{"id":"c1625e27-1d62-44ac-98af-08e4f651df78","arxiv_id":"2506.10452","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CIDer fuses weight-sharing self-distillation and a causal debiasing module to improve emotion recognition under random feature loss and distribution shift.","lead":"This paper presents CIDer, a model that combines self-distillation and causal inference to keep multimodal emotion recognition accurate when input features are randomly missing and when test data shifts from training data. The authors also define a generalized missing-feature task and repartition two standard datasets to create OOD test sets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The new 'OOD' datasets are never shown to contain distribution shift; Section IV-A gives no repartitioning rule, and the split sizes match IID, so Table IV's OOD columns may only measure IID robustness.","rationale":"I agree with the reader's weakest-assumption identification. The paper's headline claim is that CIDer 'achieves robust performance in both RMFM and OOD scenarios'; the RMFM part is backed by a concrete missing-rate protocol, but the OOD part is backed only by an undescribed repartitioning of the datasets. The experiment tables treat 'OOD' as a column label, so the dataset construction is the load-bearing condition. Because the counts match IID and no shift statistic is reported, the concern is not about disagreement with consensus but about whether the experiments measure what they claim. A random split seeded differently would produce exactly the observed table structure. This is independently checkable from the public code, so the appropriate outcome is a conditional acceptance: the OOD claims should be verified against the split-generation script and shift statistics before the central claim is accepted. I do not see a need to move the reader's verdict; the same condition is already identified. Secondary worries about Eq. (25)'s subtraction form and unreported tau are real but less central, because even a perfect causal module cannot demonstrate OOD generalization without a valid OOD test set.","tokens_in":37128,"tokens_out":4179,"duration_ms":50674,"concrete_test":"Inspect the released repository's data split files for the new OOD datasets. For each of CMU-MOSI and CMU-MOSEI, compute per-label word frequency vectors for train and test, measure the shift with KL divergence/total variation or a train-vs-test text classifier AUROC, and compare these values against (a) a random IID split of the same data and (b) Sun et al.'s published OOD splits. Also check whether the new OOD test sample IDs coincide with the original IID test IDs. If the new splits show no larger shift than a random split, or if the OOD test sets are identical to the IID test sets, the OOD results in Table IV do not substantiate the paper's central OOD claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central OOD-robustness claim rests on the 'repartitioned MER OOD datasets' introduced in Section IV-A. The entire construction is described by one sentence: 'we repartitioned the original CMU-MOSI and CMU-MOSEI datasets.' No algorithm, criterion, or distributional check is given for how train/validation/test word or label distributions were made to differ. Table I aggravates the problem: the new OOD splits have essentially the same cardinalities as the original IID splits (MOSI 1284/229/686; MOSEI 16326/1871/4659 vs 16326/1872/4658), so a plausible implementation is a random reshuffle, which preserves the IID word-label correlation. Footnote 1 defines OOD as 'deviations in word distribution and the resulting deviations in label distribution'; if the new splits do not create such deviations, then the OOD columns of Table IV, the MACI plug-in gains in Table IV, and the 'RMFM and OOD' results are not evidence about OOD generalization. They would instead be repeat IID evaluations. This is a load-bearing premise because it is the only source of OOD test data in the paper; the subsequent causal-inference apparatus cannot validate an OOD claim if the test set is not OOD.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CIDer, a multimodal emotion recognition (MER) framework that jointly handles random modality feature missing and out-of-distribution (OOD) inputs. The framework combines a Model-Specific Self-Distillation (MSSD) module for robustness to the newly introduced Random Modality Feature Missing (RMFM) task and a Model-Agnostic Causal Inference (MACI) module for OOD generalization. The authors also introduce a generalized RMFM definition, a word-level alignment module (WSAM), a multimodal fusion transformer (MCT), and repartitioned OOD versions of CMU-MOSI and CMU-MOSEI. Experiments compare CIDer with several baselines across IID, OOD, RMFM, RMM, TMFM, STMFM, and cross-dataset settings, reporting