Re:Astra can read and write weights to improve Qwen
by porridgeraisin on Tuesday November 03, 2026 @12:22AM
(#50348410)
> For Qwen, the model fixed a looping bug that may have been known apriori
It was included in the prompt itself. The authors say they used claude code for the experiment, and claude wrote the prompt for astra, which included details of the looping bug. Then astra seems to have done h(looping) - h(loop fixed) and then gotten the usual steering vector.
It was included in the prompt itself. The authors say they used claude code for the experiment, and claude wrote the prompt for astra, which included details of the looping bug. Then astra seems to have done h(looping) - h(loop fixed) and then gotten the usual steering vector.