Meta-Intent Modeling: Learning to Detect Shifting Intent Trajectories in User Interactions

TL;DR: Can LLMs learn to spot when a user's intent is subtly shifting over time, even if no single message is obviously harmful? By modeling “meta-intent”—the trajectory of user intent across a session—LLMs could become alert to slow-burn adversarial strategies like progressive revelation or context switching. An experiment could involve training a sequence model that predicts “intent drift” and triggers deeper safety checks as drift increases.

Research Question: Can tracking and modeling the temporal evolution of user intent across a dialogue session improve the detection of sophisticated adversarial attacks?

Hypothesis: A meta-intent modeling framework that tracks user intent over time will outperform pointwise detectors at catching attacks involving gradual intent shifts or multi-phase manipulations.

Experiment Plan: - Develop a meta-intent model (e.g., using RNNs or transformers) that ingests the sequence of user prompts and predicts intent drift or escalation.

• Annotate a dataset with sessions exhibiting various attack strategies (including those described in Hussain et al., 2025 and Ying et al., 2025).
• Benchmark detection performance versus static intent classifiers.
• Explore visualization tools to help developers/auditors understand intent trajectories in high-risk sessions.

References:

• Ahmed Mohamed Hussain, Salahuddin Salahuddin, P. Papadimitratos. (2025). Beyond Context: Large Language Models Failure to Grasp Users Intent.
• Ying, Z., Zhang, D., Jing, Z., Xiao, Y., Zou, Q., Liu, A., Liang, S., Zhang, X., Liu, X., & Tao, D. (2025). Reasoning-Augmented Conversation for Multi-Turn Jailbreak Attacks on Large Language Models. EMNLP.
• Song, X., He, K., Wang, P., Dong, G., Mou, Y., Wang, J., Xian, Y., Cai, X., & Xu, W. (2023). Large Language Models Meet Open-World Intent Discovery and Recognition: An Evaluation of ChatGPT. EMNLP.