Gradient-based steering like ReFT may remove the need for localization
arunasank · x · 2026-10-27
arunasank responds to johnhewtt: behavior is evidence that mediators are effective, and gradient-based steering methods like ReFT (from their paper) work without needing localization in limited settings. The exchange touches the core tension of johnhewtt's blog on separating forms from estimators — interpretable localization+steering vs. direct fine-tuning.
Related event: Interpretability debate: steer by difference of means or by gradients?(6 posts)→