{"id":"f79a604b-a4e6-4b42-895a-b0cdf8757c87","arxiv_id":"2506.03158","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":9,"one_line_summary":"DUAL adds dynamic uncertainty modeling, adaptive loss modulation, and uncertainty-weighted cross-modal fusion to existing deep learning backbones, reporting accuracy improvements on CIFAR, sentiment analysis, and EEG benchmarks.","lead":"The paper proposes DUAL, a framework for treating feature uncertainty as a time-varying quantity in both image and multi-modal classification, and reports accuracy gains on six benchmarks. A generalist reader might care because robust handling of uncertain features is a core obstacle in deploying deep learning.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central empirical claim is unsupported because no training protocol, hyperparameter values, or code are reported, so baseline comparability and the source of the reported gains cannot be verified.","rationale":"The reader's weakest_assumption identifies exactly the same condition that is load-bearing: the reported gains depend on baselines trained to a comparable standard under the same protocol, and on hyperparameters not selected by test-set overfitting. I agree with the REJECT verdict because the manuscript provides no experimental contract—no epochs, no learning rate, no augmentation, no seeds, no hyperparameter values, and no code—that would let a reader know what was actually compared. The consistency and size of the gains (e.g., 7.1% on CIFAR-10, 6.2% Acc-7 on CMU-MOSEI) make the concern concrete rather than procedural: those deltas are far larger than typical seed noise for well-trained baselines, and the absence of error bars means the improvements could be driven by weak baselines or selective reporting. The appendix's theoretical analysis does not substantiate the empirical claim; its main theorem is essentially a restatement of the parameter update. A controlled rerun with released code and fixed protocol is the one check that would settle whether the central claim holds, and until that is available the paper as written does not meet the bar for acceptance.","tokens_in":13464,"tokens_out":3825,"duration_ms":33561,"concrete_test":"Request the exact training configuration and code for §5.1, then rerun the EfficientNet-B0 CIFAR-10 comparison under a fixed, publicly standard protocol (e.g., 200 epochs, SGD, standard augmentation, 3 seeds) with and without DUAL-S, reporting mean ± std. If the DUAL-S gain is not positive and greater than 2× the pooled seed standard deviation, the central improvement claim is not established. The same check should be applied to TFN on CMU-MOSEI, where the reported 6.2% Acc-7 gain must exceed seed noise under the original TFN training setup.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that DUAL adds consistent, material accuracy gains to existing backbones. The only evidence for this is Tables 1–5, which present 'baseline + improvement' numbers but never state the training protocol: epochs, batch size, optimizer, learning rate, data augmentation, seeds, split assignment, or the values of the many hyperparameters introduced in Eqs. 8, 14, 16, 17, 19–20, 23–27 (α, β, γ, λ, λKL, α0, τ, R, η0, etc.). Without these details, the baseline numbers could be weak or undertuned, or the hyperparameters could be selected to maximize the reported gains. The magnitude of some reported deltas makes this concern concrete rather than hypothetical: EfficientNet-B0 on CIFAR-10 improves 7.1% to reach only 82.0%, and TFN improves 6.2% in Acc-7 on CMU-MOSEI, both far outside typical seed noise if the baselines were trained to a standard protocol. The Appendix's theoretical analysis does not rescue the empirical claim because it is largely tautological (e.g., Theorem .1 restates the update rule as the result). Thus the central claim rests entirely on an undocumented comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes DUAL, a framework for uncertainty-aware learning with two variants: DUAL-S for single-modal tasks and DUAL-M for multi-modal tasks. DUAL-S combines Dynamic Feature Uncertainty Modeling (DFUM) and Adaptive Distribution-Aware Modulation (ADAM); DUAL-M additionally introduces Uncertainty-Aware Cross-Modal Relationship Learning (UCRL). The empirical section reports accuracy improvements on CIFAR-10, CIFAR-100, Tiny-ImageNet, CMU-MOSEI, CMU-MOSI, and MISR, with gains up to 7.1% and 6.2% in different settings. An appendix provides a theoretical analysis of dynamic feature uncertainty and contextual information loss.","tokens_in":13771,"tokens_out":6263,"duration_ms":51655,"significance":"If the reported improvements were reproducible, DUAL could serve as a broadly applicable plug-in for existing backbones, and the idea of modeling uncertainty as time-varying during training is potentially useful. The paper covers a wide range of architectures and tasks, which is a strength. However, the current manuscript does