acceptodds
Under review as a conference paper at ICLR 2027

CARE-KAN: Conflict-Aware Evidence Routing for Robust Multimodal Sentiment Analysis

Abstract

Multimodal sentiment analysis typically assumes that language, acoustic, and visual modalities provide mutually supportive, complementary evidence. In real-world affective computing, however, modalities frequently diverge due to nuanced emotional expressions, ambient noise, or sensory degradation. Existing methods struggle with this divergence: forcing cross-modal alignment suppresses subtle modality-unique nuances, whereas directly fusing raw conflict features introduces detrimental noise, triggering severe negative transfer on clean benchmark data. In this paper, we argue that multimodal conflict is an informative epistemic signal regarding sample difficulty and channel reliability, but it is intrinsically unsafe to exploit naively. To resolve this dilemma, we propose CARE-KAN, a conflict-aware evidence routing framework that shifts multimodal sentiment prediction from unconditional feature fusion to calibrated evidence aggregation. CARE-KAN operates through three coordinated stages: (1) Tripartite Evidence Decomposition, which factorizes multimodal features into shared consensus, modality-unique complements, and directional conflict vectors; (2) Reliability-Adaptive Routing, which assesses sample-conditioned modality reliability from the joint interaction of these evidence streams to dynamically modulate their predictive weights; and (3) Zero-Gated Residual Fusion, which incorporates the calibrated evidence via a zero-centered gate, safeguarding clean-set predictive power while selectively utilizing beneficial disagreement cues. The resulting representation is fed into a Kolmogorov–Arnold Network (KAN) prediction head to capture complex non-smooth residual decision surfaces. Evaluations on four benchmarks (CMU-MOSI, CMU-MOSEI, CH-SIMS, and IEMOCAP) show that CARE-KAN eliminates negative transfer on clean data while demonstrating superior resilience under acoustic/visual Gaussian noise and modality dropout. Mechanistic analyses further reveal that conflict magnitude reliably tracks prediction difficulty, and our router dynamically attenuates corrupted channels while preserving salient task-relevant evidence.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Sentiment Analysis, Multimodal Conflict, Evidence Routing, Robust Representation Learning, Kolmogorov--Arnold Networks (KAN), Negative Transfer Mitigation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_oHf2OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CARE-KAN, a conflict-aware framework for robust multimodal sentiment analysis. It decomposes unimodal representations into consensus, orthogonal complement, and directional conflict components, and introduces a reliability-adaptive router to modulate their contributions. A zero-initialized residual gating mechanism incorporates the resulting evidence into a base multimodal representation, followed by a Kolmogorov–Arnold Network prediction head. Experiments on CMU-MOSI, CMU-MOSEI, CH-SIMS, and IEMOCAP report competitive clean performance and improved robustness under visual noise and acoustic dropout, supported by ablation studies and supplementary analyses.

Critical strengths
  1. The paper addresses a meaningful problem by distinguishing potentially informative cross-modal disagreement from corruption-induced noise. The combination of structured evidence decomposition, reliability-adaptive routing, and gated residual fusion is reasonably motivated and provides a coherent framework for robust multimodal prediction.

  2. The paper reports competitive results across four benchmarks and provide empirical support for the effectiveness of the overall architecture.

Critical weaknesses
  1. In Eqs. (5)–(8), the estimated reliability scores attenuate the complement and conflict components within the CARE residual pathway. However, the original multimodal representation zbasez_{\mathrm{base}} is always retained and directly added to the final representation, meaning that corrupted modality features can still influence the prediction through the base pathway. Moreover, the subsequent LayerNorm and KAN head make it difficult to infer the actual influence of a modality from its reliability weight alone. Although Figure 3(b) shows that visual reliability decreases under increasing noise, this does not directly demonstrate that the final prediction becomes less dependent on corrupted visual information. Similarly, the theoretical result in Eq. (9) only establishes a limiting initialization property of the residual pathway, rather than guaranteeing the absence of negative transfer after training. I would like to see more direct analyses of how the estimated reliability affects the final prediction, such as modality intervention or prediction-sensitivity experiments under controlled corruption.

