PsychoSafe: Eliciting Psychologically-Informed Refusals in Large Language Models

2026-06-08

AI Summary

This work addresses the frequent lack of psychological support in large language models when declining high-risk user requests. The authors propose PsychoSafe, a novel framework that systematically integrates evidence-based psychological intervention principles into the refusal mechanism. By constructing a corpus encompassing five categories of psychological risk and leveraging prompt engineering alongside parameter-efficient fine-tuning (PEFT), they optimize the Qwen-3.5-27B model. Experimental results on a balanced validation set of 500 samples demonstrate a 28.1% overall improvement in refusal quality, with external resource recommendations and psychological justifications increasing by 46.8% and 34.8%, respectively, without compromising performance on non-refusal tasks. After fine-tuning, refusals and resource recommendations approach near-perfect quality, albeit with a slight reduction in relevance.

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.

Problem

• refusal
• psychological support
• large language models
• harm prevention
• crisis intervention

Innovation

• psychologically-informed refusal
• supportive communication
• parameter-efficient fine-tuning
• harm prevention
• evidence-based intervention