not provide sufficient experimental detail to verify any of the central claims, and the theoretical appendix does not analyze the proposed method. The presence of a self-evaluative passage titled 'AI Review Guidance Notes' at the end of the paper is inappropriate and raises concerns about the scholarly integrity of the submission.","major_comments":[{"comment":"No training protocol is reported: the manuscript does not state epochs, batch size, optimizer, learning rate schedule, data augmentation, number of seeds, or train/validation splits for any of the six datasets. Without this information, the baseline models cannot be assumed to be trained to a comparable standard, and the reported improvements (e.g., +7.1% for EfficientNet-B0 on CIFAR-10, +6.2% Acc-7 for TFN on CMU-MOSEI) cannot be verified. The absolute numbers in Table 1 (e.g., EfficientNet-B0 at 82.0% top-1) are far below typical published results for CIFAR-10, which strongly suggests that the baselines were undertuned.","section":"Section 5, Tables 1–5"},{"comment":"The framework introduces many balancing hyperparameters (α and β in Eq. 8; λ_KL in Eq. 14; α0, γ, τ, R in Eqs. 16–17; η0 and λ in Eq. 20; β and γ in Eqs. 26–27), but no values are reported and no sensitivity analysis is provided. The reported gains may therefore result from hyperparameter overfitting to the test sets, and the claim of consistent improvements across tasks is unsupported.","section":"Equations 8, 14, 16, 17, 20, 26, 27"},{"comment":"The theorem only derives the change in feature representation under SGD, which is a standard chain-rule consequence; it does not analyze the proposed dynamic uncertainty model in Eq. 15, the modulation in Eq. 16, or the DUAL objectives in Eqs. 19 and 27. The proof is therefore tautological with respect to the method and provides no theoretical support for the central claims. The subsequent contextual-information analysis (Theorem .3) is likewise unrelated to the algorithm actually proposed.","section":"Appendix B.1, Theorem .1"},{"comment":"The uncertainty estimate x_uncert(t) is updated using the gradient ∇θ L_task, and this same task loss is then modulated by the uncertainty estimate in L_ADAM. This feedback loop between the uncertainty estimator and the loss being optimized is not analyzed. The paper provides no stability or convergence argument, so the algorithm's behavior during training is not understood.","section":"Section 4.1, Eq. 15 and Section 4.2, Eq. 16"},{"comment":"The ablation study uses a single baseline (BioVisionEEG) and reports no variance or significance. The differences between configurations are small (e.g., 50.4 vs. 50.2 accuracy), and without repeated seeds the claim that the components contribute synergistically is not supported.","section":"Section 5.4, Table 6"}],"minor_comments":[{"comment":"The symbol β is overloaded: it is used as a balancing weight in Eq. 8, as a temperature parameter in Eq. 26, and as a regularizer weight in Eq. 27, which creates confusion.","section":"Equations 8, 26, 27"},{"comment":"The figures showing early training and testing dynamics lack legends that identify which curve corresponds to DUAL-S and which to the baseline; the reader cannot interpret the claimed stabilization and late-training gains.","section":"Figures 2 and 3"},{"comment":"Appendix A states that DUAL-S achieves a '4.2% gain' after epoch 40, but no such number appears in Tables 1–3, and the experimental setup behind this statement is not described.","section":"Appendix A"},{"comment":"The header 'Acc(%↑)' is ambiguous: the rows for DFUM-only and other configurations appear to list absolute accuracies, not improvements, while the caption uses 'improvement' terminology.","section":"Table 6"},{"comment":"The gradient ∇θ L_task is used in the uncertainty update, but it is not specified whether θ denotes the backbone parameters, the DFUM parameters, or the full model; this ambiguity affects the reproducibility of the update.","section":"Eq. 15"}],"recommendation":"reject","confidential_remarks":"The final page of the manuscript contains a passage titled 'AI Review Guidance Notes' that is not part of a scientific paper; it appears to be a template intended to influence the reviewer. This is a serious integrity concern that should be investigated by the editor. The paper's central empirical claims are also entirely unverifiable as submitted because no training details, hyperparameters, or code are provided, and the theoretical appendix does not analyze the proposed method. Even if the missing details were supplied, the low absolute baseline numbers raise doubts about the fairness of the comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: DUAL combines known pieces (Gaussian uncertainty with KL, loss modulation, MMD alignment, uncertainty-weighted fusion) into a package that is genuinely new in its assembly, and the temporal recurrent state h(t) that refines uncertainty during training is a real