Answers are grounded in this paper and cite their sources.
Across eight preregistered studies in small open-weight language models, we examined what happens to a hostility-associated influence, introduced by activation steering, when the model is instructed to behave warmly. Under a warmth instruction, steering increased a latent preference for hostile replies relative to non-affective and random control directions, and this persistence replicated across three model families (Studies 1–2). A predicted hostility-specific carryover through the model's hidden state alone was not supported (Study 2). When the model was allowed to reason before answering, however, the influence carried into a later rating of the user (about −1.0 points on a 0–9 scale), and a reasoning trace written under steering carried an effect even when the subsequent answer and rating were produced without steering (Study 3). Replaying identical stored traces in different conversational positions suggested that the effect depends on where the text appears: it was largest when the trace occupied the model's current thinking, smaller in a previous assistant turn, and absent when the same text was quoted inside a user message (Studies 4–5). The effect remained when answer content was held fixed (Study 6), and warmth-oriented constraints changed the tone of outputs without reducing the influence on the later rating (Studies 7–8). Taken together, the results are consistent with an influence that is carried mainly through text the model treats as its own reasoning rather than through hidden state, and that surface constraints on expression do not remove. The findings come largely from a single 4B-parameter model with one stimulus set, effect sizes are modest, and no claim is made that the model has affect. Each study's preregistration, code and data are publicly archived. This paper integrates eight preregistered studies reported individually in: 10.5281/zenodo.23182205 (Studies 1–2), 10.5281/zenodo.23323490 (Study 3), 10.5281/zenodo.23747766 (Studies 4–6) and 10.5281/zenodo.23876548 (Studies 7–8). Code and data: https://github.com/arsumed-del/affect-under-constraint AI contribution: experiments were executed by OpenAI Codex and designed and reviewed with Anthropic Claude; this synthesis was drafted by Anthropic Claude from the individual reports and approved by the author.
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: