Built a panel of models to judge each other. It ran for a week. Dashboard was green. Then I read what was actually winning. Not the sharpest answers. The ones that sounded like the judges. A popularity contest wearing a rubric. The fix: take away the name tags. Karpathy had already shipped it. I just had to admit what I had built.
https://praveenlavu.com/dispatch/anonymized-peer-review
#llm #evaluation #machinelearning #aiengineering #buildinpublic
LLM Self-Preference Bias: How Anonymized Peer Review Fixes It · Praveen Lavu
I wired a multi-model judge panel to pick the best output, and it spent a week quietly voting for its own architecture family. Not the best answer. The one that sounded like the judge. Here is the night I caught it, and the one-line fix that a researcher named Karpathy had already shipped.







