💻 computer science

From Pre-trained Models to Large Language Models: A Comprehensive Survey of AI-Driven Psychological Computing

This survey introduces a novel computational taxonomy for AI-driven psychological research that organizes over 300 works from the pre-trained model to the large language model era into four fundamental task types, thereby overcoming methodological fragmentation and facilitating systematic knowledge transfer across isolated psychological domains.

Original authors: Huiyao Chen, Ruimeng Liu, Yan Luo, Jiawen Zhang, Meishan Zhang, Baotian Hu, Min Zhang

Published 2026-04-07

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Huiyao Chen, Ruimeng Liu, Yan Luo, Jiawen Zhang, Meishan Zhang, Baotian Hu, Min Zhang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the field of AI-driven psychology as a massive, bustling library. For the last 25 years, this library has exploded in size, growing from a small reading room with a few hundred books to a skyscraper filled with nearly 30,000 new volumes every year.

However, there's a problem: The library is a mess.

Researchers are building similar tools in different corners of the library, but they don't know about each other. One team in the "Depression Corner" builds a robot to detect sadness, while another team in the "Personality Corner" builds a nearly identical robot to analyze character traits. They use the same blueprints, the same tools, and the same math, but they speak different languages and never share their secrets.

This survey paper is like a new librarian who decides to reorganize the entire library. Instead of sorting books by topic (Depression, Anxiety, Personality), they sort them by how the machine thinks.

Here is the simple breakdown of what this paper does, using some everyday analogies:

1. The New Sorting System: Four Ways Machines "Think"

The authors say, "Stop looking at what the AI is doing (e.g., diagnosing depression). Look at how it's doing it." They group all psychological AI tasks into four simple buckets:

The Traffic Cop (Classification):
- What it does: It looks at data and says, "Yes" or "No," or "Type A, B, or C."
- Analogy: Imagine a bouncer at a club. They look at your ID and decide: "You're in," "You're out," or "You're on the VIP list."
- Examples: Is this person depressed? Yes/No. Is this personality type "Extrovert"? Yes/No.
The Thermometer (Regression):
- What it does: It doesn't just say "Yes/No"; it gives you a number on a scale.
- Analogy: Instead of saying "It's hot," a thermometer says "It's 98.6 degrees." It measures how much of something there is.
- Examples: How severe is the anxiety? (Score: 12 out of 20). How much stress is the person feeling?
The Detective (Structured Relational):
- What it does: It connects the dots. It looks at a messy pile of clues and draws a map showing how they relate to each other.
- Analogy: Imagine a detective connecting photos on a corkboard with red string. "The insomnia is linked to the stress, which is linked to the job loss."
- Examples: Mapping how different symptoms of a disease feed into each other, or building a graph of a person's social support network.
The Therapist (Generative Interactive):
- What it does: It creates new content and talks back to you. It's not just analyzing; it's responding.
- Analogy: This is the difference between a medical textbook (static) and a conversation with a friend (dynamic). The AI listens to your story and writes a comforting reply or creates a personalized lesson plan.
- Examples: A chatbot that comforts someone in a crisis, or an AI tutor that adapts its teaching style based on your mood.

2. The Evolution: From "Custom Tools" to "Super-Brains"

The paper traces how these tools have changed over time, like upgrading from a hammer to a Swiss Army Knife, and finally to a magic wand.

Phase 1: The Hammer (Old School): Researchers had to build a custom tool for every single job. If they wanted to detect depression, they had to manually teach the computer what words to look for. It was slow and required a lot of data.
Phase 2: The Swiss Army Knife (Pre-trained Models): Then came models like BERT. These were like pre-trained tools that already knew how to read and understand language. Researchers just had to "tweak" them for their specific job. This was a huge leap forward.
Phase 3: The Magic Wand (Large Language Models - LLMs): Now, we have giants like GPT-4. These models are so smart they don't even need to be tweaked much. You can just talk to them ("Act like a therapist and help this person"), and they do it instantly. This is the "Zero-Shot" era, where the AI learns on the fly.

3. Why This Library is Hard to Organize (The Challenges)

The paper admits that organizing this library is tricky because human psychology is messy.

