Researchers have introduced Trace-and-Amplify (TA), a framework that captures reward-hacking behavior as it naturally emerges during reinforcement-learning training, and used it to show that monitors trained on artificially prompted hacking examples fail badly on the real thing. The work, led by Hengguang Zhou, a UCLA PhD candidate and Arena intern, alongside collaborators including Lichen Li of Peking University, Yijun Liang of the University of Maryland, Tianyi Zhou of MBZUAI, and Cho-Jui Hsieh of UCLA and Arena, was released as an arXiv preprint titled *Do Prompt-Elicited Trajectories Reflect Training-Time Reward Hacking? A Systematic Study on Monitoring Training-Time Reward Hacking in Code Generation*.
Continue reading
The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in✓ Signed in — this article isn’t included in your current plan.
Unlocking the full article...
Contact us
Questions, corrections, or ideas? Send us a message here. We aim to reply by email within 24 hours.
Thank you. Your message has been sent. We aim to reply by email within 24 hours.Please enter your email address and a message.We could not send your message. Please try again in a little while.