Multimodal models aren't actually 'thinking' in a unified way; they're just pretending to share parameters.
April 15, 2026
Original Paper
Pseudo-Unification: Entropy Probing Reveals Divergent Information Patterns in Unified Multimodal Models
arXiv · 2604.10949
The Takeaway
This study reveals 'pseudo-unification,' where vision and language follow completely different entropy trajectories even within the same model. Sharing parameters doesn't mean the model is integrating the two modalities into a single concept. This explains why multimodal models often fail at simple tasks that require true 'vision-text' fusion. For practitioners, this is a sign that 'unified' architectures are still deeply fragmented internally. We need to move beyond simple parameter sharing toward models that truly synchronize the 'entropy' of different data types.
From the abstract
Unified multimodal models (UMMs) were designed to combine the reasoning ability of large language models (LLMs) with the generation capability of vision models. In practice, however, this synergy remains elusive: UMMs fail to transfer LLM-like reasoning to image synthesis and exhibit divergent response behaviors. We term this phenomenon pseudo-unification. Diagnosing its internal causes is important, but existing probing methods either lack model-internal insight or ignore prompt-response depend