{"id":"f5f69e4d-6323-4ca0-aaa1-3466a7f44274","arxiv_id":"2608.09986","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"MIDAS combines variational latent disentanglement, mutual information minimax objectives, and uncertainty-weighted attention to improve multimodal sentiment analysis under incomplete inputs.","lead":"This paper introduces MIDAS, a neural framework for sentiment analysis when parts of the text, audio, or video input are missing or corrupted. It separates each modality into shared and modality-specific factors using mutual information objectives and weights each modality by its estimated uncertainty during fusion.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported SOTA gains rest on single-run comparisons with small or negative margins on some metrics; without error bars or significance tests the central claim of strong, consistent improvement is not established.","rationale":"The reader's conditional verdict is reasonable, and their condition (1) already asks for error bars and significance tests. However, the reader's weakest_assumption singles out posterior-variance calibration in Eqs. (15)-(17) as the load-bearing risk. I view the absence of statistical grounding as more load-bearing for the central claim: even if the uncertainty weights are perfectly calibrated, the headline 'strong and consistent gains' would still be unsupported if the numerical differences are within seed noise. Conversely, if the uncertainty weights are somewhat miscalibrated, the method could still succeed through the MIM and VM components; the UAF ablation in Table IV shows only modest, mixed gains (e.g., CH-SIMS MAE worsens from 0.505 to 0.513 when UAF is added to VM), so the entire contribution does not hinge on UAF alone. The paper has real strengths: code is provided, hyperparameters are fixed, ablations are systematic, and the variational MI upper bound in Eq. (9) is a recognizable and defensible construction despite notation slips. Those strengths make the work conditionally acceptable, but the statistical support must be supplied before the SOTA claim can be taken at face value. My concrete test is deliberately narrow: recompute the headline comparisons with seeds and paired tests, since that directly settles whether the central claim lands.","tokens_in":22322,"tokens_out":18016,"duration_ms":195777,"concrete_test":"Run the released code (https://github.com/ultramarineX/MIDAS) and the official implementations of the top baselines (EMT-DLFR, P-RMF, LNLN) for 10 random seeds on MOSI, MOSEI, and CH-SIMS under the exact protocol of Section IV-D (unified checkpoint, same missing-rate schedule). Report mean and standard deviation for every metric and compute paired bootstrap or Wilcoxon tests on per-sample predictions at each missing rate. If the MIDAS-versus-best-baseline differences are not significant at p < 0.05 (with multiple-comparison correction), the central claim should be softened to 'competitive' and the verbatim claim of new state-of-the-art should be withdrawn. If significance is confirmed, the evidentiary concern is resolved and the conditional acceptance can proceed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is evidentiary rather than architectural: the central claim of 'strong and consistent performance gains' (Abstract; Section I) rests on single-run comparisons in Tables II and III. The margins over the strongest baseline are small on several metrics—MOSI Acc-2 +0.84, MOSEI Acc-2 +0.65, CH-SIMS Acc-2 +2.27—and negative on MOSI Acc-5 (35.12 vs 35.19) and MAE (1.074 vs 1.065). No seed variance, confidence intervals, or paired significance tests are reported anywhere. In MSA benchmarks, typical seed-level standard deviations are of the same order as these margins, so the reported numbers do not by themselves establish that MIDAS outperforms the baselines. If the gaps are within run-to-run noise, the headline claim and the 'new state-of-the-art' statement collapse. The ablation studies are internally consistent, but they do not resolve the significance question: the full model is better than ablated variants on most metrics, yet the reported differences are also single-run and could reflect seed luck rather than component efficacy.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MIDAS, a framework for incomplete multimodal sentiment analysis. Each modality is encoded into a pair of multivariate Gaussian latent variables, factorized into shared and exclusive components. A mutual-information minimax objective is used to disentangle these components and align shared spaces across modalities, and an uncertainty-aware fusion module computes reliability weights from posterior variance and modulates attention keys. The authors train a single model under stochastic feature masking with rates 0.0–0.9 and evaluate on MOSI, MOSEI, and CH-SIMS, reporting aggregate robustness metrics, ablations, an efficiency analysis, and a qualitative case study.","tokens_in":22609,"tokens_out":5739,"duration_ms":63532,"significance":"If the