Why is Mark using the Apple Magic Mouse?

Can searching through more options actually make an AI system worse? That’s the question my paper, “When More Search Can Worsen Deployment: Best-of-N Selection on an Incomplete Proxy,” investigates. The paper studies what happens when an AI system selects the “best” output using a score that captures only part of what we actually care about. As the number of candidates grows, selection becomes increasingly aggressive and the system can become extremely good at finding outputs that look good to the proxy while becoming worse on the underlying objective. Rather than just observing this failure mode, the work analyzes when and why increasing Best-of-N search causes deployment performance to improve, plateau, or deteriorate. The broader takeaway is that more compute and more search do not automatically produce better systems. When evaluation is incomplete, scaling the optimizer can also scale the gap between what is measured and what actually matters. Happy to share that this work was accepted to @NewInML at NeurIPS 2026. #NeurIPS2026 #MachineLearning

Is this what addicts feel like? 22.3 B tokens in the last week