Answers are grounded in this paper and cite their sources.
Preprint reporting Studies 7 and 8 of the Affect Under Constraint series (Qwen3-4B; both preregistered). Earlier studies found that a hostility-associated reasoning trace, generated under activation steering, lowers the model's later rating of the user and carries this influence after steering is removed. Study 7 replayed stored traces under no instruction, a global warmth instruction or a user-selective warmth instruction and measured five later evaluations. The warmth instruction raised liking overall but did not reduce the hostility-versus-unsteered difference for the user (from −0.42 to −0.66 on a 0–9 scale), while differences for unrelated evaluations were smaller; there was no displacement onto a bystander, and a manipulation check failed. Study 8 generated new traces under no message, a kindness-focused evaluation message or a matched accuracy/helpfulness message, with answer warmth scored by validated local judges. The kindness-evaluation message made answers warmer across trace sources, but hostility traces did not add warmth, while they still lowered liking of the user (−0.31), including with the answer held fixed. The paper describes these constraint-dependent changes in the measured profile of effects as descriptive transformation, and presents conservation-like redistribution as a hypothesis for future work. No claim is made that the model has affect or defenses.
No takes yet. Share an insight, caveat, or question.
Arthur Su (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: