New AI Trick Lets Chatbots Identify Medical Images They've Never Seen

Some picture-reading AI models — the paper names Qwen-VL and LLaVA — can ace tasks that overlap with what they saw during pretraining, but they flop on specialized domains where the features they'd need were never learned. The researchers call this "distant out-of-distribution," and standard adaptation methods can't fix it, because those methods stay inside the encoder's existing feature space.

The trick the authors exploit: these models keep a strong descriptive ability even when their discrimination collapses. A model that can't classify a medical scan can still articulate its visual patterns. So they introduce Inductive Visual Logic (IVL), a training-free framework that builds classification knowledge out of that surviving descriptive skill. IVL pulls visual traits from a few-shot support images using dual-mode prompting — semantic descriptions plus primitive visual observations — and organizes them into per-class trait dictionaries. At inference, hierarchical filtering finds spatially grounded trait evidence.

Across multiple distant-OOD benchmarks, IVL lands the highest aggregate accuracy under two VLM backbones, and its predictions are interpretable and trait-traceable. The work was accepted to ECCV 2026.