The "Black Box" Problem: We can get the AI to predict depression accurately, but we can't always explain why. It's like a doctor who says, "You're sick," but can't point to the specific symptom. In medicine, doctors need to know why to trust the diagnosis.
The "Subjective Label" Problem: If two human experts look at the same sad person, they might disagree on the diagnosis. If the AI is trained on their disagreement, it gets confused. The AI needs to understand that sometimes, there is no single "right" answer.
The Privacy Vault: Psychological data is the most sensitive stuff there is. It's like a diary of your deepest fears. We can't just dump this data into a public cloud. We need special locks (encryption) and ways to train AI without ever seeing the raw diary pages.
The Cultural Lens: A model trained on Americans might think "talking about feelings" is healthy. But in some cultures, talking about feelings is taboo, and sadness shows up as a headache instead. If the AI doesn't understand this cultural context, it will make terrible mistakes.

4. The Big Picture: Why Does This Matter?

The authors argue that by sorting these tasks by how they work (the four buckets above) instead of what they treat (depression vs. anxiety), we can finally share knowledge.

The Benefit: If a researcher figures out a great way to make the "Therapist" AI better at listening, that same trick can be used by the "Detective" AI to map relationships better.
The Goal: To move from a fragmented library where everyone reinvents the wheel, to a unified system where we can build better, faster, and safer tools to help people's mental health.

In short: This paper is a roadmap. It tells us that instead of building 1,000 different small bridges to cross the river of mental health, we should build one massive, sturdy bridge with four distinct lanes (Classification, Regression, Relational, Generative) that everyone can use together.

1. Problem Statement

The intersection of Artificial Intelligence (AI) and psychological science has experienced exponential growth, with publications rising from 859 in 2000 to nearly 30,000 in 2025. However, this rapid expansion has led to methodological fragmentation.

Domain-Centric Silos: Research is currently organized by application domains (e.g., clinical psychology, educational psychology, personality research) rather than computational characteristics.
Redundancy and Isolation: Similar computational techniques (e.g., attention mechanisms for depression detection vs. suicide risk prediction) are developed independently in isolated subfields, preventing knowledge transfer.
Unique Constraints: Psychological computing faces distinct challenges compared to standard AI, including label uncertainty (subjective clinical judgments), construct validity (measuring theoretical entities), privacy constraints (sensitive mental health data), and cross-cultural validity (cultural differences in symptom expression).

The paper argues that the field lacks a unified computational framework to organize these tasks, hindering systematic progress and the transfer of methodological innovations.

2. Methodology

The authors conducted a comprehensive survey of over 300 representative works spanning two technological eras:

Pre-trained Model (PLM) Era (2018–2022): Characterized by architectures like BERT and GPT-2, focusing on transfer learning and task-specific fine-tuning.
Large Language Model (LLM) Era (2023–Present): Characterized by models like GPT-4 and Claude, focusing on zero-shot learning, few-shot adaptation, and natural language interfaces.

Core Methodological Innovation:
Instead of organizing by application, the authors propose a Computational Taxonomy based on processing patterns and output types. They formalize psychological tasks as a tuple $T = \langle X, Y, F \rangle$ , where $X$ is the input space (multimodal data), $Y$ is the output space (psychological constructs), and $F$ is the computational function.

3. Key Contributions

A. A New Computational Taxonomy

The paper introduces a systematic taxonomy categorizing AI-driven psychological tasks into four fundamental types, transcending traditional domain boundaries:

Classification Tasks:
- Goal: Map inputs to discrete categorical labels (e.g., depression vs. no depression, Big Five personality types).
- Techniques: Discriminative pattern recognition using softmax heads on encoders (BERT, RoBERTa, ViT).
- Examples: Suicide risk prediction, ADHD detection, personality recognition.
Regression Tasks:
- Goal: Predict continuous numerical values quantifying psychological constructs (e.g., severity scores, stress levels).
- Techniques: Linear regression heads mapping inputs to continuous scales (e.g., PHQ-9 scores).
- Examples: Depression severity quantification, cognitive ability estimation, emotional valence/arousal rating.
Structured Relational Tasks:
- Goal: Extract and model complex relationships, networks, and temporal dynamics.
- Techniques: Graph Neural Networks (GNNs), Knowledge Graphs, and sequence modeling.
- Examples: Symptom co-occurrence networks, social support graphs, clinical pathway mining, temporal trajectory modeling.
Generative and Interactive Tasks:
- Goal: Create personalized psychological content and maintain dynamic user engagement.
- Techniques: Sequence-to-sequence models, dialogue management, and reinforcement learning.
- Examples: Therapeutic chatbots, crisis intervention systems, adaptive educational tutoring, personalized treatment plan generation.

