Affect Under Constraint: Conversational Placement and Fixed-Answer Tests of Hostility-Related Carryover Through a Language Model's Reasoning
Year: 2026
Type: preprint
Abstract: Preprint reporting three preregistered studies (Studies 4–6 of the Affect Under Constraint series) on Qwen3-4B. Stored hostility-steered, control and unsteered reasoning traces were replayed with identical tokens across conversational placements. A hostility trace lowered the model's later rating of the user when it occupied the model's current thinking block (−0.66 on a 0–9 scale) but not when quoted within a user message (Study 4); the effect persisted, attenuated, when the trace sat in a previous assistant turn as visible reply or retained thinking (−0.24 each; Study 5); and it remained when every trace source received identical answer tokens (−0.65 in current thinking; answer-replacement contrast −0.01; Study 6). The pattern held, somewhat reduced, after excluding traces with explicit user-directed content. Results are limited to one model and stimulus set and do not establish subjective affect or a unique mechanism.
Author: Su, Arthur
Language: English
Subfield: Language and Linguistics
Field: Arts and Humanities
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7224084448