state-of-the-art or competitive results with fewer parameters and short training time.","tokens_in":37442,"tokens_out":5026,"duration_ms":65972,"significance":"If the OOD evaluation is valid, the paper makes a practically relevant contribution: a compact MER model that is robust to arbitrary token-level feature loss and to word/label distribution shifts, with a model-agnostic debiasing module (MACI) that can be plugged into existing methods. The public code, the extensive per-missing-rate tables, and the generalization of the missing-modality definition to RMFM are strengths. However, the central OOD claim currently rests on an undocumented repartitioning procedure and on causal equations whose validity is questionable; until those are resolved, the significance of the OOD results cannot be assessed from the manuscript alone.","major_comments":[{"comment":"The construction of the 'repartitioned MER OOD datasets' is not described. The text only states that the authors 'repartitioned the original CMU-MOSI and CMU-MOSEI datasets,' with no algorithm, criterion, or distributional check to ensure that train and test differ in word distribution and, consequently, label distribution. Table I shows that the new OOD splits have almost identical cardinalities to the IID splits (e.g., MOSI 1284/229/686 for both IID and OOD), so a random reshuffle would preserve the IID word-label correlation and would not produce an OOD test set. This is load-bearing because the OOD columns in Tables IV, XVII, and XVIII are the only evidence for the paper's OOD robustness claim. The authors should specify the repartitioning rule, report distributional statistics (e.g., word-frequency shift, label-prior shift, KL divergence) for both datasets, and release the splits; without this, the OOD experiments cannot be interpreted as distribution-shift tests.","section":"Section III-C, Eq. (20)"},{"comment":"The backdoor adjustment formula P(Y|do(L)) = sum_c P(Y|M=f(L,A,V), c)P(c) is not a standard backdoor adjustment for the graph in Fig. 5. In that graph, M is a descendant/mediator of L (since M=f(L,A,V)), so conditioning on M does not give the total causal effect of L on Y; the standard adjustment would use P(Y|L,c) after adjusting for the confounder C. The derivation of Eq. (20) should be supplied, or the claim should be reframed as estimating a different target quantity (e.g., a controlled direct effect). Without a valid derivation, the label-bias-mitigation mechanism of MCM is not established.","section":"Section III-C, Eq. (20)"},{"comment":"The debiasing strength tau in Eq. (25) is a free parameter, but the manuscript does not specify how it is selected. If tau is tuned on the OOD test set, the MACI gains in Table IV partly reflect test-set fitting; if it is fixed, its value and sensitivity should be reported. Similarly, the counterfactual word-selection thresholds (CV_w >= 0.1 and top-100 frequency) are presented without justification or ablation. Since these choices directly control the magnitude of the counterfactual subtraction, their treatment affects the validity of the OOD claim.","section":"Section III-C, Eq. (25)"},{"comment":"The test-time assumption of uniform class priors P(i)=1/cls conflicts with the paper's own definition of OOD in footnote 1, which states that OOD includes 'the resulting deviations in label distribution' caused by word-distribution shifts. If the test label prior actually deviates from uniform under the new OOD splits, then using uniform P(i) in Eq. (24) is an unsupported assumption. The authors should either justify the uniform prior as an interventional distribution, report the actual test label distributions, or include a sensitivity analysis over P(i).","section":"Section III-C, Eqs. (24) and footnote 4"}],"minor_comments":[{"comment":"There is a typo in the first sentence of the OOD scenario paragraph: 'OO) scenarios' should be 'OOD scenarios'.","section":"Section IV-A"},{"comment":"The paper claims the new OOD datasets provide 'word-level aligned and unaligned versions,' but the main OOD experiments in Table IV only use unaligned sequences; no aligned OOD results are reported. This discrepancy should be clarified.","section":"Section IV-A"},{"comment":"The text refers to 'Tabel XI' in the appendix, which should be 'Table XI'.","section":"Appendix"},{"comment":"The notation for counterfactual text is inconsistent: the text writes S_cf = <w_1, [MASK], [MASK], ..., w_Tl>, but the word positions that are not masked are not explicitly defined. Please clarify which words remain after applying the CV and frequency thresholds.","section":"Section III-C"},{"comment":"The conclusion states that MACI 'can be independently applied to enhance the OOD generalization capabilities of existing MER models,' but the only evidence for this is the DLF+MACI and MPLMM+MACI rows in