// 01 / About this session
Most vision systems look at an image or video once and commit to an answer. When they fail, it is usually because that answer was never fully grounded in the visual input. Andrew shows how to design systems that go back for more evidence before they respond, drawing on lessons from building Glimpse, the visual reasoning model his team at Elorian recently launched.
You will learn how to audit a model for shortcuts: remove an input and see if the answer changes, make small counterfactual edits that should flip a result, and add distractors to test whether the model is reasoning or estimating. Andrew demonstrates these techniques with DecoyCount, Elorian’s counting benchmark, then shows why models that revisit their inputs, by zooming into part of an image or pulling key frames from a video, give more accurate answers.
You will leave with a checklist for auditing any model or agent you use and a set of design patterns for adding a second look where it counts.
What you’ll learn
Auditing a model for shortcuts with ablations, counterfactual edits and distractors
Why models that revisit their inputs give more accurate answers
A checklist and design patterns for adding a second look
Skills you’ll practice
Evaluation-driven development
Machine learning foundations
// 02 / Speakers

Andrew Dai
Co-founder and CEO, Elorian
Spent over a decade at Google Brain and DeepMind as a core area lead on Gemini, then co-founded Elorian to build AI that thinks visually.
AI DEV · BY DEEPLEARNING.AI
Join us at AI Dev
Meet the engineers shipping AI to production, and leave ready to do it yourself.