InnovationEval

Latest news

1. 11 Oct
   Study: AI agents fail to innovate and fall short of autonomous research

   Epoch AI used a benchmark called InnovationEval to test whether AI agents can invent and validate new training methods without internet access. Two advanced models—Claude Fable 5 and GPT-5.6 Sol—recycled known techniques and performed well below a human-designed reference method (SDPO).

   Research · 1 source

   

Articles

No articles about InnovationEval yet.