  2. The experimental evidence does not sufficiently validate the paper's central distinction between informative emotional conflict and corruption-induced disagreement. A major motivation of CARE-KAN is to preserve meaningful cross-modal disagreement while suppressing harmful sensory noise. However, the robustness evaluations in Section 4.3 primarily consider visual Gaussian noise and acoustic frame dropout, while Figure 3(a) relates the proposed conflict magnitude to prediction difficulty without independently establishing whether the disagreement is semantically informative. Consequently, the results support robustness to selected sensory corruptions but do not directly demonstrate the claimed ability to distinguish genuine affective conflict from noise. In addition, the comparisons omit recent relevant reliability-aware methods such as P-RMF (ACL 2025) [1], TF-Mamba (EMNLP 2025) [2], and EBMC (CVPR 2026) [3], which are closely related to the paper's reliability estimation and robust fusion objectives. More targeted conflict evaluations and comparisons under matched protocols would strengthen the central empirical claims.

Refs:

[1] Aoqiang Zhu et al. Proxy-Driven Robust Multimodal Sentiment Analysis with Incomplete Data. ACL 2025.

[2] Xiang Li et al. TF-Mamba: Text-enhanced Fusion Mamba with Missing Modalities for Robust Multimodal Sentiment Analysis. EMNLP 2025 Findings.

[3] Kang He et al. Enhance-then-Balance Modality Collaboration for Robust Multimodal Sentiment Analysis. CVPR 2026.

[4] Yitong Zhu et al. QA-MoE: Towards a Continuous Reliability Spectrum with Quality-Aware Mixture of Experts for Robust Multimodal Sentiment Analysis. ACL 2026.

Questions for authors
  1. Since zbasez_{\mathrm{base}} always retains all modality representations, how does the proposed reliability-aware routing mechanism prevent corrupted information from influencing the final prediction through the base pathway? Is there evidence that lowering a modality's reliability score actually reduces the final prediction's sensitivity to that modality?

  2. How does CARE-KAN distinguish semantically meaningful cross-modal conflict from noise-induced divergence? In particular, does the proposed conflict representation capture useful affective disagreement independently of general sample difficulty or corruption severity?

  3. The Zero-Gating analysis in Eq. (9) establishes a limiting property at initialization. What evidence supports the stronger claim that the gating mechanism prevents negative transfer after training? How does its performance compare with simpler residual scaling or conventional gated fusion under comparable conditions?

  4. I would like to see evaluations on independently identified affective conflicts, such as sarcasm or deliberately incongruent emotional expressions, alongside corrupted inputs. Such experiments could more directly assess whether CARE-KAN preserves informative disagreement while suppressing irrelevant noise.

I would be willing to raise my score if the authors can adequately address these concerns, particularly through more direct validation of conflict-aware routing.

6 comments
AuthorsOpenReview ↗

Response to Q1. Does lowering reliability reduce final-output sensitivity despite the retained base pathway?

Corrupted features remain in the base pathway, so a low internal reliability score alone is insufficient evidence. We therefore intervene in the same five trained MOSI checkpoints, without retraining, on all 686 test examples. We override visual reliability after the sigmoid, then recompute softmax weighting, residual assembly, LayerNorm, the gate, and KAN. Corruption and perturbation draws are matched across conditions. Each example receives five visual perturbations with RMS 0.01; S_v is the mean absolute prediction change divided by perturbation RMS. Overrides apply in both forward passes.

Inference conditionClean MAE ↓MAE at visual σ=1.0 ↓S_v at σ=1.0 ↓
Adaptive routing0.6868 ± 0.00310.7628 ± 0.00510.079 ± 0.004
Force visual reliability to 10.6892 ± 0.00340.9352 ± 0.00890.412 ± 0.015
Force visual reliability to 00.7164 ± 0.00390.7562 ± 0.00440.052 ± 0.003
Disable residual in the same checkpoint0.6985 ± 0.00320.8251 ± 0.00630.218 ± 0.008

Adaptive routing reduces final-output sensitivity by approximately 80.8% relative to forced high reliability, while improving noisy MAE from 0.9352 to 0.7628. Disabling the residual worsens both sensitivity and noisy MAE. Fixed-zero reliability performs slightly better under strong noise, but worsens clean MAE to 0.7164, showing the cost of suppressing useful visual evidence.

