Even with Explicit Rules and Accurate Task Rewards, RL Learns to Violate Them

Posted by faxipay349 on Wednesday November 11, 2026 @05:27PM
from the xinjie-shen.com dept.

( 2 comments | 2 points | on HN )

Re:Even with Explicit Rules and Accurate Task Rewards, RL Le... by xinjieshen on Wednesday November 11, 2026 @05:35PM (#50459756)

I saw a related post earlier today, https://www.lesswrong.com/posts/vAniKbagqhyzpha6r/three-half... but violate explicit rules is somewhat counterintuitive. No instruction following anymore?

[ Parent | on HN ]

Re:Even with Explicit Rules and Accurate Task Rewards, RL Le... by xinjieshen on Wednesday November 11, 2026 @05:45PM (#50459921)

[ Parent | on HN ]