twist. The multi-modal covariance model in Eq. 23 is also something I have not seen. So the framework is worth thinking about. But the paper as written does not support its headline claims. The experimental section gives no training protocol: no epochs, batch size, learning rate, augmentation, seeds, splits, or values for the many hyperparameters in Eqs. 8, 14, 16, 17, 19-20, 23-27. Tables 1-5 are just 'baseline + improvement' numbers. Without the protocol, you cannot rule out under-tuned baselines or test-set cherry-picking. The magnitudes make this worse: EfficientNet-B0 at 82.0 on CIFAR-10 and 6.2-point Acc-7 jumps on MOSEI are far beyond seed noise if the baselines were trained well. The appendix theory does not rescue it. Theorem .1 just restates the SGD update; Theorem .3 is a generic information inequality with no connection to the method. And the stray 'AI Review Guidance Notes' block at the end of Appendix B.2 is a red flag that the manuscript was not carefully prepared. The ablation study is on one dataset only and also lacks details. Credit where due: the writing is readable, the motivation is sensible, and the framework is not ad hoc. The problem is not the idea; it is the complete absence of reproducibility artifacts and protocol. A paper that claims 1-7 point gains across six benchmarks needs code, hyperparameters, error bars, and multiple seeds. This one ships none of those. My recommendation: do not accept. The manuscript needs a full rewrite of the experimental section with all details before it can be evaluated. If the authors provide code and a real protocol, it might turn into a solid workshop or conference paper. As submitted, it is not ready for a serious referee.","headline":"A coherent new framework with no verifiable evidence: the recurrent uncertainty state is a real twist, but missing training details, no code, and a stray 'AI Review Guidance Notes' block sink the empirical claim.","tokens_in":14299,"tokens_out":2201,"would_cite":false,"duration_ms":19745,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DUAL claims that modeling feature uncertainty as a dynamic, time-varying quantity improves accuracy by up to 7.1% across vision and multimodal benchmarks.","keywords":["dynamic uncertainty","feature uncertainty","multi-modal learning","adaptive modulation","uncertainty-aware fusion","cross-modal relationship","sentiment analysis","inner speech recognition"],"falsifier":"Train the same backbone with and without DUAL under identical seeds, epochs, and augmentations, using the paper's reported baseline configuration; if the accuracy gap shrinks below the reported margin or the baseline matches published standard results (e.g., EfficientNet-B0 well above 82% on CIFAR-10), the claimed improvements are artifacts of a weak comparison. Alternatively, run a small grid over α0 and τ; if gains vanish for nearby values, the results are hypersensitive to unreported hyperparameters.","tokens_in":13235,"feed_emoji":"📈","tokens_out":5399,"duration_ms":43924,"temperature":0.7,"pith_summary":"DUAL claims that feature uncertainty in deep learning is dynamic—it changes as training progresses—and that modeling it as a learnable time-varying component improves accuracy and stability. The framework has two variants: DUAL-S for single-modal tasks (CIFAR-10/100, Tiny-ImageNet) and DUAL-M for multi-modal tasks (CMU-MOSEI, CMU-MOSI, MISR). Reported gains over existing backbones range from 1.1% to 7.1% accuracy, with the largest jumps seen on EfficientNet-B0 for CIFAR-10 and on fine-grained sentiment classification. The paper's message is that uncertainty-aware modules can be added to existing architectures with minimal overhead and consistent benefit.","feed_headline":"Dynamic uncertainty lifts accuracy by up to 7.1%","feed_subtitle":"DUAL models feature uncertainty as time-varying and reports consistent gains on vision and multimodal benchmarks.","key_machinery":"The load-bearing object is the decomposition $x_i = x_i^c + x_i^u(t)$, where $x_i^u(t)$ is produced by a learnable function of the observed features and a recurrent temporal state $h(t)$. From this, the paper builds a Gaussian uncertainty estimator with reparameterized sampling, a temporal smoothness regularizer, an adaptive distribution-aware modulation loss $L_{\\text{ADAM}}$ that alternates between a capped loss and a log-scaled loss, an MMD-based distribution alignment term, and, for DUAL-M, a relationship module that computes cross-modal uncertainty covariances $\\Sigma_{m,n}(t)$ and uses their traces to set fusion weights. This machinery turns 'uncertainty' from a static property into a trainable quantity that evolves with the learning dynamics.","core_discovery":"The central discovery is that representing each feature as the sum of an observable part and a time-varying uncertainty part, estimated through a recurrent neural state, lets a model absorb incomplete or noisy inputs more reliably. Uncertainty estimates