Affect Under Constraint: Examining Expression and Persisting Influence in Language Models

Abstract

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). In Qwen3-4B, when the model reasoned before answering, 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 answer and rating were produced without steering (Study 3). Replaying identical stored traces in different conversational positions showed that the effect depended on where the text appeared: it was largest when the trace occupied the model’s current thinking, smaller in a previous assistant turn, and near zero when the same text was quoted inside a user message (Studies 4–5). The effect remained when answer content was held fixed (Study 6). A kindness-focused evaluation context made answers warmer without reducing the influence on the later rating (Study 8); a warmth instruction showed a similar pattern, but its manipulation check failed (Study 7). An exploratory analysis found that hostility-steered reasoning traces were consistently much longer than control traces, that this difference was larger under evaluative contexts, and that a descriptive, non-causal adjustment for trace length removed most of the hostility effect on the rating. Taken together, the results show a robust placement effect, in which the influence was carried by text in the model’s own reasoning positions rather than by hidden state alone, with the amount of that reasoning as a possible visible marker. Studies 3–8 used a single 4B-parameter model with one stimulus set, effect sizes are modest, and no claim is made that the model has affect. Preregistrations, code and data are publicly archived. Version of record: https://doi.org/10.5281/zenodo.24096561.

Keywords