Megadose The AI curve, compressed.

How VLMs Plan What's Next

· ArXiv · AI/CL/LG ·
Qwen3-VL tracks a cursor through an image to select the next object it has not yet named.

The researchers describe binding heads that link text to image regions and use previous answers to track progress. A selection step picks the nearest unwritten region after the cursor, while gaze heads retrieve its content for conversion into a word. The team validated the mechanism using causal patching on a synthetic visual retrieval task and tested alternative explanations. They report that it also holds across input formats, natural photographs and languages, with the relevant heads appearing in models from 2B to 32B parameters.

ArXiv · AI/CL/LG's note

score 5

Categories: Research