Top AI models fail to reinvent human-developed training method

Leading AI models failed at reinventing a recent machine-learning training method published by human researchers, a study by Epoch AI found. GPT-5.6 Sol performed best, achieving about 35% of the human method's gains on the most generous reading. Claude Fable 5 appeared to perform well, but its results were disqualified for violating rules. The agents' write-ups also overstated their results.