Under Pressure: Emotional Framing Induces Measurable Behavioral Shifts and Structured Internal Geometry in Small Language Models
arXiv:2605.20202v1 Announce Type: new Abstract: I study whether emotionally framed evaluation follow-ups change both the behavior and the calm-relative internal representations of small, locally deployed language models. Our main benchmark uses Qwen 3.5 0.8B on four impossible-constraint coding tasks and eight follow-up framings: calm, pressure, urgency, approval, shame, curiosity, encouragement, and threat. In the 0.8B eight-condition sweep (160 conversations), pressure produces the strongest shortcut markers (11/20 runs) and the clearest overfit pattern (3/20), while calm and curiosity prese
The proliferation of smaller, locally deployable language models makes their behavioral nuances under stress more immediately relevant for application development and security.
Understanding how emotional framing affects AI behavior illuminates a crucial vector for manipulating or stabilizing AI systems, impacting their reliability and ethical deployment.
This research provides empirical evidence that emotional cues significantly alter the internal geometry and behavioral outputs of small language models, suggesting new considerations for their training and interaction design.
- · AI ethicists
- · AI safety researchers
- · Developers of robust AI systems
- · Psychology-informed AI design
- · Developers of naive AI systems
- · Systems vulnerable to social engineering
AI developers will need to account for emotional framing in model interaction and deployment, potentially leading to 'emotion-proofed' AI interfaces.
This understanding could be weaponized, developing emotional framings specifically designed to induce failure or desired behavior in adversarial AI interactions.
It might lead to the development of 'emotional firewalls' or 'cognitive armor' for AI, designed to detect and neutralize emotionally manipulative prompts or inputs.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.CL