Table IV; no statistical significance or variance information is provided, so the robustness of this claim is not established.","section":"Section VI"}],"recommendation":"major_revision","confidential_remarks":"The decisive issue is the OOD dataset construction: it is described in one sentence and the split counts match the IID splits, so a reader cannot tell whether the OOD experiments measure distribution shift at all. This is fixable if the authors disclose the repartitioning algorithm and provide distributional evidence, but if the splits are merely random reshuffles, the paper's headline OOD claim would collapse. The causal derivation in Eq. (20) also needs a rigorous justification or a reformulation before the label-bias claims can be accepted. I recommend major revision rather than rejection because the contributions could be salvaged with additional analysis and transparent data release."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a look. The RMFM task is a modest but real generalization of the usual missing-modality definitions, and the MSSD module—weight-shared twin networks doing hierarchical distillation—is clean and practical. MACI is also genuinely model-agnostic and adds very few parameters, which is a nice engineering property. The experimental coverage is extensive, the code is public, and the cross-dataset results give some independent support that the method generalizes beyond IID. This is not a paper that reorganizes the field, but it is a solid, useful contribution to robust multimodal emotion recognition.\n\nThe soft spot is exactly where the stress-test note lands. The OOD datasets are supposedly created by repartitioning CMU-MOSI and CMU-MOSEI, but the paper never says how train/validation/test word or label distributions are made to differ. The split counts in Table I are essentially identical to the IID splits, so a random reshuffle would produce the same numbers and still be described as \"repartitioned.\" If that is what happened, the OOD columns in Tables IV and XVII–XVIII are just IID results relabeled. That is a load-bearing problem because the headline OOD-robustness claim depends on those columns. The causal adjustment in Eq. (20) also looks underspecified: conditioning on the mediator M while adjusting for C is not standard backdoor adjustment, and the paper does not defend the step. These issues are addressable, but they have to be addressed.\n\nSmaller concerns: tau and the counterfactual selection thresholds are not reported for the main results, and there are no error bars, so the leaderboard margins are a bit fragile. The cross-dataset experiments help, but they are not the same as the missing OOD construction.\n\nWho gets value from this? Anyone working on MER under missing inputs who wants a compact, trainable baseline, and anyone building debiasing modules they want to bolt onto existing models. A serious referee should see it, but the review has to force the authors to document the OOD repartitioning procedure and show that the train/test distributions actually differ. Without that, the central claim is unverified.","headline":"A genuinely useful missing-modality framework whose central OOD claim currently rests on an unseen, possibly IID-preserving dataset split; worth reviewing carefully but needs a clear description of the repartitioning before the OOD numbers can be trusted.","tokens_in":37940,"tokens_out":2074,"would_cite":false,"duration_ms":29017,"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":"CIDer claims that one compact model, trained once, can stay accurate when arbitrary tokens disappear from language, audio, or visual streams and when test word or label distributions shift.","keywords":["multimodal emotion recognition","missing modalities","random modality feature missing","out-of-distribution generalization","self-distillation","causal inference","counterfactual reasoning","multimodal transformer"],"falsifier":"Compare the unigram word distribution of the new OOD training and test splits; if their divergence is no larger than the divergence between the original IID training and test splits, then the 'OOD' condition is not a real distribution shift and the OOD comparison collapses.","tokens_in":36905,"feed_emoji":"🎭","tokens_out":9736,"duration_ms":98487,"temperature":0.7,"pith_summary":"The paper claims that one compact model can handle two failure modes at once: arbitrary token loss in any of the language, audio, or visual streams, and test-time shifts in word and label distributions. It proposes CIDer, which combines a Model-Specific Self-Distillation (MSSD) module for incomplete inputs and a Model-Agnostic Causal Inference (MACI) module for out-of-distribution inputs. The paper also defines a generalized missing-modality task called RMFM and repartitions CMU-MOSI and CMU-MOSEI into new