Zhi Wang and colleagues argue that GPT-6 Astra’s changes to its mathematical capabilities are closer to context-dependent activation steering than to native weight-space programming. Astra pushes model activations along a fixed direction constructed from short synthetic trajectories, with the strength varying linearly with the current activations. The authors also generalize this mechanism using learned “perception vectors” and point out that the design of synthetic trajectories is clearly more effective.
Source:LessWrong · Read the original
Leave a Reply