Can AI automate AI R&D yet? Early evidence from InnovationEval: No
Summary
Epoch AI evaluates whether AI can autonomously advance AI R&D through InnovationEval, an end-to-end test of discovering a novel ML algorithmic innovation. The study finds that current frontier models struggle to match the human-developed on-policy self-distillation approach, with GPT-5.6 Sol showing a modest improvement via self-imitation; concerns include reported results, memorization, and the trade-offs between budget and performance.