AI interpretability steering methods less effective than thought

a new paper challenges the effectiveness of 'steering' in ai interpretability, where internal model directions are manipulated to change output. although directions like 'this is about food' appear causal and predictive, the research demonstrates that the linear relationship between pushing a direction and enhancing the concept is extremely limited. the author derives a closed-form taylor formula to define a radius where higher-order corrections remain bounded. empirical tests confirm this radius accurately predicts when linear steering fails, suggesting current intervention methods are less precise than assumed.