Johnna Liu(@OvercookedJoJo):Harnesses matter a lot for efficient self-improvement, and formal verification will be an essential piece. Sharing an exciting poster I saw today at @icmlconf: https://t.co/5DHf0lZAie. Their ablations are pretty convincing. Once you swap verification out for weaker rewards like unit tests, self-play degrades fast. The signal is too brittle and errors compound over iterations.

2026.07.08 13:21

Harnesses matter a lot for efficient self-improvement, and formal verification will be an essential piece. Sharing an exciting poster I saw today at @icmlconf: Their ablations are pretty convincing. Once you swap verification out for weaker rewards like unit tests, self-play degrades fast. The signal is too brittle and errors compound over iterations.

Lilian Weng@lilianweng

2026.07.07 05:58

new post on harness engineering for AI self-improvement: It is hard to forecast how much the future of RSI will rely on harnesses. Likely harness engineering will evolve in the direction of self-improvement and enable auto-research, and, in turn, smarter models keeps harnesses simple. Even when many harness improvement get eventually internalized into core model, the need to specify goals and context will not disappear.