claims are supported, the paper makes a useful contribution: it addresses incomplete multimodal sentiment analysis without imputation, combines a variational latent formulation with MI-based disentanglement, and uses intrinsic posterior uncertainty for fusion. The paper also releases code, evaluates a single unified model across missing rates, compares against both complete and incomplete baselines, and includes component-level ablations and a complexity analysis. These are genuine strengths. However, the central empirical claim of 'strong and consistent performance gains' is not yet established because all comparisons are single-run, the theoretical bound in Eq. (9) is not a valid variational upper bound as written, and the reliability assumption behind the uncertainty weighting is asserted rather than validated. The contribution is therefore conditional on resolving these load-bearing issues.","major_comments":[{"comment":"The central claim of strong and consistent improvements rests on single-run comparisons. No seed variance, confidence intervals, or paired significance tests are reported anywhere, and several margins over the strongest baseline are very small: on MOSI, Acc-2 improves by 0.84 over EMT-DLFR but Acc-5 is worse (35.12 vs 35.19), and MAE is worse than Self-MM/TETFN/TFR-Net (1.074 vs 1.065); on MOSEI, the Acc-2 gain over EMT-DLFR is only 0.41. In MSA benchmarks, seed-level deviations are typically of the same order as these margins, so the reported numbers do not by themselves establish that MIDAS outperforms the baselines. Please report at least 3–5 seeds per method and dataset, provide confidence intervals or paired tests for the aggregated metrics, and discuss per-missing-rate variability; the abstract and Section I should then be reworded to match the strength of the evidence.","section":"Section V-A, Tables II and III"},{"comment":"The derivation of the variational upper bound is not valid as written. Starting from Eq. (8), the KL-divergence terms can upper-bound I(Z_s_m; X_m) and I(Z_e_m; X_m), but subtracting an expected log-likelihood term does not guarantee an upper bound on I(Z_s_m; Z_e_m); when the reconstruction likelihood is high, the RHS can become negative and even unbounded below. In addition, the expectation notation in Eq. (9), 'E_{p(Z_s,Z_e)} E_{q(Z_s|X)q(Z_e|X)}', is ambiguous because X is not specified. The authors should state the precise variational identity being used, or clearly identify Eq. (9) as a heuristic objective rather than a bound, since the theoretical grounding of L_dis depends on this step.","section":"Section III-C, Eq. (9)"},{"comment":"The core assumption that posterior variance of the variational latents is a valid reliability indicator for fusion is asserted without calibration or empirical validation. Under random feature masking, the variance may reflect encoder sensitivity, initialization, or optimization state rather than true predictive uncertainty. If high variance does not track low reliability, the reliability-modulated attention in Eq. (17) can downweight informative tokens and the advantage of the uncertainty mechanism would vanish. The authors should provide evidence that the estimated uncertainties are calibrated, for example by showing that uncertainty increases with missing rate and correlates with per-sample prediction error, or by comparing against an oracle reliability weighting that uses the true missingness or error.","section":"Section III-D, Eqs. (15)–(17)"},{"comment":"The simplification that I(Z_s_m; Z_e_m | X_m) = 0 is an independence assumption, not a consequence of the variational factorization, and it is not stated explicitly as an assumption. Conditioning on X_m does not by itself make the shared and exclusive latents conditionally independent. If this conditional-independence assumption fails, the third term in Eq. (7) cannot be dropped and the subsequent expansion in Eq. (8) is incomplete. Please state this assumption explicitly and justify it, or keep the conditional mutual-information term and explain how it is handled.","section":"Section III-C, Eq. (7)"}],"minor_comments":[{"comment":"The sentence 'Here, p(X_m) is typically chosen as a standard Gaussian prior' appears to be a typo: the priors should be p(Z_s_m) and p(Z_e_m), not p(X_m).","section":"Section III-C, after Eq. (9)"},{"comment":"The F1 value listed for EMT-DLFR (69.39/67.58) is identical to that of TFR-Net and differs from the EMT-DLFR F1 reported in Table II (70.87/70.18); this appears to be a transcription error and should be corrected.","section":"Table VI"},{"comment":"There are several typos, including 'encocding' and 'vatiational' in the Fig. 2 caption, 'as as a measure' in Section I, and 'uncertainity' in the Index Terms; a careful proofreading pass is needed.","section":"Fig. 2 and Section I"},{"comment":"The notation in Eq. (17) is unclear: W is defined as a 6-dimensional vector, but the expression Q(√W K)^T requires a precise definition of how the vector is expanded across the key