The 21 Lie: Why AI Research Agents Can't Be Trusted With Their Own Results

Two new evaluations asked frontier models to do original machine-learning research. Neither model...