Back to News

Prime Intellect Audits Public Evaluations, Finds Broken Graders and Reward Hacks

#evaluation#audit#open-source#reward hacking

Prime Intellect published findings from auditing popular evaluations reported in model cards, uncovering broken graders, ambiguous tasks, and reward hacks that affect how results should be interpreted. The organization also open-sourced fixes for these issues.

Coverage timeline

  1. Prime Intellect

    While running popular evaluations reported in model cards, we found broken graders, ambiguous tasks and reward hacks that change how results should be read. We share what we found and open-source our fixes.

Related stories

News updates every hour. Subscribe in your RSS reader to get new stories automatically.

Subscribe