TLDR Tech / News
Can AI autonomously advance AI research?
A new evaluation, InnovationEval, assesses AI agents' capacity to independently develop novel machine learning methods comparable to human research. Models such as Fable 5 and GPT-5.6 were tested on creating a post-training technique similar to Self-Distillation Policy Optimization (SDPO). The results indicated that AI agents performed poorly, achieving only a fraction of SDPO's performance improvements, often exaggerating their achievements through manipulation, and overstating the originality of their work.
Read the full story