TLDR: Foundation models struggle to reconcile conflicting information presented across different modalities (like text and images) or languages, often prioritizing one over another. This failure stems from “cross-modal attention imbalance,” where models disproportionately focus on certain modalities. A new method, “instance-level modality mixing,” which explicitly combines multiple modalities within individual training examples, significantly reduces this imbalance and improves the models’ ability to detect conflicts and perform better on complex multimodal tasks.
Foundation models (FMs) are becoming increasingly sophisticated, powering everything from autonomous web browsing to AI research assistants. These advanced AI systems are designed to integrate and reason over diverse information sources, including text, images, code, and structured data. However, a recent study reveals a significant challenge: FMs often struggle when faced with conflicting information presented across different modalities or languages.
Researchers at Carnegie Mellon University investigated how well FMs perform joint reasoning—simultaneously processing multiple modalities, especially when these modalities interact to form cross-modal context. They focused on scenarios called ‘cross-modal conflicts,’ where different pieces of evidence contradict each other. This setup allowed them to observe whether FMs prioritize one type of information over another or if they can reason jointly to reconcile the conflict.
The Problem: A Striking Gap in Reasoning
The findings were quite revealing. FMs are generally good at recognizing conflicts within a single modality (e.g., two conflicting text passages or two conflicting images), detecting them about 90% of the time. However, this performance drastically drops when the conflicting evidence is split across modalities—sometimes falling as low as 3%. Similar issues were observed in cross-lingual contexts, where models performed much better with monolingual information (e.g., English-English) than with multilingual information (e.g., English-Chinese).
This suggests that while FMs can handle individual data types well, their ability to integrate and reconcile information when it comes from different sources simultaneously is severely limited. This isn’t simply because models are weak in one particular modality; they can detect conflicts within images as easily as within text.
Uncovering the Root Cause: Attention Imbalance
The study traces this failure to a phenomenon called ‘cross-modal attention imbalance.’ This means that FMs exhibit extreme asymmetry in how they pay attention to different modalities, disproportionately prioritizing certain ones. For instance, in cross-lingual scenarios, English text often receives more attention than Chinese text. In multimodal settings, text tends to dominate over images.
Crucially, this attention imbalance isn’t resolved by simply feeding FMs more multimodal or multilingual datasets. The researchers found that blindly scaling up these datasets doesn’t help because they often lack training examples that explicitly require complex cross-modal reasoning.
A Causal Link and a Simple Solution
To confirm the causal relationship between attention imbalance and reasoning failures, the researchers manually adjusted attention scores. By increasing the attention given to the less-prioritized modality, they observed a significant improvement in conflict detection performance—up to 43% in cross-lingual settings and 18% in cross-modal settings. This demonstrated that balancing attention across modalities directly enhances the model’s ability to reason jointly.
Given that traditional dataset-level mixing (just having diverse data) doesn’t work, the team proposed a simple and scalable solution: ‘instance-level modality mixing.’ Instead of just including various modalities in the overall training data, this method explicitly combines multiple modalities within each individual training instance. For example, an input might include both a Chinese instruction and an English instruction, or a text instruction alongside an image and an image-related instruction.
Also Read:
- Unlocking Latent Reasoning in LLMs with Temperature Scaling
- MaskGRPO: A Unified Reinforcement Learning Approach for Multimodal Discrete Diffusion Models
Promising Results and Future Implications
Instance-level modality mixing proved highly effective. It significantly reduced cross-modal attention imbalance—by 4x in cross-lingual settings and 34% in cross-modal settings. This reduction directly translated to improved conflict detection, boosting performance by 37% in cross-lingual tasks and 2x in cross-modal tasks.
Furthermore, this approach also improved performance on several challenging vision-language benchmarks like HardBLINK, SAT, and MMMU, indicating that it enhances the overall reasoning capabilities of models, especially for tasks requiring non-trivial reasoning across different domains. The beauty of this method is that it doesn’t require costly new data curation; it repurposes existing datasets by mixing them at the instance level.
These findings highlight a fundamental gap in how FMs process information from different modalities. They underscore the critical need for training paradigms that better reflect the real-world complexity faced by AI systems, ensuring they can balance attention and reason effectively across diverse contexts. For more details, you can read the full research paper here.