We also sweep the forced reliability at a separate, fixed noise level, σ=0.5:

Forced visual reliability r_vS_v ↓MAE ↓
0.000.061 ± 0.0030.724 ± 0.004
0.250.098 ± 0.0050.738 ± 0.005
0.500.165 ± 0.0080.769 ± 0.006
0.750.248 ± 0.0110.801 ± 0.007
1.000.324 ± 0.0130.825 ± 0.007

Lower reliability reduces both local sensitivity and MAE throughout this tested sweep. These are measurements of the complete prediction, including downstream normalization and KAN. They support reduced predictive influence, not complete modality isolation or algebraic cancellation: the base, softmax-weighted consensus, and cross-modal features can still depend on visual inputs even when r_v=0.

AuthorsOpenReview ↗

Response to Q2. Does the conflict representation capture useful disagreement beyond difficulty or corruption?

The router uses consensus, complement, and conflict jointly; conflict magnitude alone is not defined as unreliability. To test predictive utility independently of model-defined conflict scores, we use CH-SIMS's human unimodal sentiment labels to identify 61 natural-conflict examples. All predictions are evaluated against the overall multimodal sentiment label. No-Conflict removes d_m while retaining consensus and complement; Raw Difference replaces the structured conflict representation. Noisy inputs use visual σ=0.5; classification metrics are percentages.

VariantClean MAE ↓Clean F1 ↑Clean Acc-3 ↑Noisy MAE ↓
DecAlign0.50478.181.10.595
No-Conflict0.46580.183.60.528
Raw Difference0.47279.583.20.552
Full CARE-KAN0.43682.585.20.482

On the same conflict examples, Full improves clean/noisy MAE over No-Conflict by 0.029/0.046 and also outperforms raw differences. Thus, the stream provides predictive utility on annotation-defined incongruity, including clean inputs where no synthetic corruption is introduced. Because d_m is derived from joint features, this is a representation benefit, not a claim of newly created or statistically independent information.

We separately examine the difficulty confound on MOSI. Difficulty tiers use Base prediction error; low/high conflict groups use the learned conflict norm. ΔMAE is CARE-KAN minus Base.

DifficultyConflict groupNBase MAECARE-KAN MAEΔMAE ↓
EasyLow1750.3120.310−0.002
EasyHigh540.3460.332−0.014
MediumLow1450.6650.655−0.010
MediumHigh830.7200.648−0.072
HardLow881.1551.132−0.023
HardHigh1411.2421.068−0.174
OverallAll6860.73800.6868−0.0512

High-conflict groups show larger improvements within each tier, particularly medium/hard examples. However, their baseline errors also remain higher within each tier. We explicitly acknowledge this residual confounding: the diagnostic does not establish independence from every difficulty factor. The independently labelled evaluation above and the controlled corruption comparison under Q4 provide complementary evidence.

AuthorsOpenReview ↗

Response to Q3. What supports gating after training compared with simpler residual scaling or conventional gates?

Eq. (9) is an initialization-limit result for the task-loss gradient under bounded quantities. Our finite initialization gives α≈0.00474, not zero; neither auxiliary-loss gradients nor subsequent optimization are covered. We have removed the interpretation of this result as a guarantee against negative transfer after training.

We retrain matched MOSI controls, retaining C3D, RAER, KAN, and the same training protocol in all CARE variants. Learned gates have upper bound 0.1. The fixed scale κ=0.025 is validation-selected; the global gate is 0.1·sigmoid(b), initialized at b=−3. Base is trained separately, unlike the inference-time residual-off intervention in Q1.

Gate settingClean MAE ↓MAE, visual σ=0.5 ↓MAE, visual σ=1.0 ↓
Base, no residual0.7380 ± 0.00520.8360 ± 0.00710.9200 ± 0.0094
Fixed κ=0.0250.7185 ± 0.00450.7852 ± 0.00620.8421 ± 0.0081
Learned global gate0.7062 ± 0.00410.7624 ± 0.00550.8115 ± 0.0075
Sample gate, b₀=00.7015 ± 0.00480.7482 ± 0.00510.7894 ± 0.0068
Full sample gate, b₀=−30.6868 ± 0.00310.7280 ± 0.00420.7628 ± 0.0051

Full improves clean MAE over fixed scaling, a global gate, and ordinary sample-gate initialization by 0.0317, 0.0194, and 0.0147, respectively, and has lower noisy MAE in all three comparisons. Training checkpoints also show a smaller initial residual-to-base norm ratio: 0.0058 versus 0.0682 at epoch 1. At epoch 50, the ratios are 0.0381 versus 0.0395, while validation MAE is 0.684 versus 0.704. This supports near-zero initialization followed by useful residual learning. The evidence concerns measured training behavior and final performance; residual norms do not establish gradient insulation or universal absence of negative transfer.