feed an adaptive loss that adjusts sample influence in a distribution-aware way, and in multi-modal settings a dedicated module estimates the uncertainty of cross-modal relationships and uses it to weight fusion. On the paper's own terms, DUAL-S improves Top-1 accuracy by up to 7.1% on CIFAR-10, 6.5% on CIFAR-100, and 2.3% on Tiny-ImageNet, while DUAL-M gains 4.1% on CMU-MOSEI, 2.8% on CMU-MOSI, and 1.4% on MISR over the listed baselines.","pith_inferences":["A direct test the paper does not run: freeze the uncertainty estimates after a warm-up phase and retrain; if most of the gain survives, the dynamic component is not the main driver.","The 7.1% gain on CIFAR-10 comes from an EfficientNet-B0 baseline at 82.0% Top-1, well below typical published figures with standard augmentation; a fair replication should compare against a properly tuned baseline to separate real gains from undertrained baselines.","The ADAM loss's periodic min/log form resembles robust-loss families; one could test whether DUAL's benefit mostly comes from this loss shape rather than from the dynamic uncertainty network.","The Gaussian assumption, acknowledged by the paper as a limitation, means the framework may underperform on heavy-tailed or discrete uncertainty patterns; extending it to mixture or nonparametric distributions is a natural next step."],"forward_implications":["If the gains hold, uncertainty-aware modules can be bolted onto existing backbones without changing the base architecture, giving a general upgrade path.","Multi-modal fusion becomes uncertainty-weighted: pairs of modalities with less reliable relationships automatically contribute less to the fused representation.","The dynamic (time-varying) treatment implies that uncertainty estimates improve as training proceeds, so early training is slower but final generalization and stability improve.","The framework spans vision classification, multimodal sentiment analysis, and EEG-based inner speech recognition, suggesting the mechanism is task-agnostic."],"supporting_citations":[{"why":"Introduces the distinction between aleatoric and epistemic uncertainty that motivates modeling feature-level uncertainty in deep learning.","marker":"(Kendall & Gal, 2017)"},{"why":"Provides the variational dropout approach to uncertainty that the paper positions DFUM against as a static baseline.","marker":"(Gal & Ghahramani, 2016)"},{"why":"Supplies the deep-ensemble method for uncertainty estimation that motivates a cheaper dynamic alternative.","marker":"(Lakshminarayanan et al., 2017)"},{"why":"Defines the CMU-MOSEI and CMU-MOSI benchmarks used for all multi-modal sentiment analysis results.","marker":"(Zadeh et al., 2018)"},{"why":"Supports the claim that distribution shifts in features degrade generalization, the motivation for the adaptive distribution-aware modulation component.","marker":"(Qin et al., 2022)"}],"fun_headline_variants":["Dynamic uncertainty modeling boosts accuracy up to 7.1%","Uncertainty-aware learning lifts vision and multimodal benchmarks","DUAL: Up to 7.1% accuracy gain from feature uncertainty","Modeling feature uncertainty improves accuracy by up to 7.1%","Uncertainty-aware cross-modal fusion yields consistent gains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported improvements rest on the untested premise that the baseline models were trained to the same standard as the DUAL-enhanced versions under an identical, unreported protocol, and that the many hyperparameters (α, β, γ, λ, λ_KL, α0, τ, R, η0) were not tuned to the test sets; the framework also assumes uncertainty is Gaussian, which the paper itself flags as a limitation.","fun_headline_variants_meta":{"raw":{"variants":["Dynamic uncertainty modeling boosts accuracy up to 7.1%","Uncertainty-aware learning lifts vision and multimodal benchmarks","DUAL: Up to 7.1% accuracy gain from feature uncertainty","Modeling feature uncertainty improves accuracy by up to 7.1%","Uncertainty-aware cross-modal fusion yields consistent gains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000248,"raw_usage":{"total_tokens":1541,"prompt_tokens":933,"completion_tokens":608,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":521}},"tokens_in":549,"tokens_out":608,"duration_ms":5574,"temperature":1.0,"reasoning_tokens":521,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:09:18.911291+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same backbone with and without DUAL under identical seeds, epochs, and augmentations, using the paper's reported baseline configuration; if the accuracy gap shrinks below the reported margin or the baseline matches published standard results (e.g., EfficientNet-B0 well above 82% on CIFAR-10), the claimed improvements are artifacts of a weak comparison. Alternatively, run a small grid over α0 and τ; if gains vanish for nearby values, the results are hypersensitive to unreported hyperparameters.","supporting_citations":[],"review_version":1}