A new training technique nudges multimodal AI models to actually look at the parts of an image that matter, and it measurably improves their benchmark scores.
Researchers built a framework called Selective Probability Mass Concentration, or sPMC, that targets only the attention heads inside a multimodal model that already respond to visual grounding - typically just 3% to 15% of all heads in the model. During training, those heads are nudged to concentrate their attention on the image regions a segmentation model flags as relevant, while every other head is left alone. The team tested the approach across six benchmark suites and multiple MLLMs. It delivered an average zero-shot accuracy improvement of 3%, with gains on individual benchmarks reaching as high as 11.3%.
Most fixes for multimodal reasoning either retrain the whole model on reasoning-labeled data or bolt on extra inference-time steps, both of which cost real compute. This method instead makes a narrow, targeted change to the small subset of heads doing the actual visual-grounding work, leaving the rest of the model untouched. That's a cheaper lever, and it hints that a chunk of multimodal reasoning failures come from models not attending to the right pixels, rather than from a shortage of reasoning ability.
One caveat worth flagging: the paper measures benchmark accuracy, not hallucination rates, so any claim that this technique directly cuts hallucinations is getting ahead of the data.