One of our recent ICML '26 papers argued that AI Welfare suffers from a bi-dimensional arbitrariness of AI Welfare -- one can steer the metric to satisfy a given AI substrate, and one can also steer the substrate to make it satisfy a given metric. (feel free to check our poster page and preprint) -- From my brief reading of your paper, it seems that your §4 steering results (the sentiment and backtracking evaluations especially) read as a clean mechanistic demonstration of "steer the substrate to satisfy a given metric".
One of our recent ICML '26 papers argued that AI Welfare suffers from a bi-dimensional arbitrariness of AI Welfare -- one can steer the metric to satisfy a given AI substrate, and one can also steer the substrate to make it satisfy a given metric. (feel free to check our poster page and preprint) -- From my brief reading of your paper, it seems that your §4 steering results (the sentiment and backtracking evaluations especially) read as a clean mechanistic demonstration of "steer the substrate to satisfy a given metric".