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

Zuo Fei; Kezhi Wang; Xiaomin Chen; Yizhou Huang

cs.AIMar 5, 2026

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

Authors: Zuo Fei, Kezhi Wang, Xiaomin Chen, Yizhou Huang

Organizations: School of Computer Science and Technology, East China Normal University, 3663 North Zhongshan Road, Shanghai 200062, China · Department of Computer Science, Brunel University of London, UB8 3PH, UK · Department of Computer Science, University of Reading, RG6 6UR, UK

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 & tables

Fig. 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.

Fig. 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 $\Pi_{\text{util}}$ .

Param.	Range	Cognitive Interpretation
$A_{rew}$	[0, 1]	Reward learning rate
$A_{pun}$	[0, 1]	Punishment learning rate
$K$	[0, 5]	Forgetting/perseveration decay
$\beta_{F}$	$\mathbb{R}$	Frequency weight (wins vs. losses)
$\beta_{P}$	$\mathbb{R}$	Perseveration weight
$\theta$	[0, $\infty$ )	Inverse temperature (exploration)

TABLE I: ORL Model Parameters and Their Cognitive Interpretations

Dataset	N	Population
Ahn et al. (HC) [ 6 ]	48	Healthy Controls
Horstmann et al. [ 27 ]	162	Healthy Controls
Kjome et al. [ 37 ]	19	Healthy Controls
Maia & McClelland [ 38 ]	40	Healthy Controls
Ahn et al. (Amphetamine) [ 6 ]	38	Amphetamine Users
Ahn et al. (Heroin) [ 6 ]	43	Heroin Users

TABLE II: Breakdown of Datasets Used in the Current Study

Fig. 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).

Fig. 4: Behavioral trajectory validation using DeepSeek backend. Same setup as Fig. 3 , demonstrating consistent performance across LLM backends with comparable trajectory tracking fidelity to GPT-4o, validating the generalizability of the hybrid approach.

Dataset	GPT-4o						DeepSeek
	Neutral			Noisy			Neutral			Noisy
	r	MAE	B.Diff	r	MAE	B.Diff	r	MAE	B.Diff	r	MAE	B.Diff
Ahn et al. (Amphetamine) [ 6 ]	0.958	0.073	0.005	0.957	0.146	0.108	0.944	0.090	0.055	0.919	0.119	0.054
Ahn et al. (Heroin) [ 6 ]	0.954	0.066	0.001	0.938	0.133	0.097	0.912	0.097	0.060	0.874	0.133	0.069
Ahn et al. (HC) [ 6 ]	0.943	0.084	0.005	0.927	0.140	0.089	0.928	0.108	0.068	0.897	0.161	0.081
Horstmann et al. [ 27 ]	0.981	0.073	0.001	0.979	0.126	0.095	0.959	0.111	0.067	0.919	0.167	0.081

TABLE III: Model Consistency Analysis: Traditional RL vs. BioLLMAgent Performance

Fig. 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).

Fig. 6: Parameter recovery analysis. Scatter plots of true vs. recovered parameters: (a) $A_{rew}$ , (b) $A_{pun}$ , (c) $K$ , (d) $\beta_{F}$ , (e) $\beta_{P}$ . Bottom: Pearson correlations with confidence intervals. Diagonal line indicates perfect recovery.

Dataset	$A_{rew}$	$A_{pun}$	$K$	$\beta_{F}$	$\beta_{P}$
Amphetamine [ 6 ]	0.709	0.696	0.719	0.863	0.928
Heroin [ 6 ]	0.615	0.680	0.751	0.841	0.938
HC [ 6 ]	0.701	0.684	0.664	0.909	0.905
Average	0.676	0.687	0.711	0.871	0.924

TABLE IV: Parameter Recovery Performance Summary

Model	Instr.	Chi-sq	KL	Std
		$p$	Div.	Dev.
GPT-4o	w/o	0.105	0.124	0.125
GPT-4o	w/	0.988	0.002	0.017
DeepSeek	w/o	$<$ 0.001	0.615	0.282
DeepSeek	w/	0.604	0.031	0.062
Llama-3.2 (3b)	w/o	$<$ 0.001	1.386	0.433

TABLE V: Prompt Ablation Study: Effect of Uniform Distribution Instruction

Fig. 7: Choice distributions from prompt ablation study. Action probabilities for (a) GPT-4o, (b) DeepSeek-Chat, (c) Llama-3.2, (d) Gemma-3.27b with/without uniform instruction. Red line: perfect uniform (0.25). Larger models show better compliance.

Rank	Model	Temp.	MSD	SD	N Datasets
1	GPT-4o	0.5	1.00	0.65	6
2	GPT-4o	0.0	1.12	1.23	6
3	GPT-4o	1.0	1.47	0.92	6
4	DeepSeek	0.0	1.59	1.97	6
5	DeepSeek	1.0	1.61	1.56	6
6	DeepSeek	0.5	2.01	2.22	6

TABLE VI: Performance Ranking of LLMs within the BioLLMAgent Framework

Fig. 8: LLM performance across temperature settings (0.0, 0.5, 1.0). Performance as

1/\text{MSD}

. Models: GPT-4o (blue), DeepSeek-Chat (orange), Llama-3.2 (green), Gemma-3.27b (red).

