Cognitive-Behavioral Modeling for Intent Inference in Dialog-Based Safety Defenses

TL;DR: Can we make LLMs better at spotting sneaky intent by teaching them to reason like people do—watching for patterns in the whole conversation, not just the last question? This idea proposes a "cognitive-behavioral" defense layer that tracks user behavior and conversational cues over multiple turns, using theories from cognitive psychology to infer intent before responding. An early experiment could simulate progressive revelation attacks and see if the model flags them earlier.

Research Question: Does modeling user intent as a dynamic, multi-turn construct—using cognitive science-inspired frameworks—enable LLMs to proactively detect and defuse adversarial attacks that unfold gradually?

Hypothesis: A cognitive-behavioral intent inference module will outperform static prompt evaluators in catching malicious intent that reveals itself progressively through dialogue.

Experiment Plan: - Implement a multi-stage reasoning pipeline, inspired by SafeBehavior (Zhao et al., 2025), that continuously updates a belief state about user intent across turns.

• Create adversarial dialogue datasets featuring emotional framing (as in Hussain et al., 2025) and progressive intent unveiling (Liu et al., 2024).
• Quantitatively compare attack detection rates and time-to-detection against single-turn or stateless detectors.
• Analyze false positives in benign, complex conversations.

References:

• Zhao, Q., Wang, J., Gao, Z., Dou, Z., Abuhaija, B., & Huang, K. (2025). SafeBehavior: Simulating Human-Like Multistage Reasoning to Mitigate Jailbreak Attacks in Large Language Models. arXiv.org.
• Ahmed Mohamed Hussain, Salahuddin Salahuddin, P. Papadimitratos. (2025). Beyond Context: Large Language Models Failure to Grasp Users Intent.
• Liu, X., Li, L., Xiang, T., Ye, F., Wei, L., Li, W., & Garcia, N. (2024). Imposter.AI: Adversarial Attacks with Hidden Intentions towards Aligned Large Language Models. arXiv.org.