For long horizon tasks or optimizations at some point the model hyperfocuses on a solution and it ignores better solutions that might require some lateral thinking. And the way to get out of this is you just delete the context. So fork and flush called it flushing. Shenen called it forgetting. what I loved about Shenhen's paper is it talked about how he showed that the longer your context got the less the model thought like the reasoning tokens were shorter as it went on and the size of the edits which to me is like the biggest issue got shorter and shorter and shorter and shorter and if you forget delete the context it kind of resets it and bumps it back up.
★ In the digest▶ 48:05