Increasing Computation Resolves Conflicts in Vision Language Models

A research team led by Bingyang Wang and nine collaborators has published a groundbreaking study demonstrating that Vision Language Models (VLMs) exhibit human-like cognitive control when resolving conflicts between visual and textual information. Their paper, 'Increasing Computation Resolves Conflicts in Vision Language Models,' systematically tested 47 different VLMs across 4,410 tasks spanning seven conflict paradigms including Stroop and Flanker tests. The key finding reveals that larger models with more parameters show significantly better conflict resolution capabilities, establishing parameter count as a direct proxy for cognitive control capacity in artificial systems.

The study's most striking discovery is that VLMs reproduce the fine-grained demand-resource relationship observed in human psychology: larger models actually drop below chance performance on high-conflict incongruent trials, while smaller models fail to engage meaningfully and perform at chance level. This mirrors human behavior at short processing times and suggests that adaptive flexibility under conflict emerges naturally through optimization dynamics in scaled neural networks. The research provides the first systematic evidence that human-like cognitive control can arise from pure scaling, with implications for developing more robust multimodal AI systems that can handle real-world ambiguity and conflicting information sources.