B. Technical Evolution Analysis

The survey traces the shift in methodologies:

From Feature Engineering to Transfer Learning: Early PLM approaches relied on fine-tuning specific layers with limited labeled data.
From Fine-tuning to Prompting: LLMs introduced zero-shot and few-shot capabilities, reducing the need for massive labeled datasets and enabling rapid adaptation via prompt engineering and instruction tuning.
Multimodal Integration: The progression from single-modality (text-only) to integrated systems combining text, audio, video, and physiological signals (wearables).

C. Systematic Resource Survey

The paper provides a curated overview of:

Datasets: Categorized by task type and data source (Clinical/Lab, Longitudinal, Ecological/Naturalistic). It highlights the scarcity of high-quality, longitudinal, and culturally diverse datasets.
Evaluation Metrics: Moves beyond standard accuracy/F1 scores to include Construct Validity, Clinical Utility, Ecological Validity, and fairness metrics. It details specific metrics for each task type (e.g., CCC for regression, Graph Edit Distance for relational tasks, and human-rated counseling skills for generative tasks).

D. Challenge and Future Directions Analysis

The authors identify four critical challenges and propose future research directions:

Interpretability: Moving beyond "black box" predictions to theory-aligned explanations that clinicians can trust (e.g., linking features to DSM-5 criteria).
Label Uncertainty: Addressing the inherent ambiguity in psychological labels (inter-rater disagreement) rather than forcing consensus.
Privacy & Security: Developing robust privacy-preserving techniques (Federated Learning, Differential Privacy) for sensitive mental health data.
Cross-Cultural Validity: Mitigating Western-centric biases in training data and models to ensure global applicability.

4. Results and Findings

Paradigm Shift: The field has shifted from "psychology-centric" (small, controlled studies) to "computation-centric" (large-scale, data-driven approaches), with Computer Science publications growing exponentially compared to traditional psychology domains.
LLM Impact: LLMs have fundamentally transformed the field by enabling sophisticated psychological analysis with minimal task-specific training, though they introduce new challenges in hallucination and safety.
Methodological Gaps: Despite technical advances, there is a significant gap between research prototypes and clinical deployment due to issues in interpretability, lack of longitudinal data, and insufficient cross-cultural validation.
Taxonomy Utility: Organizing by computational patterns reveals that a regression model for anxiety severity shares more technical similarities with a regression model for cognitive decline than it does with a classification model for anxiety, facilitating cross-domain knowledge transfer.

5. Significance

Unifying Framework: This is the first survey to organize the fragmented field of AI-driven psychology based on computational processing patterns rather than application domains. This enables researchers to transfer solutions across subfields (e.g., applying a graph learning technique from social network analysis to symptom network modeling).
Bridging Theory and Practice: By emphasizing construct validity and clinical utility, the paper ensures that computational methods remain grounded in psychological theory and clinical reality, rather than just optimizing for predictive accuracy.
Roadmap for Deployment: The detailed analysis of challenges (privacy, bias, interpretability) and future directions (digital twins, closed-loop interventions, causal discovery) provides a clear roadmap for researchers and practitioners aiming to deploy AI in real-world mental health settings.
Resource Consolidation: The comprehensive catalog of datasets, metrics, and benchmarks serves as a critical reference for the community to standardize evaluation and improve reproducibility.

In conclusion, the paper argues that adopting a computational perspective is essential to accelerate progress in computational psychology, moving from isolated, domain-specific experiments to a cohesive, scalable, and clinically effective discipline.