TLDR: This paper introduces “modality sabotage,” a diagnostic failure mode in multimodal AI where a highly confident error from one data source (e.g., audio) overrides correct information from other sources, leading to an incorrect overall prediction. The authors propose a model-agnostic framework that treats each modality as an independent agent, allowing for the identification of “saboteurs” and “contributors” to a decision. Applied to emotion recognition, the framework reveals systematic reliability patterns and highlights recoverable uncertainty in multimodal large language models.
Multimodal Large Language Models (MLLMs) are rapidly advancing, combining information from various sources like text, audio, and vision to perform complex tasks. However, understanding how these models arrive at their decisions, especially when different data streams provide conflicting information, remains a significant challenge. Often, it’s unclear which modality influences a prediction most, how disagreements are resolved, or if one data type dominates the others.
Prior research has touched upon related issues such as ‘modality collapse,’ where models over-rely on text, or ‘unimodal bias,’ where one data stream consistently dominates across an entire dataset. This new research introduces a distinct diagnostic failure mode called modality sabotage. Unlike systematic biases, modality sabotage focuses on individual instances where a highly confident error from a single modality not only fails locally but actively overrides other correct evidence, leading the overall fused prediction astray.
Understanding Modality Sabotage
Imagine an MLLM trying to understand an emotional scene. If the audio component confidently misinterprets a sound as anger, even when visual cues clearly suggest happiness, and this audio error then forces the model to predict anger, that’s modality sabotage. It’s a specific, instance-level problem where an overconfident, incorrect unimodal input derails the final decision.
A Diagnostic Framework: Modalities as Agents
To analyze and diagnose such dynamics, researchers propose a lightweight, model-agnostic evaluation layer. This framework treats each modality—Text (T), Audio (A), and Vision (V)—as an independent ‘agent.’ A joint view (TAV) also acts as an agent. Each agent produces candidate labels for a given task (like emotion recognition) along with a confidence score and a brief self-assessment of data quality.
A simple fusion mechanism then aggregates these outputs. This process makes it transparent which modalities are ‘contributors’ (supporting correct outcomes) and which are ‘saboteurs’ (leading the model astray). The framework doesn’t require retraining or architectural changes to the MLLM, making it a plug-and-play diagnostic tool.
How the Fusion Works
For each sample, the Text, Audio, Vision, and joint view agents are queried with structured prompts. They return a ranked set of candidate labels with confidence scores (1-100) and a data-quality report. The confidence reflects the agent’s belief in its labels, while the data-quality report attempts to capture if the LLM can self-diagnose input issues (e.g., noisy audio, occluded faces). These confidences are then aggregated and normalized to produce a final ranked prediction.
The researchers distinguish between ‘potential sabotage’ (a high-confidence unimodal error) and ‘successful sabotage’ (when that error also causes the fused model to make the same wrong prediction). For diagnostic purposes, the focus is primarily on potential sabotage, as it provides a clearer upper bound on a modality’s tendency towards overconfident errors.
Recoverable Uncertainty with Top-k Reasoning
Modality sabotage can lead to a situation where the top prediction is wrong, but the correct label might still be ranked highly among other possibilities. To diagnose this, the framework uses ‘Top-k reasoning.’ This quantifies whether the true label is present within the top-k hypotheses ranked by the fused scores. This helps identify ‘recoverable uncertainty’—cases where the model’s internal ranking still preserves the correct hypothesis despite sabotage, suggesting that with better conflict resolution, the correct answer could be found.
Also Read:
- Enhancing Multimodal Sentiment Analysis with a Double Information Bottleneck
- Unlocking Reliable AI in Medical Imaging with Limited Data: A New Theoretical Framework
Case Study: Multimodal Emotion Recognition
The framework was applied to three widely used multimodal emotion recognition benchmarks: MER, MELD, and IEMOCAP. The findings revealed systematic reliability profiles across different datasets and model backbones (GPT-5-nano and GPT-4o-mini).
- Top-1 vs. Top-k: The fusion generally maintained baseline Top-1 accuracy while significantly improving Top-k coverage across datasets. This indicates that even when the primary prediction was incorrect due to sabotage, the correct answer was often still present among the leading options, highlighting recoverable uncertainty.
- Data Quality Weighting: Interestingly, weighting the fusion by self-reported data quality did not consistently improve accuracy, suggesting that while models can self-diagnose input issues, these signals are only weakly aligned with correctness.
- Modality Behavior: Across the benchmarks, audio was frequently identified as the primary saboteur, while text often acted as a strong contributor. This provides valuable insight into which components of a multimodal pipeline might need refinement.
- Dataset-Specific Profiles: The study also highlighted how dataset characteristics influence modality reliability. For example, MER suffered from noisy speech recognition but benefited from rich video cues, while MELD’s sitcom-style videos with exaggerated cues could sometimes mislead vision.
This research offers a valuable diagnostic scaffold for multimodal reasoning, supporting principled auditing of how different data streams interact and informing potential interventions to improve the reliability and interpretability of MLLMs. For more details, you can read the full paper here.
