BioLLMAgent: A Hybrid Framework with Enhanced Structural Interpretability for Simulating Human Decision-Making in Computational Psychiatry

TL;DR

Validated across reward-punishment learning and temporal discounting tasks, BioLLMAgent provides a structurally interpretable "computational sandbox" for testing mechanistic hypotheses and intervention strategies in psychiatric research.

Abstract

Computational psychiatry faces a fundamental trade-off: traditional reinforcement learning (RL) models offer interpretability but lack behavioral realism, while large language model (LLM) agents generate realistic behaviors but lack structural interpretability. We introduce BioLLMAgent, a novel hybrid framework that combines validated cognitive models with the generative capabilities of LLMs. The framework comprises three core components: (i) an Internal RL Engine for experience-driven value learning; (ii) an External LLM Shell for high-level cognitive strategies and therapeutic interventions; and (iii) a Decision Fusion Mechanism for integrating components via weighted utility. Comprehensive experiments on the Iowa Gambling Task (IGT) across six clinical and healthy datasets demonstrate that BioLLMAgent accurately reproduces human behavioral patterns while maintaining excellent parameter identifiability (correlations $>0.67$). Furthermore, the framework successfully simulates cognitive behavioral therapy (CBT) principles and reveals, through multi-agent dynamics, that community-wide educational interventions may outperform individual treatments. Validated across reward-punishment learning and temporal discounting tasks, BioLLMAgent provides a structurally interpretable "computational sandbox" for testing mechanistic hypotheses and intervention strategies in psychiatric research.

Figures (13)

• Figure 1: Network-level intervention effects. Top: PCA visualization of behavioral states under (a) No Intervention (0.700), (b) Targeted CBT (0.750), (c) Network Hub (0.630), (d) Community Education (0.950). Bottom: Network structures with health levels. Community Education achieves highest performance with cohesive behavioral convergence and uniform network-wide improvements.
• Figure 2: BioLLMAgent framework architecture integrating IGT Environment, Internal RL Engine (ORL model), External LLM Shell, and Decision Fusion mechanism. Parameter $\omega$ controls RL-LLM balance. The Internal RL Engine processes experience to generate Expected Value (EV), Expected Frequency (EF), and Perseveration (PS) utilities, while the External LLM Shell uses persona prompts to simulate complete IGT trials, generating probability distributions that are averaged and converted to static utility-scale priors $\Pi_{\text{util}}$.
• Figure 3: Behavioral trajectory validation using GPT-4o. Choice patterns for (a) Amphetamine, (b) Heroin, (c) Healthy controls, (d-f) Additional datasets. Lines: Pure ORL (red), BioLLMAgent neutral (green), noisy control (orange), human data (black).
• Figure 4: Behavioral trajectory validation using DeepSeek backend. Same setup as Fig. \ref{['fig:traj_gpt4o']}, demonstrating consistent performance across LLM backends with comparable trajectory tracking fidelity to GPT-4o, validating the generalizability of the hybrid approach.
• Figure 5: Fusion mechanism comparison across five alternatives. Violin plots show advantageous choice rate distributions for Linear (0.967, our choice), Multiplicative (0.991), Bayesian (0.982), Attention-based (0.954), and Gated (0.949) fusion methods. Linear balances performance with interpretability ($\omega$ as intuitive dosage parameter) and parsimony (single parameter vs. 3-5 in alternatives).