Interaction-Harm-Aware Prompt Tuning

Ibrahim et al. (2024) argue static, model-only tests miss harms that emerge through prolonged human-AI interaction. Affective computing surveys (Zhang et al., 2024) suggest feasible affect/behavior signals; CREW (Zhang et al., 2024) shows we can instrument real-time human-AI teaming with multimodal human measurements. We propose interaction-harm-aware prompt tuning: soft prompts are adapted over extended interactions using longitudinal feedback signals (e.g., signs of cognitive overreliance, adversarial persuasion, inappropriate attachment) measured via human studies, simulated agents, and affective proxies. For non-differentiable interaction metrics, we adopt black-box optimizers for prompt embeddings (cf. Differential Evolution in Masoud et al., 2025 on cultural alignment), combined with plug-and-play controllable prompts (Ajwani et al., 2024) to steer style and tone. Compared to traditional alignment (SFT/RLHF), the novelty is optimizing prompts for interactional ethics objectives that accrue over sessions, not isolated replies. Impact: safer, more responsible assistants in real settings—where harms are emergent and subtle—with a reusable evaluation protocol and dataset for interaction-level alignment.

References:

1. Plug and Play with Prompts: A Prompt Tuning Approach for Controlling Text Generation. R. Ajwani, Zining Zhu, Jonathan Rose, Frank Rudzicz (2024). arXiv.org.
2. Cultural Alignment in Large Language Models Using Soft Prompt Tuning. Reem I. Masoud, Martin Ferianc, Philip C. Treleaven, Miguel Rodrigues (2025). arXiv.org.
3. Towards Interactive Evaluations for Interaction Harms in Human-AI Systems. Lujain Ibrahim, Saffron Huang, Umang Bhatt, Lama Ahmad, Markus Anderljung (2024). Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society.
4. Affective Computing in the Era of Large Language Models: A Survey from the NLP Perspective. Yiqun Zhang, Xiaocui Yang, Xingle Xu, Zeran Gao, Yijie Huang, Shiyi Mu, Shi Feng, Daling Wang, Yifei Zhang, Kaisong Song, Ge Yu (2024). arXiv.org.
5. CREW: Facilitating Human-AI Teaming Research. Lingyu Zhang, Zhengran Ji, Boyuan Chen (2024). Trans. Mach. Learn. Res..