BJBrian Jalaian
All news

· Video

SCoOP explained in about a minute

"SCoOP: Semantic Consistent Opinion Pooling for Uncertainty Quantification in Multiple Vision-Language Model Systems", led by my Ph.D. student Chung-En Johnny Yu with Nathaniel D. Bastian, was a poster at the ICLR 2026 Workshop on Agentic AI in the Wild: From Hallucinations to Reliable Autonomy. This video explains it in 68 seconds.

When several vision-language models work together, combining their answers the wrong way can amplify errors and spread hallucinations. SCoOP is training-free. It treats each model as an expert, samples several answers from each, and weights each model by how consistent its answers are (inverse entropy). It then pools the weighted opinions and uses the uncertainty of the pooled vote as a system-level score for detecting hallucinations and for abstaining. On ScienceQA, SCoOP detects hallucinations with an AUROC of 0.866, against 0.732 to 0.757 for majority voting and naive selection, and its aggregation step costs only microseconds more than those baselines.

The apple question in the first part of the video is a made-up illustration, not a result from the paper; the bar chart at the end shows the paper's numbers. The narration voice and music are AI-generated.

For the full write-up, see When a Team of AI Models Is Unsure.