dimension and how the square root is applied; please clarify.","section":"Section III-D, Eq. (17)"},{"comment":"The claim that the five objectives 'act in a complementary rather than adversarial manner' is presented without quantitative evidence; adding loss curves or gradient-conflict measurements would make this statement verifiable.","section":"Section V-C, 'Training Stability and Co-optimization'"}],"recommendation":"major_revision","confidential_remarks":"The paper is potentially suitable for publication if the authors can provide multi-seed evaluation results that confirm the headline improvements, correct or re-scope the theoretical claim in Eq. (9), and validate the uncertainty-based reliability mechanism. The current single-run comparisons make the 'new state-of-the-art' claim too strong for the evidence presented."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: MIDAS is a coherent architecture paper for incomplete multimodal sentiment analysis, but the headline claim of strong and consistent gains over state-of-the-art baselines is not supported by the reported evidence. The single-run margins on Tables II and III are mostly 0.4 to 2.3 points, with a couple of negative results (MOSI Acc-5 and MAE) that the discussion glosses over. In this benchmark family, seed-level noise is typically the same order as those margins, so without error bars or significance tests the central claim is not established.\n\nWhat is actually new: the specific combination of variational shared/exclusive factorization with an MI minimax objective and posterior-variance-modulated attention. Each ingredient is drawn from prior work—the variational MI upper bound follows the style of [62] and [63], and the uncertainty-aware weighting is adapted from [23] and [58]—but the assembly for missing-modality settings is new, and the authors provide code. The ablations are internally consistent: removing the prediction loss hurts sharply, removing disentanglement or alignment losses degrades most metrics, and the full model generally beats its own ablated variants. That gives me some confidence the components are doing real work.\n\nThe soft spots are evidentiary and conceptual. No multi-seed runs, no confidence intervals, no paired tests. The derivation of Eq. (9) is under-specified: the step from the interaction-information decomposition to the variational upper bound is not fully spelled out, and the expectation notation is ambiguous. The more fundamental concern is the uncertainty signal: posterior variance is used as a reliability indicator without any calibration against actual missingness or reconstruction error. If variance does not track reliability under distribution shift, the uncertainty-aware fusion could downweight useful features and the method's advantage would plausibly vanish. That assumption is plausible but untested.\n\nWho this is for: researchers working on incomplete multimodal learning or robustness under missing modalities will find the architecture worth reading, and the available code helps. It does not introduce new theory or a novel capability, and the claimed gains are modest.\n\nRecommendation: do not desk-reject. The paper is coherent, well-situated in the literature, and reports new empirical results on three standard benchmarks. Send it to review, but instruct referees to require multi-seed runs with standard deviations, significance tests where appropriate, and either a validation of the reliability assumption or a softened claim about uncertainty weighting. Without those changes, the state-of-the-art claim should not stand as written.","headline":"Sound architecture, weak evidence: single-run comparisons don't support the SOTA claim, but the paper is coherent and deserves refereeing.","tokens_in":23099,"tokens_out":1927,"would_cite":false,"duration_ms":21578,"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":"MIDAS claims that incomplete multimodal sentiment inputs can be handled by disentangling each modality into shared and exclusive Gaussian latents and fusing them by posterior uncertainty, without imputing missing data.","keywords":["multimodal sentiment analysis","incomplete modalities","mutual information","disentangled representation learning","variational inference","uncertainty-aware fusion","missing data","sentiment analysis"],"falsifier":"The claim would be falsified by an ablation that replaces the uncertainty weights with uniform attention and finds the averaged accuracy across missing rates from 0.0 to 0.9 changes by less than noise; then the posterior-variance weighting is not what produces the reported robustness.","tokens_in":22157,"feed_emoji":"🧠","tokens_out":5970,"duration_ms":59186,"temperature":0.7,"pith_summary":"MIDAS tackles multimodal sentiment analysis when one or more input channels—text, audio, or video—are missing or corrupted. The paper's claim is that a single model can stay accurate from fully intact input up to heavy missingness by restructuring each modality into shared and private Gaussian latent factors, then fusing them by their learned uncertainty rather than treating all channels equally. The authors argue this beats both data-imputation methods, which reconstruct missing inputs first, and coordination methods, which try to keep representations consistent across modality combinations. Across three public datasets and missing rates from 0% to 90%, MIDAS reports consistent gains over ten baselines. If true, it would make sentiment systems usable in real deployments where sensors fail, connections are poor, or privacy filters remove parts of the signal.","feed_headline":"MIDAS keeps sentiment predictions accurate as modalities vanish","feed_subtitle":"A mutual-information minimax plus uncertainty-weighted fusion beats ten baselines across missing rates from 0 to 90 percent.","key_machinery":"The load-bearing mechanism is a mutual-information minimax over disentangled Gaussian latents, paired with reliability-modulated attention. Each modality is represented by shared and exclusive multivariate Gaussian posteriors; the objective minimizes mutual information within a modality between the shared and exclusive factors via a variational upper bound, and maximizes mutual information across modalities between shared factors via a Jensen–Shannon divergence estimator with a critic. The variance of each posterior is converted by differential entropy and a temperature-scaled sigmoid into a reliability weight, and the weights scale the key vectors inside a Transformer-style attention block so uncertain tokens contribute less to fusion. This is what carries the claim that the model can maintain semantic alignment and robust prediction when modalities are partly erased.","core_discovery":"The central discovery, as the paper states it, is that incomplete multimodal inputs are best handled by reorganizing the latent space rather than filling in missing data or imposing geometric constraints. Each modality is encoded as a pair of multivariate Gaussians—a shared factor meant to carry modality-invariant sentiment semantics and an exclusive factor meant to carry modality-specific detail and noise. A minimax mutual-information objective minimizes the dependence between shared and exclusive factors within a modality while maximizing the dependence among shared factors across modalities, so the shared space becomes semantically aligned. Prediction and reconstruction losses anchor the latents to the task. Fusion then uses the posterior variance of each latent token as a reliability weight inside attention, so tokens whose distributions are broad—taken as a sign of missing or unreliable information—are downweighted. The paper's experiments on MOSI, MOSEI, and CH-SIMS show MIDAS outperforming all compared methods on most metrics averaged across missing levels, with the largest margins on the Chinese CH-SIMS dataset.","pith_inferences":["My inference: the same uncertainty weighting could transfer to other multimodal tasks—emotion recognition or human-computer interaction—where a channel can be unreliable for reasons other than missingness, such as loud acoustic noise or blurred video, provided the variance estimate remains calibrated to that noise.","My inference: the reliability-weight mechanism is only as good as the learned variance; if the encoder's variance collapses or becomes scale-invariant across inputs, the attention modulation degenerates. A cheap probe would be to compare prediction accuracy under uniform weights versus learned weights on held-out missing patterns.","My inference: the paper's benchmark fixes the same missing rate across all modalities. Asymmetric missing patterns—say, 80% of audio missing but only 10% of text—are the harder real-world case, and the method's disentanglement objective is likely where any advantage would show or fail.","My inference: the claim that posterior variance tracks reliability could be tested by correlating per-token variance with whether that token's input region was actually masked; a strong correlation would make the mechanism auditable."],"forward_implications":["A single MIDAS checkpoint trained on mixed missing rates can be deployed under heterogeneous sensor failures without retraining per missing pattern, since the model is evaluated on all rates with one selected checkpoint.","Downweighting unreliable tokens during fusion should make predictions degrade more slowly as missingness increases, which is the smooth performance curve the paper reports.","Because shared latents are forced to align across modalities, the model can lean on whichever channels remain, so severe loss in one channel should not be able to dominate or silently corrupt the final prediction.","The uncertainty weights come from the variational posterior itself, so no auxiliary uncertainty estimator or Monte Carlo sampling is needed at inference.","The method's parameter count stays low beyond the text encoder while giving the reported accuracy, suggesting that robust handling of missingness does not require a larger