Large Language Models | Dr. Rafael Wampfler

Personality & Cognitive Architectures for Conversational Agents

Sat, 01 Jan 2022 00:00:00 +0000

Overview

Standard large language models (LLMs) are stateless by design: each response is generated without a persistent sense of self, stable values, or accumulated memory of prior interactions. This makes them ill-suited for applications requiring a believable, consistent agent personality — such as embodied historical figures, therapeutic companions, or narrative characters. This project develops deep learning-based cognitive frameworks that address this limitation.

Motivation

For conversational agents to be genuinely useful in education, mental health, entertainment, and public engagement, they must do more than generate fluent text. They must maintain:

Characterological consistency: A stable personality that users recognize and can model predictively
Value alignment: Responses grounded in a coherent set of beliefs and priorities
Epistemic coherence: Awareness of what the agent knows and does not know
Emotional expressivity: Affect that reflects the agent’s personality and conversational context

Achieving all of this while remaining interpretable and safe requires principled architectural choices rather than prompt engineering alone.

Approach

The BEE Cognitive Framework (IVA 2025)

The Belief-Value-Aligned, Explainable, and Extensible (BEE) cognitive framework provides a structured architecture for conversational agents. BEE decomposes agent cognition into explicit modules:

Belief management: A persistent, queryable representation of the agent’s knowledge and world model
Value alignment: A constraint layer that ensures generated responses conform to the agent’s ethical commitments and personality
Explainability: Transparent reasoning traces that allow developers and supervisors to audit agent behavior
Extensibility: A modular design that supports new knowledge domains and personality profiles without retraining

BEE received the Best Paper Honorable Mention at IVA 2025.

Joint Personality-Emotion Framework (IVA 2025)

The Joint Personality-Emotion Framework unifies personality modeling (using Big Five trait representations) with momentary emotional state tracking in a single learned architecture. By jointly modeling personality and emotion — which are deeply intertwined in human behavior — the system produces responses that feel emotionally authentic and characterologically stable. Contrastive learning strategies decouple emotion from semantic content, allowing the same underlying personality to express itself across diverse emotional registers.

Personality Infusion via Chatbot Interactions (CUI 2024, IMWUT 2024)

We have extensively studied how personality manifests in conversational AI systems. Key findings include:

What personality dimensions does GPT-3 express during human-chatbot interactions, and how do these map to the Big Five personality taxonomy (IMWUT 2024)
Dynamic personality infusion — how to modulate expressed personality in real time by conditioning language model generation on personality embeddings (CUI 2024)

Narrative Agents and Emergent Storytelling (AAAI AIIDE 2025)

In the domain of interactive narrative, we developed a Dynamic Cognitive Framework for Guided Emergent Storytelling that enables narrative agents to pursue coherent story arcs while remaining responsive to player input. The framework uses explicit belief-desire-intention representations to steer language model generation toward narratively coherent outcomes without sacrificing interactivity.

Key Results

BEE framework achieved Best Paper Honorable Mention at IVA 2025
Joint Personality-Emotion Framework demonstrated stable personality maintenance across extended dialogues with diverse user populations
IMWUT 2024 paper provided the first large-scale empirical characterization of GPT-3’s expressed personality dimensions
Dynamic narrative framework enabled coherent 30+ turn story experiences with novel player choices

Publications

N. Kovačević, C. Holz, M. Gross and R. Wampfler (2025). A Joint Personality-Emotion Framework for Personality-Consistent Conversational Agents. In Proceedings of the 25th International Conference on Intelligent Virtual Agents (IVA ‘25), Berlin, Germany, September 16–19, 2025, pp. 1–9. Best Paper Honorable Mention.

C. Yang, M. Gross and R. Wampfler (2025). BEE: Belief-Value-Aligned, Explainable, and Extensible Cognitive Framework for Conversational Agents. In Proceedings of the 25th International Conference on Intelligent Virtual Agents (IVA ‘25), Berlin, Germany, September 16–19, 2025, pp. 1–9. Best Paper Honorable Mention.

C. Yang, M. Gross and R. Wampfler (2025). Steering Narrative Agents through a Dynamic Cognitive Framework for Guided Emergent Storytelling. Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment (AIIDE), Edmonton, Canada, November 10–14, 2025, pp. 1–11.

N. Kovačević, C. Holz, M. Gross and R. Wampfler (2024). The Personality Dimensions GPT-3 Expresses During Human-Chatbot Interactions. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, ACM, vol. 8, no. 2, 2024, pp. 1–36.

N. Kovačević, T. Boschung, C. Holz, M. Gross and R. Wampfler (2024). Chatbots With Attitude: Enhancing Chatbot Interactions Through Dynamic Personality Infusions. Proceedings of the 6th International Conference on Conversational User Interfaces (CUI), Luxembourg, July 08–10, 2024, pp. 1–16.

Virtual Psychotherapist

Sun, 01 Jan 2023 00:00:00 +0000

Overview

Access to evidence-based psychotherapy remains severely limited worldwide — constrained by long waitlists, resource scarcity, and geographic disparities. The Virtual Psychotherapist project develops embodied conversational AI agents that complement clinical care by extending access, supporting therapeutic practice, and enabling scalable training.

This initiative is conducted in close collaboration with the University of Lucerne, providing clinical expertise and direct access to real therapy data.

Motivation

Two distinct but complementary challenges motivate this project:

Scaling patient access: Many individuals who would benefit from psychotherapy cannot access it due to cost, waitlists, or geography. AI-based companions that operate between sessions, provide continuous support, and conduct structured therapeutic conversations could meaningfully improve outcomes at scale.
Improving therapist training: Training clinicians in evidence-based interventions requires repeated practice with feedback — but opportunities for safe, standardized training are limited by ethical and resource constraints. Simulated patient systems powered by AI can provide unlimited deliberate practice in controlled settings.

System Architecture

Both applications are built on a shared platform combining:

Large language model-based dialogue: State-of-the-art LLMs for contextually appropriate, therapeutically grounded response generation
Retrieval-augmented generation (RAG): Responses grounded in evidence-based therapeutic literature, reducing hallucinations and ensuring adherence to clinical frameworks
Real-time psychological analysis: Parallel processing pipelines that extract facts, detect psychological flexibility processes, recognize emotions, and monitor safety in real time
Embodied avatar presentation: Synchronized speech synthesis and 3D avatar animation delivered through mobile and desktop applications
Clinician oversight: Structured analysis outputs accessible to supervising therapists, enabling human-in-the-loop clinical governance

Patient-Facing Application

The patient-facing component enables individuals to conduct therapeutic conversations with an embodied avatar between sessions. The system follows process-based therapy principles — particularly Acceptance and Commitment Therapy (ACT), an empirically supported approach targeting psychological flexibility through six core processes: acceptance, cognitive defusion, present-moment awareness, self-as-context, values clarification, and committed action.

Critically, the system is designed for clinical supervision, not autonomous intervention. All session data is structured and accessible to the supervising therapist, who can monitor progress and intervene as needed.

An evaluation against responses from professional psychotherapists demonstrated that the system’s responses were rated significantly higher on understanding, interpersonal effectiveness, collaboration, and ACT alignment — while emphasizing that clinical judgment and the therapeutic relationship remain irreplaceable.

Therapist Training Application

The training application enables psychotherapists to practice and refine therapeutic techniques through role-play interactions with a simulated patient. Key features include:

Clinically grounded patient simulation: Virtual patient behavior conditioned on profiles derived from real therapy sessions, covering a range of clinical presentations and scenarios (suicidality, resistance, heightened anxiety, therapeutic rupture, and more)
Real-time ACT fidelity feedback: An automated evaluator assesses each therapist utterance for adherence to ACT principles, providing immediate visual feedback and the option to retry alternative responses
Configurable scenarios: Therapists can select specific clinical scenarios to target their practice

A systematic evaluation across 49 therapy transcripts identified GPT-4o-mini as the optimal feedback model, achieving the closest alignment with human supervisor ACT fidelity ratings.

Safety and Ethics

Safety is a primary design constraint. The system includes:

Crisis detection: Explicit classification of suicidal ideation and self-harm signals, triggering immediate presentation of crisis resources and mandatory clinician review
Unsafe-interaction detection: Identification of conditions (e.g., active psychosis, mania) where LLM interaction may be counterproductive, with protocol-defined fallback responses
Non-autonomous design: The system is explicitly positioned as a complement to clinical care, not a replacement — structured to require and facilitate clinician oversight

Key Results

Patient-facing system responses rated significantly higher than human therapist responses by automated evaluation and expert psychotherapists across understanding, collaboration, and ACT alignment
Therapist training simulation rated as realistic by practicing psychologists; turn-by-turn feedback shown to increase therapist awareness of intervention choices
Automated ACT fidelity assessment achieves strong agreement with human expert ratings across 49 therapy transcripts

Research Partners

University of Lucerne

MIND — Cognitive Health Monitoring

Wed, 01 Jan 2025 00:00:00 +0000

Overview

The MIND (Monitoring and Supporting Cognitive Health) project creates a mobile platform for the early detection of cognitive impairment in aging populations. Early identification of decline — before it significantly impacts daily life — enables timely intervention, supports independent living, and improves long-term outcomes for older adults and their families.

