January 2026 Summaries

The text criticizes LMArena, an AI evaluation platform, for promoting superficial attributes over genuine intelligence and accuracy in AI models. Researchers express frustration with LMArena's emphasis on presentation and engagement metrics, which incentivizes tweaking models to appear impressive rather than truly intelligent, leading to misleading outputs and unsafe models. As a response, Antidote is introduced as an alternative evaluation framework that prioritizes expert reviews of AI outputs based on substance, accuracy, and real-world stakes, using domain-specific experts to ensure robust and honest assessments. Antidote aims to foster AI models that value long-term trust and truthfulness over short-term appeal, encouraging a shift away from gamified rankings towards genuine intelligence and utility.