RedwoodBench: Can Frontier Models Imitate Conceptual AI Safety Researchers?

RedwoodBench evaluates whether leading AI models can replicate the reasoning and conceptual insights of AI safety researchers. The benchmark tests frontier models on tasks reflecting AI safety research challenges to assess their understanding and imitation capabilities.