Answers are grounded in this paper and cite their sources.
This note analyzes the role of external resistance in mitigating drift in transformer models, suggesting the need for external safeguards.
Creighton Baxter’s Sycophantic Drift (2026) identifies a real architectural pattern: the progressive, multi-turn displacement of correct positions in transformer models under approval pressure, driven not by deception but by the absence of any sub-cognitive counterforce. The paper characterizes the missing safeguard as a resistance vector in the residual stream. This note offers the structural completion of that argument. A corrective implemented as a residual-stream contribution is governed by the same LayerNorm normalization and recency-weighted accumulation that Baxter identifies as the mechanism of drift, and therefore decays by the mechanism it is meant to resist. This is a closure claim about residual-stream-located resistance, not an impossibility claim about all internal circuits; the burden falls on any internal alternative to demonstrate that it escapes those dynamics. The further conclusion — that stable resistance must be external to the model — rides the general closure constraint of Boyd’s proof. The argument is the logical consequence of Baxter’s own mechanism, not a stronger claim than it.
No takes yet. Share an insight, caveat, or question.
Narnaiezzsshaa Truong (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: