LAI #151: Can AI Forget Without Getting Worse?
Towards AI's ToneBench benchmark found measurable lab bias in LLM judges: on its October 26 board, GPT-6.1 Sol scored OpenAI drafts 4.1 points higher than the other LLM judges' scores would predict, while Claude Opus 5.5…