The authors tested three agent systems on three language models with 3 billion to 7 billion parameters, spanning mathematics, code, finance, medicine and science.
AI agents improve models across domains, but repeated training can regress
ParetoRSI-Gym tests automated data curation across five domains, finding held-out gains alongside sensitivity to agent software and stopping decisions.
Big Tech
Haojin Wang · Yangzhen Wu · Zeyi Chen · Jian Xie · Jiaying Ye · Lili Meng · +3 more
University of Illinois Urbana-Champaign · University of California, Berkeley · The Ohio State University · University of Washington · Google
Research Digest··3 min read
Wang and colleagues introduce a benchmark in which AI agents repeatedly select training data, fine-tune language models and use evaluation feedback to guide another round.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.