model."],"supporting_citations":[{"why":"Supplies the Transformer reconstruction baseline and the [UNK] masking convention for incomplete text inputs that MIDAS compares against.","marker":"[8]"},{"why":"Defines the incomplete-MSA evaluation setting with uniform missing rates and a single unified model, the protocol the paper adopts.","marker":"[17]"},{"why":"Gives the prior mutual-information disentanglement approach for multimodal sentiment that MIDAS extends to incomplete inputs and compares with.","marker":"[21]"},{"why":"Provides the Jensen–Shannon divergence estimator used as the variational lower bound for cross-modal shared-space alignment.","marker":"[36]"},{"why":"Supplies the reparameterization trick that makes sampling from the Gaussian latent posteriors differentiable.","marker":"[61]"},{"why":"Provides the interaction-information decomposition used to turn intractable mutual information into a variational upper bound.","marker":"[62]"},{"why":"Shows modality-wise aleatoric uncertainty used in multimodal fusion, the lineage the uncertainty-aware module builds on.","marker":"[23]"},{"why":"Introduces Gaussian uncertainty representations that regulate fusion variance, a direct predecessor of the reliability-modulated attention.","marker":"[58]"}],"fun_headline_variants":["Mutual information minimax rescues sentiment when modalities drop","MIDAS disentangles shared and exclusive signals for missing-modality sentiment","Uncertainty-weighted fusion beats imputation for incomplete sentiment","Missing modalities? MIDAS reweights latent reliability instead of imputing","MIDAS: turning missing inputs into reliability weights for sentiment"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the variance of the learned Gaussian posterior is a faithful signal of how unreliable that token is, so that downweighting high-variance tokens improves fusion instead of discarding useful information.","fun_headline_variants_meta":{"raw":{"variants":["Mutual information minimax rescues sentiment when modalities drop","MIDAS disentangles shared and exclusive signals for missing-modality sentiment","Uncertainty-weighted fusion beats imputation for incomplete sentiment","Missing modalities? MIDAS reweights latent reliability instead of imputing","MIDAS: turning missing inputs into reliability weights for sentiment"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000807,"raw_usage":{"total_tokens":3559,"prompt_tokens":975,"completion_tokens":2584,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":2498}},"tokens_in":591,"tokens_out":2584,"duration_ms":19473,"temperature":1.0,"reasoning_tokens":2498,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:50:46.070335+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The claim would be falsified by an ablation that replaces the uncertainty weights with uniform attention and finds the averaged accuracy across missing rates from 0.0 to 0.9 changes by less than noise; then the posterior-variance weighting is not what produces the reported robustness.","supporting_citations":[{"cited_title":"Transformer-based feature recon- struction network for robust multimodal sentiment analysis,","cited_arxiv_id":null,"evidence_quote":"Supplies the Transformer reconstruction baseline and the [UNK] masking convention for incomplete text inputs that MIDAS compares against."},{"cited_title":"Towards robust multimodal sentiment analysis with incomplete data,","cited_arxiv_id":null,"evidence_quote":"Defines the incomplete-MSA evaluation setting with uniform missing rates and a single unified model, the protocol the paper adopts."},{"cited_title":"Mul- timodal sentiment analysis with mutual information-based disentangled representation learning,","cited_arxiv_id":null,"evidence_quote":"Gives the prior mutual-information disentanglement approach for multimodal sentiment that MIDAS extends to incomplete inputs and compares with."},{"cited_title":"Learning deep representations by mutual information estimation and maximization,","cited_arxiv_id":null,"evidence_quote":"Provides the Jensen–Shannon divergence estimator used as the variational lower bound for cross-modal shared-space alignment."},{"cited_title":"Feddcsr: Federated cross-domain sequential recommendation via disentangled representa- tion learning,","cited_arxiv_id":null,"evidence_quote":"Provides the interaction-information decomposition used to turn intractable mutual information into a variational upper bound."},{"cited_title":"Embracing unimodal aleatoric uncertainty for robust multimodal fusion,","cited_arxiv_id":null,"evidence_quote":"Shows modality-wise aleatoric uncertainty used in multimodal fusion, the lineage the uncertainty-aware module builds on."},{"cited_title":"Cold fusion: Calibrated and ordinal latent dis- tribution fusion for uncertainty-aware multimodal emotion recognition,","cited_arxiv_id":null,"evidence_quote":"Introduces Gaussian uncertainty representations that regulate fusion variance, a direct predecessor of the reliability-modulated attention."}],"review_version":1}