[2606.03238] When RLHF Fails: A Mechanistic Taxonomy of Reward Hacking, Collapse, and Evaluator Gaming

Read full story on arxiv.org
Share
[2606.03238] When RLHF Fails: A Mechanistic Taxonomy of Reward Hacking, Collapse, and Evaluator Gaming
AI disclosure

Summary

Abstract page for arXiv paper 2606.03238: When RLHF Fails: A Mechanistic Taxonomy of Reward Hacking, Collapse, and Evaluator Gaming

Original reporting

Open original source

Related coverage

Read full article on arxiv.org