A new distillation method teaches AI models to diagnose whether a wrong answer came from misreading an image or misapplying logic, then fixes only the reading problem.
Researchers describe Perception-Correction Distillation (PCD), a label-free technique for training smaller multimodal models to match larger ones. Standard on-policy distillation cannot tell whether a bad answer stemmed from faulty perception or faulty reasoning, and existing perception success metrics conflate the two. PCD combines two signals, a downstream answer failure and disagreement between teacher and student, multiplying them into a soft AND gate that only strengthens correction when both signals agree perception is the problem. The method also separates perception and reasoning rollouts and uses mean-preserving weights so the reasoning objective stays untouched. Across eight benchmarks, it lifted an 8B-to-2B distillation's macro average from 44.50 to 47.28, and a 32B-to-8B run from 56.94 to 61.22; removing the technique or the separated rollouts in matched tests cut held-out scores by 2.22 and 0.88 points.
The result matters because distillation is the standard way labs shrink expensive frontier models into cheaper ones that still perform well, and multimodal models fail in two very different ways, seeing wrong or thinking wrong, that most training pipelines currently treat identically. Targeting corrections at the actual point of failure, rather than applying a single blunt reward signal across the whole reasoning chain, is a more surgical approach than the field's usual methods.
Still, this is a benchmark-average story, not a deployed-product one, and the teacher-student disagreement signal only works if the teacher model is itself reliably right about what it saw.