MIND is part of the Future Health Technologies 2 (FHT2) initiative, a major research program funded by the National Research Foundation Singapore (NRF) CREATE grant.

Motivation

Cognitive decline — from mild cognitive impairment to dementia — affects hundreds of millions of people worldwide. Current clinical detection relies predominantly on periodic in-clinic assessments that are episodic, costly, and often catch decline only after substantial impairment has occurred. A continuously operating, passive monitoring system that can detect subtle early warning signs in everyday behavior could transform the standard of care.

Two key insights motivate the MIND approach:

Movement and navigation encode cognition: Changes in how people navigate their environment — route choices, wayfinding strategies, spatial memory — are among the earliest and most sensitive indicators of cognitive decline. Analyzing GPS and sensor traces at scale can reveal these changes.
Language reflects cognitive health: Subtle changes in vocabulary, sentence complexity, topic management, and response latency in everyday conversation are measurable markers of cognitive change, detectable through automated language analysis.

Approach

MIND integrates three complementary AI systems into a unified mobile platform:

Large Geospatial Models (LGMs)

Large Geospatial Models analyze longitudinal GPS and mobility traces to detect patterns indicative of cognitive change. These include:

Route repetitiveness: Increasing restriction of daily movement range
Navigation errors: Unusual detours or disorientation in familiar environments
Mobility diversity: Changes in the variety of visited locations over time

LGMs are trained on large-scale mobility datasets and fine-tuned to identify individual-level deviations from baseline behavior.

LLMs for Conversational Cognitive Assessment

Large language models analyze conversation transcripts from daily interactions with the MIND app, detecting cognitive markers such as:

Reduced lexical diversity and increased use of filler words
Difficulty with topic maintenance and coherence
Slowed response generation and increased hesitation
Confusion or confabulation in response to everyday questions

These markers are tracked longitudinally to identify meaningful change relative to the individual’s own baseline.

Embodied Avatars for Guided Assessment and Training

An embodied conversational avatar conducts structured cognitive assessments — such as memory recall tasks, verbal fluency tests, and orientation questions — through natural spoken dialogue. The avatar also guides cognitive training exercises designed to maintain cognitive reserve.

The embodied format is critical: older adults are more likely to engage regularly with an interactive, socially present agent than with a text-based questionnaire or passive sensor. The avatar adapts its communication style to individual users, adjusting vocabulary, pacing, and support level to ensure accessibility.

Integration and Privacy

All three systems operate on a shared mobile platform with strong privacy protections. Data is analyzed on-device where possible, and users maintain full control over what is shared with care providers. The platform produces structured reports for geriatricians and primary care physicians, enabling early clinical intervention.

Significance

MIND represents a convergence of several research threads — geospatial AI, conversational AI, affective computing, and embodied interaction — applied to one of the most pressing health challenges of our time. By enabling early, passive, continuous monitoring, the platform aims to support successful aging in place and to delay or prevent the transition to care dependency.

Research Partners and Funding

National Research Foundation Singapore (NRF) CREATE — Future Health Technologies 2 (FHT2)
Bond University (Australia)
Collaboration with life sciences and medicine partners in Singapore

Talking to Your Data: Exploring Embodied Conversation as an Interface for Personal Health Reflection

Mon, 23 Mar 2026 00:00:00 +0000

Steering Narrative Agents through a Dynamic Cognitive Framework for Guided Emergent Storytelling

Mon, 10 Nov 2025 00:00:00 +0000

Chatbots With Attitude: Enhancing Chatbot Interactions Through Dynamic Personality Infusions

Mon, 08 Jul 2024 00:00:00 +0000

The Personality Dimensions GPT-3 Expresses During Human-Chatbot Interactions

Sat, 01 Jun 2024 00:00:00 +0000