Encoded but Disconnected: Decomposing Vision-Language Model Failures under a Patching Null
The research tested if information encoded in model mid-layers is actually used to make final predictions across three different vision-language models. The finding is a universal negative: this mid-layer encoding, though often present in errors, does not causally drive the final output. Researchers decomposed failures into distinct modes and developed a routing framework that predicts which specific intervention fixes each failure type.