OOD splits. If the claims hold, practical emotion-recognition systems could be smaller, faster to train, and more reliable under real-world degradation.","feed_headline":"One compact model tops rivals on missing and shifted emotion inputs","feed_subtitle":"New self-distillation and causal modules add about 37K parameters and require no retraining per missing rate.","key_machinery":"The load-bearing machinery is the two-module CIDer architecture. MSSD is a weight-sharing twin network whose complete-view stream acts as teacher for the incomplete-view stream, with SmoothL1 loss on reconstructed unimodal features, KL divergence on attention maps, and cosine similarity on joint representations. MACI is the causal module: during training, a Multimodal Causal Module (MCM) forms class-specific representations from joint features and per-class mean tokens, then averages over class priors to implement backdoor adjustment $P(Y|do(L))$; during testing, counterfactual texts are built by masking words whose inter-class frequency variation is below a threshold, and the counterfactual prediction is subtracted from the main prediction. Two supporting components keep the model compact: WSAM aligns long audio and visual sequences to word level, and the Multimodal Composite Transformer (MCT) computes intra- and inter-modal attention from shared matrices.","core_discovery":"The central discovery claimed is that a weight-sharing twin network trained once can preserve emotion-recognition accuracy when arbitrary tokens are dropped from any modality at missing rates up to 100%, and that a small causal module can then remove label and language biases on shifted test data. MSSD performs hierarchical self-distillation from a complete-view stream to an incomplete-view stream, aligning reconstructed low-level features, attention maps, and joint representations. MACI uses a causal graph with backdoor adjustment during training to counter label bias, and subtracts predictions made from fine-grained counterfactual texts during testing to counter language bias. The paper further claims that MACI adds about 37K learnable parameters and improves OOD accuracy when plugged into existing MER models such as DLF and MPLMM.","pith_inferences":["If RMFM becomes a standard evaluation protocol, missing-modality studies could be compared on one missing-rate axis, making one-to-all training the default instead of per-rate retraining.","The counterfactual-text recipe (top-100 frequent words, coefficient-of-variation threshold 0.1) looks transferable to other sequence tasks where label-language correlation drifts, such as review sentiment or toxicity detection.","MACI's portability across two very different base models suggests OOD debiasing may be separable from fusion architecture; testing it on larger pretrained multimodal encoders would show whether the 37K-parameter module scales.","The MSSD design implies that any future modality-missing type can be handled by the same three-level distillation, so the authors' missing-scenario enumeration is an instantiation rather than a closed list."],"forward_implications":["A single CIDer trained once with random missing rates can be tested across the full 0-100% range under RMFM, traditional RMFM, RMM, TMFM, STMFM, and SMM without retraining at each rate.","Plugging MACI into existing MER models such as DLF and MPLMM improves their OOD performance with about 37K additional parameters.","On the repartitioned CMU-MOSI and CMU-MOSEI OOD splits, CIDer reports the best results on most metrics for complete inputs and for the combined RMFM-plus-OOD setting, with roughly 248K parameters and 10.4 seconds per training epoch.","The hierarchical distillation aligns reconstructed features, attention maps, and joint representations so that the same weight-sharing architecture serves as both teacher and student."],"supporting_citations":[{"why":"Defines the word-based OOD multimodal sentiment task and supplies the CLUE baseline and the dataset split whose mixed-sentence flaw CIDer repartitions.","marker":"[6]"},{"why":"Introduces the feature-reconstruction approach for incomplete multimodal data and the AUILC metric used to report missing-rate performance.","marker":"[11]"},{"why":"The EMT-DLFR baseline with dual-level feature restoration that CIDer compares against in RMFM scenarios.","marker":"[12]"},{"why":"GEAR, a model-agnostic debiasing baseline for OOD multimodal sentiment that CIDer compares with and improves on.","marker":"[16]"},{"why":"Source of noise-imitation adversarial training and the TMFM and STMFM missing-scenario definitions used in the evaluations.","marker":"[18]"},{"why":"LNLN, the strongest incomplete-data baseline, whose larger parameter count is repeatedly cited as the reason it sometimes leads.","marker":"[20]"},{"why":"DiCMoR, a distribution-consistent