Interpretability debate: steer by difference of means or by gradients?

johnhewtt and arunasank are debating the distinction between "forms" and "estimators" in interpretability research. On October 27, johnhewtt published a new blog post, "separating forms from estimators, or, why steer by difference of means?", arguing that "forms" such as steering and neuron analysis are often conflated with "estimators" like difference of means (DoM) or SGD on log-loss: much steering work treats "constructing steering vectors via difference of means" as the method itself. He argues that if gradients are available, one should use gradient descent directly rather than defaulting to DoM.

Confirmed

• johnhewtt's core claim: localization methods (which can be wrong) and activation modification approaches (where DoM can also be wrong) do not provide mutual validation; the fact that DoM works better on the layers selected by localization does not prove either is correct.
• He further argues that DoM is no more transparent than gradient-based selection of steering vectors—both ultimately require interpreting nonlinearities of a deep network, so neither is more interpretable.
• arunasank responded: behavioral performance is evidence that an intermediate variable is effective; in their paper, gradient-based steering methods such as ReFT work in limited settings without requiring localization, which is one of the selling points of gradient approaches.

Why It Matters

• The debate strikes at the methodological foundations of interpretability research: if DoM is neither more transparent nor more effective than gradient-based methods, then the large body of steering and localization work that relies on "difference-of-means steering vectors" as the mainstream tool needs to re-examine its justification.
• The "forms vs estimators" distinction gives the community a clearer conceptual framework for assessing what various intervention experiments actually do and do not demonstrate.

2026-10-27 ~ 2026-10-27 · 6 related posts

• [source] Separate Forms from Estimators: Why Steer by Difference of Means When You Can Use SGD? — johnhewtt · 2026-10-27
• [source] New blog: separate forms from estimators — when you can learn by SGD, just should — arunasank · 2026-10-27
• [source] Gradient-based steering like ReFT may remove the need for localization — arunasank · 2026-10-27
• DoM steering success isn't strong evidence for interpretability localization — johnhewtt · 2026-10-27
• DoM steering working at a localized layer doesn't prove localization is correct — johnhewtt · 2026-10-27
• Researcher argues DoM steering is no more transparent than gradient-based vector selection — johnhewtt · 2026-10-27