PsychoSafe: Eliciting Psychologically-Informed Refusals in Large Language Models
2026-06-08 • Computation and Language
Computation and Language
AI summaryⓘ
The authors created PsychoSafe, a system that helps large language models say no safely and supportively in risky situations, like crises or coercion. Instead of just refusing, PsychoSafe guides responses to be more helpful and grounded in psychological methods. They trained this system on many examples and found it improved refusal quality and made better referrals to outside help, without hurting other tasks. However, it works best on situations similar to its training, showing the need for more diverse data in the future.
large language modelsrefusal strategiespsychological interventionfine-tuningresource referralQwen 3.5 27Bpromptingmodel evaluationrisk domains
Authors
Gianluca Barmina, Federico Torrielli, Sven Harms, Jacob Nielsen, Felix Mächtle, Stine Lyngsø Beltoft, Peter Schneider-Kamp, Thomas Eisenbarth, Lukas Galke Poech, Anne Lauscher
Abstract
Large language models (LLMs) routinely face requests that should be refused, creating a trade-off between helpfulness and harm prevention. However, refusals themselves can be helpful. In high-risk interactions involving crisis, coercion, or escalating intent, blunt non-compliance may prevent direct harm while still failing to support the needs of the person behind the request. We present PsychoSafe, a psychologically-informed refusal framework that reframes refusal as structured supportive communication grounded in evidence-based intervention strategies. To develop PsychoSafe, we construct a corpus of 8019 prompt-response pairs spanning five psychologically salient risk domains and apply prompting and parameter-efficient fine-tuning to Qwen 3.5 27B. On a balanced validation set of 500 prompts, evaluated with an LLM judge and validated through human ratings, PsychoSafe prompting improves overall refusal quality by 28.1% over a generic baseline, with particularly strong gains in external resource referral (+46.8%) and psychological grounding (+34.8%), while preserving downstream performance on non-refusal tasks. Fine-tuning achieves near-perfect refusal and resource-referral rates but reduces response relevance. Additional evaluations on SORRY-Bench and XSTest show strong in-domain robustness but limited out-of-domain generalization, suggesting that future work should diversify fine-tuning data to help models apply interventions selectively rather than schematically.