#model alignment

LLM Judges Show Systematic Bias Toward Own Outputs

New research quantifies how language models favor their own responses in evaluation tasks, threatening benchmark validity.