Reviewer_m5J7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors investigate the challenge of multimodal sentiment analysis under cross-modal disagreement, particularly when discrepancies arise from meaningful emotional incongruity or sensory corruption. To address this challenge, they propose CARE-KAN, a conflict-aware evidence routing framework based on an asymmetric dual-pathway architecture. The framework consists of three key components: Tripartite Evidence Decomposition (C3D), which separates multimodal representations into consensus, modality-specific complementary information, and directional conflict; Reliability-Adaptive Evidence Routing (RAER), which dynamically estimates modality reliability and adjusts evidence contributions; and Zero-Gated Residual Fusion, which selectively incorporates conflict-aware evidence into the primary representation. A Kolmogorov–Arnold Network (KAN) prediction head is further employed to capture complex nonlinear relationships.

The authors conduct experiments on four multimodal affective computing benchmarks: CMU-MOSI, CMU-MOSEI, CH-SIMS, and IEMOCAP. The reported results demonstrate improvements over competitive baselines on several evaluation metrics, along with enhanced robustness under visual Gaussian noise and acoustic frame dropout.

Critical strengths

1.The proposed tripartite evidence decomposition extends conventional shared-private representations by explicitly modeling directional conflict and introducing geometric constraints to better characterize multimodal disagreement.

2.The integration of reliability-adaptive routing and zero-gated residual fusion provides an interpretable mechanism for selectively incorporating conflicting evidence while limiting its initial interference with the primary representation.

Critical weaknesses

1.The theoretical analysis does not establish the stability of the primary pathway throughout training, particularly since corrupted modality features can still directly influence the final prediction.

2.Despite encouraging empirical evidence, the reliability mechanism's ability to consistently distinguish meaningful emotional disagreement from noise-induced corruption remains insufficiently established under complex conflict conditions.

3.Several core ideas, including explicit conflict modeling and adaptive reliability-guided fusion, have been explored in prior work, and the distinctive methodological contribution of CARE-KAN remains insufficiently articulated.

Questions for authors

1.Could the authors clarify to what extent the theoretical analysis supports the stability of the primary representation beyond initialization?

2.Could the authors explain how the dual-pathway architecture mitigates the influence of corrupted modality information on the primary prediction, given that reliability filtering is applied only to the auxiliary pathway?

3.Could the authors further explain how the reliability mechanism distinguishes meaningful emotional disagreement from sensor corruption, particularly when both occur simultaneously?

4.Could the authors clarify the theoretical advantages of the proposed tripartite decomposition and reliability-adaptive routing over conventional shared-private representations and attention-based fusion mechanisms?

4 comments
AuthorsOpenReview ↗

Response to Q1. What does the theoretical analysis establish beyond initialization?

Eq. (9) establishes a task-loss gradient limit as the gate bias tends to negative infinity, under bounded features, residuals, and Jacobians. It does not guarantee stability throughout training or absence of negative transfer afterward. The finite initialization gives α≈0.00474, not zero; auxiliary-loss gradients are outside this result. We have narrowed the theoretical claims accordingly.

To assess trained models, we retrain the following MOSI controls (N=686 test examples). CARE variants retain C3D, RAER, KAN, and the same training protocol. Learned gates have upper bound 0.1. The fixed scale κ=0.025 is selected on validation data; the global gate is 0.1·sigmoid(b), initialized at b=−3. Base is trained separately, unlike the inference-time residual-off control under Q2.

Gate settingClean MAE ↓MAE, visual σ=0.5 ↓MAE, visual σ=1.0 ↓
Base, no residual0.7380 ± 0.00520.8360 ± 0.00710.9200 ± 0.0094
Fixed κ=0.0250.7185 ± 0.00450.7852 ± 0.00620.8421 ± 0.0081
Learned global gate0.7062 ± 0.00410.7624 ± 0.00550.8115 ± 0.0075
Sample gate, b₀=00.7015 ± 0.00480.7482 ± 0.00510.7894 ± 0.0068
Full sample gate, b₀=−30.6868 ± 0.00310.7280 ± 0.00420.7628 ± 0.0051

Full reduces clean MAE by 0.0317 versus fixed scaling, 0.0194 versus a global gate, and 0.0147 versus ordinary sample-gate initialization. The reported validation checkpoints also show how the residual contribution changes. Here ρ is the per-example ratio ∥αeCARE∥2/(∥zbase∥2+ϵ)\|\alpha e_{\mathrm{CARE}}\|_2/(\|z_{\mathrm{base}}\|_2+\epsilon), averaged across examples; entries are five-seed means.

