Limits to scalable evaluation at the frontier: LLM as Judge won't beat twice the data

Factual evidence

What the source reports

Research demonstrates fundamental scaling limits and systematic bias in LLM-as-a-judge methods compared to traditional data annotation.