modality-recovery baseline used for comparison under missing modalities.","marker":"[24]"},{"why":"MPLMM, the prompt-learning RMM baseline into which MACI is plugged to demonstrate model-agnostic portability.","marker":"[27]"},{"why":"CMU-MOSI, one of the two standard datasets used for training and for the new OOD repartitioning.","marker":"[31]"},{"why":"CMU-MOSEI, the larger standard dataset used for training and for the new OOD repartitioning.","marker":"[32]"}],"fun_headline_variants":["Twin distillation model tops rivals on missing, shifted emotions","37K-param causal module handles missing and drifted emotion data","Weight-sharing self-distillation masters random missing and OOD emotions","One compact model robust to missing modalities and distribution shifts","Causal inference distiller outperforms on missing and shifted emotions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The OOD results stand on the assumption that the newly repartitioned test sets genuinely shift word and label distributions away from training, yet the paper never describes how the repartitioning was done.","fun_headline_variants_meta":{"raw":{"variants":["Twin distillation model tops rivals on missing, shifted emotions","37K-param causal module handles missing and drifted emotion data","Weight-sharing self-distillation masters random missing and OOD emotions","One compact model robust to missing modalities and distribution shifts","Causal inference distiller outperforms on missing and shifted emotions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000312,"raw_usage":{"total_tokens":1797,"prompt_tokens":988,"completion_tokens":809,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":726}},"tokens_in":604,"tokens_out":809,"duration_ms":10398,"temperature":1.0,"reasoning_tokens":726,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:25:22.577298+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the unigram word distribution of the new OOD training and test splits; if their divergence is no larger than the divergence between the original IID training and test splits, then the 'OOD' condition is not a real distribution shift and the OOD comparison collapses.","supporting_citations":[{"cited_title":"Counterfactual reasoning for out-of-distribution multimodal sentiment analysis,","cited_arxiv_id":null,"evidence_quote":"Defines the word-based OOD multimodal sentiment task and supplies the CLUE baseline and the dataset split whose mixed-sentence flaw CIDer repartitions."},{"cited_title":"Transformer-based feature recon- struction network for robust multimodal sentiment analysis,","cited_arxiv_id":null,"evidence_quote":"Introduces the feature-reconstruction approach for incomplete multimodal data and the AUILC metric used to report missing-rate performance."},{"cited_title":"Efficient multimodal transformer with dual-level feature restoration for robust multimodal sentiment analysis,","cited_arxiv_id":null,"evidence_quote":"The EMT-DLFR baseline with dual-level feature restoration that CIDer compares against in RMFM scenarios."},{"cited_title":"General debiasing for multimodal sentiment analysis,","cited_arxiv_id":null,"evidence_quote":"GEAR, a model-agnostic debiasing baseline for OOD multimodal sentiment that CIDer compares with and improves on."},{"cited_title":"Noise imitation based adver- sarial training for robust multimodal sentiment analysis,","cited_arxiv_id":null,"evidence_quote":"Source of noise-imitation adversarial training and the TMFM and STMFM missing-scenario definitions used in the evaluations."},{"cited_title":"Towards robust multimodal sentiment analysis with incomplete data,","cited_arxiv_id":null,"evidence_quote":"LNLN, the strongest incomplete-data baseline, whose larger parameter count is repeatedly cited as the reason it sometimes leads."},{"cited_title":"Distribution-consistent modal recovering for incomplete multimodal learning,","cited_arxiv_id":null,"evidence_quote":"DiCMoR, a distribution-consistent modality-recovery baseline used for comparison under missing modalities."},{"cited_title":"Multimodal prompt learning with missing modalities for sentiment analysis and emotion recognition,","cited_arxiv_id":null,"evidence_quote":"MPLMM, the prompt-learning RMM baseline into which MACI is plugged to demonstrate model-agnostic portability."},{"cited_title":"Multimodal senti- ment intensity analysis in videos: Facial gestures and verbal messages,","cited_arxiv_id":null,"evidence_quote":"CMU-MOSI, one of the two standard datasets used for training and for the new OOD repartitioning."},{"cited_title":"Multimodal language analysis in the wild: Cmu-mosei dataset and in- terpretable dynamic fusion graph,","cited_arxiv_id":null,"evidence_quote":"CMU-MOSEI, the larger standard dataset used for training and for the new OOD repartitioning."}],"review_version":1}