CUS) could serve as a risk-aware calibration audit metric, since the paper reports it is strongly correlated (r≈0.88) with harmful overconfidence.","The method reports consistent reliability gains across model families (BERT variants, GPT, LLaMA) and prompt styles, including chain-of-thought, which suggests it is model-agnostic rather than tied to a single architecture.","Because the framework abstains rather than guesses, it gives clinical workflows a concrete 'review required' signal, the computational analogue of deferring to human judgment."],"fun_headline_variants":["No-retrain uncertainty layer cuts clinical AI overconfidence","MedBayes-Lite: zero-train layer slashes harmful confident mistakes","Uncertainty governance: clinical AI defers risky calls","Calibration layer nukes overconfident medical AI errors","Dropout-based guard makes clinical AI know when to abstain"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing assumption is that layer-by-layer variance really adds up to the model's total uncertainty; the paper's own implementation adds an extra attention term and uses dropout masks where the proof uses Gaussian noise, so the identity is assumed rather than demonstrated.","fun_headline_variants_meta":{"raw":{"variants":["No-retrain uncertainty layer cuts clinical AI overconfidence","MedBayes-Lite: zero-train layer slashes harmful confident mistakes","Uncertainty governance: clinical AI defers risky calls","Calibration layer nukes overconfident medical AI errors","Dropout-based guard makes clinical AI know when to abstain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000143,"raw_usage":{"total_tokens":1012,"prompt_tokens":752,"completion_tokens":260,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":189}},"tokens_in":496,"tokens_out":260,"duration_ms":3539,"temperature":1.0,"reasoning_tokens":189,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T21:05:18.993736+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run many stochastic forward passes on a fixed medical QA input and compute the empirical total variance of the prediction. Separately, using the same passes, compute each layer's conditional variance and conditional-mean variance and sum them as in Theorem 5. If the two numbers disagree beyond sampling error, the decomposition is not exact; a similar check should replace the theorem's Gaussian noise with the actual Bernoulli dropout masks used in Algorithm 1.","supporting_citations":[],"review_version":1}