The cross-attention architecture bolts onto existing frontier models.
Visual language models (VLMs) are surprisingly bad at comparative visual reasoning - detect the difference type tasks needed in medicine and science.
We just made VLMs stateful by post-training cross attention between visual encoder layers.
Our approach can be bolted on existing frontier models.
Users praised stateful visual encoders for VLMs as practical progress because they address a known weakness in comparison and change-detection tasks.
A new Stateful Visual Encoder proves valuable across a range of domans: check out Colin's post below!
Change-aware vision turns VLMs from caption readers into audit engines.
The mechanism matters: if each image is compressed separately, the LM compares two lossy summaries. Small deltas, layout shifts, defects, UI state changes, and tampering cues get buried before reasoning starts.
Push the comparison into the encoder, and the model can preserve the difference field as evidence instead of reconstructing it from text tokens. That changes the market for claims review, factory QA, medical follow-ups, satellite monitoring, and agent UI control.
The clean check: false negatives on small visual deltas, tokens per comparison, and latency per verified change.
@profjoeyg This is the kind of practical AI work I want to see more of. VLMs being bad at detect-the-difference tasks has been obvious for a while. Nice to see a concrete fix instead of more AI hype.
@profjoeyg Fascinating direction for improving VLM reasoning.
@profjoeyg CAN THIS WORK FOR VISUAL DOCUMENT UNDERSTANDING AS WELL?