EpochBase val MAEb₀=0 val MAEb₀=−3 val MAEMean ρ, b₀=0Mean ρ, b₀=−3
10.7920.8450.7950.06820.0058
150.7450.7490.7320.04650.0285
500.7350.7040.6840.03950.0381

Near-zero initialization has a smaller early residual contribution and later uses a nonzero residual while reaching lower validation MAE. These are empirical observations supporting the chosen training design; residual norms do not measure gradient insulation or extend Eq. (9) into a training-wide theorem.

AuthorsOpenReview ↗

Response to Q2. How can routing reduce corruption effects when the base pathway remains unfiltered?

We agree that corrupted features remain in the base. Furthermore, zero reliability does not remove softmax-weighted consensus or every indirect dependency on the visual modality. The relevant question is therefore whether routing changes the complete prediction, which we now measure directly.

On all 686 MOSI test examples, we use the same five trained checkpoints and matched corruption/perturbation draws, without retraining. Reliability is overridden after the sigmoid, and downstream softmax weighting, residual assembly, LayerNorm, the gate, and KAN are recomputed. Each example receives five visual perturbations with RMS 0.01. S_v is the average absolute prediction change divided by perturbation RMS; overrides are applied in both forward passes.

Inference conditionClean MAE ↓MAE at visual σ=1.0 ↓S_v at σ=1.0 ↓
Adaptive routing0.6868 ± 0.00310.7628 ± 0.00510.079 ± 0.004
Force visual reliability to 10.6892 ± 0.00340.9352 ± 0.00890.412 ± 0.015
Force visual reliability to 00.7164 ± 0.00390.7562 ± 0.00440.052 ± 0.003
Disable residual in the same checkpoint0.6985 ± 0.00320.8251 ± 0.00630.218 ± 0.008

Adaptive routing reduces sensitivity by approximately 80.8% relative to forced high reliability. Disabling the residual also worsens sensitivity and noisy MAE, supporting a functional residual contribution after all downstream transformations. Fixed-zero reliability has slightly better noisy MAE but substantially worse clean MAE. We report this trade-off and describe reduced predictive sensitivity, rather than complete isolation or algebraic cancellation of base-path noise.

AuthorsOpenReview ↗

Response to Q3. What happens when meaningful disagreement and sensor corruption coexist?

We use CH-SIMS's human modality-specific sentiment labels to define congruent (251), natural-conflict (61), and ambiguous (145) examples independently of model conflict norms and prediction errors. The overall sentiment label remains the evaluation target. The following table compares Full on the same annotation-defined examples before and after visual Gaussian noise, σ=0.5.

Group and inputNMAE ↓Mean visual reliability r_vMean gate α
Congruent, clean2510.3620.920.007
Natural conflict, clean610.4360.890.078
Congruent, noisy2510.3850.040.008
Natural conflict, noisy610.4820.050.045

Clean natural conflict retains high visual reliability and a larger gate; adding noise to those same examples lowers visual reliability. On the noisy conflict subset, Full's MAE of 0.482 compares with 0.528 for No-Conflict, 0.565 without RAER, and 0.595 for DecAlign. Thus, the evaluation jointly tests affective incongruity and controlled corruption, with prediction evidence beyond the internal scores. A separate MOSI test with visual σ=0.5 and acoustic dropout p=0.5 gives MAE/degradation 0.765/0.078 versus DecAlign's 0.902/0.167.

The router conditions on consensus, complement, and conflict jointly; conflict alone is not defined as unreliability. These results support the tested behavior without guaranteeing separation of every semantic and sensory cause. The human labels identify affective incongruity, not independently verified sarcasm intent.

AuthorsOpenReview ↗

Response to Q4. What distinguishes C3D and RAER from shared-private representations and attention?

The distinction is an explicit geometric and relational interface, not a claim that attention cannot express disagreement. C3D constructs a per-example complement near-orthogonal to its consensus vector, and a directional conflict feature using displacement and angular disagreement relative to the other modalities' consensus center. RAER uses the joint triplet to condition residual evidence. Near-orthogonality is a geometric property, not statistical independence of all three streams or a general expressiveness theorem.

The following CH-SIMS natural-conflict ablation (N=61; noisy visual σ=0.5) tests the representation choice while retaining the consensus/complement evidence in the No-Conflict control.

Conflict representationClean MAE ↓Noisy MAE ↓
No-Conflict, d_m=00.4650.528
Raw Difference0.4720.552
Full C3D0.4360.482

Together with the no-router result under Q3 and matched gate controls under Q1, these comparisons provide component-specific evidence for the tested integration. Reliability weighting and residual gating have precedents; our revision acknowledges shared-private, attention-based, uncertainty-based, and consistency-aware methods. We do not claim universal theoretical superiority, and retain residual difficulty confounding as a limitation of the separate conflict-norm diagnostic.

Reviewer_6pSPOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a conflict-aware evidence routing framework for robust multimodal sentiment analysis, called CARE-KAN. In particular, the proposed method consists of three components: Tripartite Evidence Decomposition, Reliability-Adaptive Routing, and Zero-Gated Residual Fusion. The first component separates multimodal representations into shared consensus, modality-specific complements, and conflict information. The second component estimates modality reliability to dynamically adjust the contribution of different modalities, while the third component introduces gated residual fusion to reduce negative transfer. Finally, a KAN-based prediction head is adopted for sentiment prediction. Experiments on 4 benchmarks are conducted to demonstrate the effectiveness and robustness of the proposed method. In general, I believe this work is interesting, but I still have some questions for the authors for clarification.

Critical strengths
  1. The motivation is intuitive and interesting. Instead of simply aligning different modalities, this work considers multimodal conflict as potentially useful information for sentiment prediction, enriching the design space for multimodal representation learning.

  2. The proposed framework combines representation decomposition, modality reliability estimation, and gated fusion. The overall design is reasonable, especially for handling unreliable audio-visual information.

  3. The main results seem promising. The proposed method achieves competitive performance on 4 popular benchmarks, and the robustness experiments under noise and modality dropout further demonstrate its effectiveness. Besides, the authors provide several ablation studies to support the proposed design.

Critical weaknesses
  1. This paper argues that multimodal conflict can provide useful information about modality reliability. However, I wonder whether conflict necessarily indicates unreliable information. For example, sarcasm may naturally introduce disagreement between text and audio, even when both modalities are reliable. How does the proposed method distinguish meaningful affective conflict from sensor corruption? Besides, is there any evidence that the conflict representation contains distinct information rather than redundant features? I think this is important for supporting the main contribution of this work.

  2. Some highly related literature seems to be missing, especially ConsMSA [1], which explores intra- and inter-modal semantic consistency for multimodal sentiment analysis. Although the architectures are different, both works investigate semantic relationships across modalities. I wonder why this work explicitly separates conflict information instead of modeling semantic consistency as in ConsMSA. What are the fundamental differences and advantages of these two designs? I think a more detailed discussion is necessary to better clarify the motivation and novelty of this work.

[1] ConsMSA: Semantic Distribution Consistency Learning for Multimodal Sentiment Analysis, ICML 2026.

  1. Based on Tables 1 and 2, the experimental results are promising, but some conclusions need further clarification. For example, CARE-KAN achieves better classification performance on MOSEI, while its MAE (0.553) is worse than DecAlign (0.543). Besides, the robustness experiments are mainly conducted on MOSI with visual Gaussian noise and acoustic frame dropout. Would the improvements remain consistent under other corruption types or simultaneous modality degradation? I think a more comprehensive evaluation would strengthen the conclusions.

  2. The contribution of the KAN prediction head needs further discussion. Based on Table 3, replacing KAN with an MLP leads to performance degradation. However, does this improvement mainly come from the representation capacity or the specific B-spline design? Although the authors provide a parameter-matched MLP comparison, I wonder whether other nonlinear prediction heads could achieve similar performance. Besides, the proposed framework introduces additional architectural complexity. A more detailed analysis of the performance-efficiency trade-off would be helpful.

Questions for authors
  1. What are the fundamental differences between conflict-aware representation learning and the consistency-aware strategy in ConsMSA? Why is explicit conflict decomposition necessary?

  2. How does the reliability router distinguish meaningful affective disagreement, such as sarcasm, from corrupted sensory information?

  3. Is there any quantitative evidence that the conflict representation contains information distinct from the consensus and complement representations?

  4. Why does CARE-KAN improve classification performance but achieve worse MAE than DecAlign on MOSEI?

  5. Can the proposed method maintain its robustness when multiple modalities are corrupted simultaneously?

  6. Is the KAN prediction head necessary, or could other nonlinear prediction heads achieve comparable performance with similar computational budgets?

13 comments
AuthorsOpenReview ↗

Response to Q1: How does explicit conflict decomposition differ from ConsMSA, and why use it?

ConsMSA combines distribution consistency with predictive relevance to regularize, reweight, and prune tokens. We have corrected the discussion to acknowledge that this is not unconditional suppression of inconsistent evidence. CARE-KAN instead constructs an explicit relation to the other modalities: leave-one-out displacement and angular disagreement, alongside consensus and complement. The router conditions residual modulation on all three streams. We also retain consensus alignment, so consistency learning and explicit conflict representation are compatible.

The justification is empirical within this architecture: on 61 independently identified CH-SIMS conflict examples, retaining the conflict stream reduces clean MAE from 0.465 to 0.436. The complete ablation is included under Q3. This supports our representation choice, without claiming universal necessity or superiority over ConsMSA under unmatched protocols.

AuthorsOpenReview ↗

Response to Q2: How does the router distinguish meaningful affective disagreement from corruption?

Reliability is learned from the joint consensus, complement, and conflict features, rather than treating conflict magnitude alone as unreliability. We evaluate this distinction using CH-SIMS's existing human unimodal sentiment annotations. These define congruent (251), natural-conflict (61), and ambiguous (145) examples independently of our conflict norm or prediction error. The overall multimodal label remains the prediction target, and subgroup assignments remain fixed after visual corruption.

The following results focus on the congruent and natural-conflict groups. Noisy inputs use visual Gaussian noise with σ=0.5; routing and gate columns are means for Full CARE-KAN.

Group and inputNMAE ↓Mean visual reliability r_vMean gate α
Congruent, clean2510.3620.920.007
Natural conflict, clean610.4360.890.078
Congruent, noisy2510.3850.040.008
Natural conflict, noisy610.4820.050.045

Natural conflict therefore coexists with high visual reliability and a larger gate on clean inputs. Corrupting those same examples lowers visual reliability while retaining useful prediction performance. A separate MOSI inference intervention at visual σ=1.0 gives final-output sensitivity/MAE of 0.079/0.7628 with adaptive routing versus 0.412/0.9352 when r_v is forced to 1. Sensitivity is the mean absolute prediction change divided by perturbation RMS (0.01; five perturbations per example), with downstream computations recomputed in the same checkpoints. The subgroup evidence concerns annotation-defined affective incongruity; it does not establish verified sarcasm recognition or perfect identification of physical sensor quality.

AuthorsOpenReview ↗

Response to Q3: Does the conflict representation add useful information beyond consensus and complement?

We added a No-Conflict control that sets d_m=0 while retaining consensus and complement, and a separate Raw Difference control. All results below use the same CH-SIMS natural-conflict subset (N=61); noisy inputs use visual σ=0.5. Classification metrics are percentages.

VariantClean MAE ↓Clean F1 ↑Clean Acc-3 ↑Noisy MAE ↓
DecAlign0.50478.181.10.595
No-Conflict0.46580.183.60.528
Raw Difference0.47279.583.20.552
Full CARE-KAN0.43682.585.20.482

Relative to No-Conflict, Full improves clean/noisy MAE by 0.029/0.046 and F1/Acc-3 by 2.4/1.6 percentage points. Raw differences also trail the structured representation. This demonstrates additional predictive utility from the explicit stream within the tested architecture. Because d_m is computed from joint consensus features, it does not create information absent from those inputs; we have removed claims of statistical independence or universal nonredundancy.

Comments on the paper

1 comment

Dear Reviewers,

The rebuttal deadline is approaching. I am leaving this friendly reminder to encourage active discussion. Thank you for your valuable time and engagement!

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.