Inductive Visual Logic for Few-Shot Out-Of-Distribution Adaptation in VLMs

Generative vision-language models (VLMs) such as Qwen-VL and LLaVA achieve strong zero-shot performance on tasks overlapping with their pretraining distribution, yet fail on specialized domains where the required discriminative features were never learned, a regime we term distant out-of-distribution (OOD). Standard adaptation methods cannot overcome this representational absence because they operate within the encoder’s existing feature space. However, VLMs retain a robust descriptive capacity even when discrimination collapses: a model that cannot classify a medical scan can still articulate its visual patterns. Exploiting this asymmetry, we introduce Inductive Visual Logic (IVL), a training-free framework that constructs classification knowledge from the model’s surviving descriptive ability. IVL extracts visual traits from few-shot support images through dual-mode prompting, combining semantic descriptions with primitive visual observations, and organizes them into per-class trait dictionaries. At inference, hierarchical filtering identifies spatially grounded trait evidence for classification. Across multiple distant-OOD benchmarks, IVL achieves the highest aggregate accuracy under two VLM backbones while producing interpretable, trait-traceable predictions.

@inproceedings{inductive-visual-logic-2026,
  title     = {Inductive Visual Logic for Few-Shot Out-Of-Distribution Adaptation in VLMs},
  author    = {Hung-Jen Chen and Yu-Heng Ho and Ting-Yao Huang and Po-Hsiang Hsu and Li-Yu Chen and Chun-Yi Lee and Min Sun},
  booktitle = {European Conference on Computer Vision (ECCV) 2026},
  year      = {2026}
}