The most interesting part isn't the self-improvement loop itself, it's that no… · @yueke · TwKit

@yueke

The most interesting part isn't the self-improvement loop itself, it's that no ground-truth grader is needed — the agent learns purely from contrasting its own runs. That lowers the barrier for domains where labeled data is scarce. Still, "better behaved on past tasks" is not the same as "better aligned"; someone still needs to watch what the harness is converging toward.

曝光 6 · 评论 0 · 点赞 0 · 书签 0