BJBrian Jalaian
All news

· Paper

Johnny Yu's ORCA paper is published by ACM

My Ph.D. student Chung-En Johnny Yu's paper "ORCA: An Agentic Reasoning Framework for Hallucination and Adversarial Robustness in Vision-Language Models" is now published in the Proceedings of the 2026 10th International Conference on Cloud and Big Data Computing (ICCBDC '26), ACM, pages 9 to 20. It is joint work with Nathaniel D. Bastian.

Vision-language models often report objects that are not in the image, and small adversarial changes to an image can push them further off. ORCA addresses this at inference time with an Observe, Reason, Critique, Act loop. The model asks a set of small vision tools (each under 3B parameters) targeted questions about the image, checks where their answers disagree, and revises its prediction. It needs no retraining and no access to model internals, and it keeps every reasoning step as a trace that can be audited. Although it was designed for object hallucination, ORCA also improves accuracy on adversarially perturbed images without adversarial training.

The video above is a 51-second explainer of how ORCA works. The scene illustrates the loop and is not a result from the paper. The narration voice and music are AI-generated.

Congratulations, Johnny.