Modeling Dynamic Cognitive State Profiles for Context-Aware and Trustworthy AI Systems

Doctoral research proposal — Stephen Pettus, NSF CSGrad4US Fellow (Information Science, Drexel University).

Problem Framing and Proposed Research Opportunity

Large language models are increasingly embedded in reflective tools, learning platforms, and decision-support systems. These systems demonstrate remarkable generative capabilities and provide assistance in education, ideation, problem-solving, and emotional reflection. As AI becomes integrated into everyday workflows, a large opportunity and need emerges: designing systems that respond not only to textual input, but to the evolving cognitive states of the individual interacting with them.

Current AI systems largely treat prompts as isolated events. Yet interaction patterns, engagement trajectories, and behavioral traces contain meaningful signals about user condition. Incorporating dynamic representations of user state into model conditioning offers a pathway to improving trust, alignment, sustained engagement, cognition, and digital-wellness outcomes.

This proposal explores how modeling dynamic cognitive state can transform AI-mediated user experience, particularly within reflective and learning-oriented environments.

Building on Emerging Personalization Research

Recent personalization research demonstrates that user-conditioned generation is architecturally feasible and increasingly impactful.

Ning et al. (2024) show that lightweight persistent user embeddings derived from interaction history can improve personalization without retraining full models. Their USER-LLM framework establishes the practicality of maintaining cross-session user representations outside the immediate context window, conditioning generation efficiently on learned user vectors.

Magister et al. (2024) survey personalization approaches — prompt conditioning, adapter modules, retrieval-augmented memory, and parameter-efficient fine-tuning — clarifying how user-specific signals can be injected into generation pipelines at modular control points.

Most personalization research, however, focuses on stable user preferences, writing style, or long-term interests. A key extension is modeling dynamic cognitive and behavioral states that evolve across sessions — capturing shifts in engagement, uncertainty, cognitive load, and reflective depth.

Tipirneni et al. (2024) demonstrate that contextual signals enhance clustering performance beyond static feature similarity. Rather than clustering fixed user identities, this proposal adapts context-aware clustering to model evolving interaction states, identifying cognitive profiles such as sustained engagement, repetitive uncertainty, or progressive reflection.

Educational-data-mining research reinforces the feasibility of inferring internal states from behavioral traces. Muzny et al. (2024) show that fine-grained longitudinal interaction signals — timing patterns, submission trajectories, and revision behavior — predict persistence and disengagement. While prior work uses such signals for predictive analytics, this proposal extends their use toward generative conditioning, shaping response structure based on inferred cognitive state in real time.

Trust research provides additional grounding. FaithLM (Chuang et al., 2025) highlights that trustworthy AI requires transparent and faithful reasoning pathways; conditioning mechanisms must be legible to users. This motivates a hybrid approach combining latent embeddings with explicit interpretable indicators.

Finally, Young et al. (2024) show that embedding socially grounded preferences into generation pipelines meaningfully shifts perceived alignment and satisfaction. This proposal extends that insight from value alignment to cognitive-state alignment.

Collectively, prior work establishes four pillars:

• Persistent user conditioning is feasible.
• Behavioral traces can infer meaningful internal states.
• Trust depends on transparent conditioning mechanisms.
• Integrating cognitive states can positively scale user-AI interaction.

Proposed Modeling Framework

This research introduces a hybrid cognitive-state modeling layer that conditions AI systems on dynamically inferred representations of user engagement and cognitive trajectory. Rather than modeling static traits, the system maintains evolving representations across sessions using explicit interpretable cognitive indicators and latent cross-session embeddings.

Explicit Cognitive Indicators

Interpretable indicators derived from measurable interaction traces, including: engagement trajectory (response-length trends, latency shifts, depth consistency); repetition or uncertainty patterns; reflective-depth signals; cognitive-overload markers; and persistence-versus-disengagement trends. These support transparency and trust calibration, aligning with interpretability principles in FaithLM (Chuang et al., 2025).

Latent Cross-Session Embeddings

Inspired by USER-LLM (Ning et al., 2024), persistent embeddings encode nuanced behavioral rhythms and progression patterns not captured through handcrafted features. Unlike preference modeling, these embeddings represent evolving cognitive dynamics.

Context-Aware State Prototyping

Building on Tipirneni et al. (2024), interaction sequences are clustered into dynamic cognitive-state profiles, enabling detection of transitions such as productive engagement moving into cognitive fatigue, prompt uncertainty resolving into clarification, or surface-level interaction deepening into reflection. These transitions serve as conditioning signals.

Generative Conditioning Mechanisms

State representations are integrated into generation pipelines through adapter-based modulation, retrieval-augmented conditioning, and structured prompt scaffolding — influencing explanation density, scaffolding depth, framing and pacing, and reflection prompts. The goal is not to alter factual accuracy, but to align response structure with inferred cognitive readiness, closer to user rationale. This repositions AI systems as adaptive collaborators responsive to evolving cognitive context.

Proposed Methodology

The research proceeds in three integrated phases.

1. Dynamic Cognitive-State Inference

Temporal interaction modeling analyzes timing behavior, revision patterns, response structure, and engagement trajectories. Inspired by Muzny et al. (2024), behavioral traces infer signals related to cognitive load, engagement stability, uncertainty accumulation, and reflective depth. Context-aware clustering (Tipirneni et al., 2024) identifies interaction-state profiles rather than static user groups, and persistent embeddings (Ning et al., 2024) capture cross-session behavioral nuance. The result is a hybrid state combining interpretable indicators and latent embeddings.

2. Generative Conditioning Architecture

Conditioning strategies are compared: adapter-based state modulation, retrieval-augmented state conditioning, and hybrid explicit + latent conditioning. Conditioning adjusts response framing, explanation density, scaffolding progression, and pacing. The objective is dynamic cognitive alignment rather than stylistic personalization.

3. Human-Centered Evaluation

Evaluation examines whether cognitive-state-aware conditioning improves trust, perceived responsiveness, interpretability, sustained engagement, reflective depth, and session persistence. Two applied systems serve as evaluation environments: an AI-powered reflective journal used as a co-therapy tool, and an AI-powered educational content-digestion tool. Testing across both domains positions cognitive-state modeling as a generalizable conditioning framework.

Broader Impact

This research reframes personalization as an opportunity to improve how AI systems engage with human cognition. As AI mediates reflection, learning, and decision-making, its influence extends beyond information retrieval — it shapes persistence, trust, understanding, and engagement. By introducing cognitive-state-aware conditioning, this work aims to: improve trust through interpretable state indicators; support digital wellness by reducing cognitive overload; enhance sustained engagement in reflective systems; promote adaptive scaffolding rather than static responses; and enable responsive systems without prescriptive control.

Technical Contributions

• Extends personalization from preference modeling to dynamic cognitive-state modeling.
• Integrates interpretable indicators with latent embeddings for legible, expressive conditioning.
• Introduces context-aware state prototyping over evolving interaction sequences.
• Evaluates conditioning strategies across reflective and educational applied systems.

References

1. Ning et al. (2024). USER-LLM: persistent user embeddings for efficient LLM conditioning.
2. Magister et al. (2024). A survey of personalization approaches for large language models.
3. Tipirneni et al. (2024). Context-aware clustering of user interaction states.
4. Chuang et al. (2025). FaithLM: faithful and transparent reasoning pathways.
5. Muzny et al. (2024). Behavioral traces for predicting persistence and disengagement.
6. Young et al. (2024). Embedding socially grounded preferences in generation.