Epoch AI's InnovationEval finds today's agents short of independent AI-algorithm research
In a small initial test, Epoch found GPT-5.6 Sol and Claude Fable 5 did not match a human-designed post-training method after correction for run selection and out-of-scope changes.