Fig. 9: Omega sensitivity analysis. Panel A: advantageous choice rate across $\omega\in[0,1]$ for clinical and healthy populations. Panel B: performance degradation with incorrect priors. Vertical line at $\omega=0.25$ balances quality gains (+8-10%) with robustness to prior errors ( $<$ 5% for healthy, $<$ 15% for clinical).

Fig. 10: Population-specific omega sensitivity. Panel A: performance trajectories across $\omega$ for healthy (near-ceiling, flat) vs. clinical (lower baseline, steep) populations. Clinical groups show larger sensitivity to external guidance (Heroin: +38% from $\omega=0$ to 0.25). Panel B: LLM guidance benefit at $\omega=0.25$ confirms clinical populations derive greater benefits.

Fig. 11: CBT intervention effects on decision trajectories. (a-f) Six datasets showing human choices (black), ORL model (red), neutral prior (green), CBT intervention (orange). CBT promotes advantageous learning, strongest in clinical populations.

Fig. 12: Network topology robustness analysis. Row 1: three network architectures (Watts-Strogatz, Barabási-Albert, Erdős-Rényi). Row 2: intervention effectiveness across topologies. Intervention ranking preserved across all networks (Community Education $>$ Targeted CBT $>$ Hub $>$ Random), demonstrating coverage breadth matters more than network structure.

Fig. 13: Framework generalization to Delay Discounting task. Panel A: delayed choice rate across $\omega$ for three populations. Clinical groups show lower baselines but steeper $\omega$ -response curves, mirroring IGT patterns. Panel B: CBT guidance benefit at $\omega=0.25$ . Clinical populations show largest relative improvements (+60%), validating cross-task generalization of fusion mechanism.

Explore similar work

CCD-CBT: Multi-Agent Therapeutic Interaction for CBT Guided by Cognitive Conceptualization Diagram
Apr 8, 2026Chang Liu, Changsheng Ma, Yongfeng Tao +2Conversational Agents Large Language Model-Based Role-Play Simulation
School of Information Science and Engineering, Lanzhou University
PsyCIDRA: A Dual-Agent Framework for Psychiatric Interviewing and Diagnostic Reasoning
Oct 5, 2026Milad Mohammadi, Fatemeh Akrami Shamsabadi, Zahra Mohseni +4LLM Agents Clinical Reasoning
School of Electrical and Computer Engineering, University of Tehran · Faculty of Psychology and Educational Sciences, University of Tehran · School of Medicine, Yazd (Shahid Sadoghi) University of Medical Sciences
LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals
Nov 15, 2024Joon Sung Park, Carolyn Q. Zou, Jonne Kamphorst +8Personality Modeling in Language Models Human Behavior Simulation
1Computer Science Department, Stanford University; Stanford, CA, 94305, USA. · Department of Communication Studies, Northwestern University; Evanston, IL 60208, USA. · 7CDSP and CEE, Sciences Po; Paris, 75007, France. +4

Apr 8, 2026J/K move · Enter open · S save

CCD-CBT: Multi-Agent Therapeutic Interaction for CBT Guided by Cognitive Conceptualization Diagram

Chang Liu, Changsheng Ma, Yongfeng Tao, Bin Hu+1

School of Information Science and Engineering, Lanzhou University

Conversational Agents Large Language Model-Based Role-Play Simulation Cognitive Behavioral Therapy Multi-Agent LLM Systems Cognitive Modeling

Open paper AskPDF arXiv

Large language models show potential for scalable mental-health support by simulating Cognitive Behavioral Therapy (CBT) counselors. However, existing methods often rely on static cognitive profiles and omniscient single-agent simulation, failing to capture the dynamic, information-asymmetric nature of real therapy. We introduce CCD-CBT, a multi-agent framework that shifts CBT simulation along two axes: 1) from a static to a dynamically reconstructed Cognitive Conceptualization Diagram (CCD), updated by a dedicated Control Agent, and 2) from omniscient to information-asymmetric interaction, where the Therapist Agent must reason from inferred client states. We release CCDCHAT, a synthetic multi-turn CBT dataset generated under this framework. Evaluations with clinical scales and expert therapists show that models fine-tuned on CCDCHAT outperform strong baselines in both counseling fidelity and positive-affect enhancement, with ablations confirming the necessity of dynamic CCD guidance and asymmetric agent design. Our work offers a new paradigm for building theory-grounded, clinically-plausible conversational agents.

Abstract

Explore similar work

CCD-CBT: Multi-Agent Therapeutic Interaction for CBT Guided by Cognitive Conceptualization Diagram

PsyCIDRA: A Dual-Agent Framework for Psychiatric Interviewing and Diagnostic Reasoning

LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals