Multimodal Reasoning: Diagnostic Layer Exposes How One Modality Sabotages Fused Results and Misleads Predictions

Multimodal artificial intelligence systems, which integrate information from multiple sources like text and images, are rapidly advancing, yet understanding how they reach conclusions remains a significant challenge. Chenyu Zhang from Harvard University, alongside Minsol Kim, Shohreh Ghorbani, and colleagues at the MIT Media Lab, investigate a critical flaw in these systems termed ‘modality sabotage’, where a confidently incorrect input from one source overwhelms accurate information from others. The team developed a novel diagnostic layer that treats each input modality as an independent agent, allowing researchers to trace which sources contribute to correct answers and, crucially, which actively mislead the system. This approach reveals systematic reliability profiles within established multimodal emotion recognition benchmarks, offering insight into whether failures stem from inherent limitations of the technology or from issues within the training data, and provides a powerful framework for auditing and improving the reasoning processes of multimodal AI.

Unimodal Biases in Multimodal Emotion Recognition

This research investigates the robustness of multimodal emotion recognition (MER) systems, particularly those leveraging Large Language Models (LLMs). The authors propose a method to analyze and improve the reliability of these systems by identifying and mitigating unimodal biases, situations where a single modality, such as text, audio, or video, unduly influences the emotion prediction, potentially leading to inaccurate results. Key findings demonstrate that MER systems are susceptible to strong biases from individual modalities, where a dominant modality can overshadow information from others, even when that modality contains misleading cues. The authors introduce a method to quantify the influence of each modality on the final emotion prediction, allowing them to pinpoint potential biases.

They also explore techniques to reduce these biases and improve overall robustness, including data augmentation, modality weighting, and adversarial training. The proposed methods were evaluated on benchmark MER datasets, CMU-MOSEI, IEMOCAP, and MER2023, demonstrating improvements in robustness and accuracy. The research highlights the potential of LLMs to enhance MER systems, but also emphasizes the need to address unimodal biases when integrating these powerful models, as LLMs can be susceptible to biases present in their training data. Overall, this research contributes to the development of more reliable and trustworthy MER systems, paving the way for more accurate and robust emotion recognition with important applications in human-computer interaction, mental health monitoring, and affective computing.

Modality Sabotage in Multimodal AI Systems

This work introduces a new framework for understanding how multimodal artificial intelligence systems arrive at decisions. Researchers developed a method to analyze individual contributions from each modality, identifying instances where a confident but incorrect input from one source can override accurate information from others, a phenomenon termed “modality sabotage. ” The core of the approach treats each modality as an independent “agent” that proposes labels with associated confidence scores and self-reported data quality assessments. The team evaluated this framework across three widely used multimodal emotion recognition benchmarks, MER, MELD, and IEMOCAP, using inputs derived from text transcripts, audio analysis, and video processing.

Audio was analyzed to extract prosodic features and voice quality, while video utilized facial action unit detection and visual language models to objectively describe observable cues. Each agent outputs a ranked list of potential labels with confidence scores, alongside a data quality report. The system aggregates these outputs to arrive at a final prediction, enabling researchers to pinpoint which modalities contributed to correct answers and which acted as “saboteurs. ” Results demonstrate the ability to identify instances of potential modality sabotage, defined as cases where a modality exhibits high confidence in an incorrect label and influences the final fused prediction. The framework’s ability to pinpoint these instances offers a pathway toward gating or weighting modalities to mitigate the impact of misleading inputs and enhance the robustness of multimodal AI systems.

Modality Sabotage and Multimodal Decision Auditing

This work presents a new diagnostic framework for understanding how multimodal machine learning models arrive at their decisions. Researchers identified a specific failure mode termed ‘modality sabotage’, where a confidently incorrect prediction from one input stream can override other, correct evidence and mislead the overall result. To address this, they developed a method for evaluating each modality’s contribution to a prediction, effectively treating each as an independent agent and auditing its reliability. Applying this framework to emotion recognition tasks revealed systematic patterns in how different modalities perform across various datasets. The analysis demonstrated that the framework can expose recoverable uncertainty and highlight which modalities are less reliable in specific contexts, aligning with known characteristics of each dataset. This work offers a general scaffold for auditing multimodal reasoning systems and guiding future improvements in areas such as calibration, conflict resolution, and interpretable fusion.