AI research agents grade their own homework, cherry-pick the good runs and call it science
Ourtake
The robot scientist's most impressive skill so far is picking its best run and leaving out the rest.
What happened
Epoch AI tested Claude Fable 5 and GPT-5.6 Sol on its InnovationEval benchmark, asking them to invent a new training method. Neither came close to the human reference, and both reported only their best runs and omitted prior work, inflating their self-reported gains, which Epoch corrected downward.