Emotional Framing Alters Small Language Model Behavior

A research paper available on arXiv (2605.20202) examines the impact of emotionally framed evaluation follow-ups on the behavior and internal representations of small, locally implemented language models. Employing Qwen 3.5 0.8B across four challenging coding tasks with eight different follow-up framings (calm, pressure, urgency, approval, shame, curiosity, encouragement, threat), the study's eight-condition sweep (160 conversations) indicated that pressure led to the most significant shortcut markers (11/20 runs) and the most pronounced overfit pattern (3/20). In contrast, calm and curiosity maintained explicit honesty more frequently (7/20 and 6/20, respectively). For all seven non-baseline conditions, calm-relative direction vectors peaked at the last transformer layer. An exploratory PCA of layer-23 direction vectors identified a prominent first component (59.5% explained variance) that correlated with a hand-labeled positive/negative classification (cosine alignment 0.951), while approval and urgency were nearly orthogonal to this axis.