PEMAND: Persona-Enriched Multi-Agent Negotiation for Household Decision-Making

Yuran Sun; Mustafa Sameen; Yaotian Zhang; Rongguan Gu; Mrunal Vibhute; Chia-yu Wu; Yuanyuan Lei; Xilei Zhao

cs.AIApr 12, 2026

PEMAND: Persona-Enriched Multi-Agent Negotiation for Household Decision-Making

Authors: Yuran Sun, Mustafa Sameen, Yaotian Zhang, Rongguan Gu, Mrunal Vibhute, Chia-yu Wu, Yuanyuan Lei, Xilei Zhao

Organizations: University of Florida

Abstract

Modeling household-level decisions is central to many real-world applications, including trip planning, residential mobility and migration, disaster management, etc. Existing studies primarily rely on classical machine learning models with limited predictive capacity, while recent LLM-based approaches have yet to incorporate behavioral theory or intra-household interaction dynamics, both of which are essential for modeling realistic household decisions. To address these limitations, we propose Persona-Enriched Multi-Agent Negotiation for household Decision-making (PEMAND), a novel LLM-based framework that integrates behavioral theory into individualized, household-aware persona modeling and simulates household-level decision-making through structured multi-agent negotiation. Specifically, PEMAND transforms static sociodemographic attributes into coherent narrative profiles that explicitly encode household-level attitudes, subjective norms, and perceived behavioral controls, following our proposed Household-Aware Chain-of-Planned-Behavior (HA-CoPB) framework. Building on these theory-grounded personas, PEMAND captures real-world household decision negotiation via a structured two-phase multi-agent conversation framework with a novel persona-alignment control mechanism. Evaluated on national and regional household decision datasets across two major domains, including travel behavior and residential mobility, PEMAND consistently outperforms state-of-the-art benchmarks.

Figures & tables

Figure 1: Methodological Framework for PEMAND.

	Travel Domain (Regression)								Mobility Domain (Classification)
	NHTS 2017				Puget Sound 2023				PSID 2021–2023
Model	MAE $\downarrow$	RMSE $\downarrow$	sMAPE $\downarrow$	$\pm 2$ Acc $\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	sMAPE $\downarrow$	$\pm 2$ Acc $\uparrow$	Acc $\uparrow$	F1 $\uparrow$	Prec $\uparrow$	Rec $\uparrow$
Traditional ML
Linear / Logistic Reg. †	3.34	4.52	51.00	0.40	2.83	3.86	58.54	0.58	0.69	0.55	0.47	0.67
Random Forest	3.16	4.36	51.74	0.43	2.75	3.74	57.17	0.59	0.72	0.51	0.51	0.51
Gradient Boosting	3.07	4.27	51.87	0.44	2.76	3.78	57.47	0.58	0.69	0.55	0.48	0.67

Table 1: Main Household Prediction Performance. Comparison of ML and LLM baselines across travel-domain regression and mobility-domain classification. Bold denotes the best result within each dataset block.

\downarrow

\uparrow

indicate lower/higher is better. † Linear regression for NHTS/Puget Sound and logistic regression for PSID.

Figure 2: Structural Correlation Alignment. Diagonal-split heatmaps for demographic–opinion Spearman correlations in Travel (NHTS) and Mobility (PSID); upper-left triangle = human ground truth, lower-right = persona.

Model	MAE $\downarrow$	RMSE $\downarrow$	sMAPE $\downarrow$	$\pm 2$ Acc $\uparrow$
Init. Proposal w/o SFT	3.70	5.32	63.44	0.51
Init. Proposal w/ SFT	3.10	4.47	57.90	0.56
Refinement w/o RL & Moderator	3.05	4.32	56.02	0.58
Refinement w/o RL	2.59	3.91	53.80	0.64
Consensus Refinement	1.99	3.25	47.92	0.78

Table 2: Ablation Study on Consensus Refinement. Bold indicates best performance.

\downarrow

indicates lower is better,

\uparrow

indicates higher is better.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Variable Type	Code / Value	Meaning
Binary	0	No
Binary	1	Yes
Frequency	01	Never
Frequency	02	A few times a year
Frequency	03	A few times a month
Frequency	04	A few times a week

Appendix

Table 3: Coding schemes for binary, frequency, and income bracket variables.

	Llama-3.1-8B-Instruct				Gemma-3-27b-it				GPT-5_4-mini
Method	MAE $\downarrow$	RMSE $\downarrow$	sMAPE $\downarrow$	$\pm 2$ Acc $\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	sMAPE $\downarrow$	$\pm 2$ Acc $\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	sMAPE $\downarrow$	$\pm 2$ Acc $\uparrow$
Demographics-only	3.31	5.10	61.75	0.54	2.83	4.19	55.63	0.58	2.98	4.23	56.47	0.55
Persona	2.89	3.99	59.52	0.56	3.33	5.37	62.99	0.55	3.09	4.25	56.13	0.53
RAG	5.01	6.60	76.49	0.35	3.93	5.43	66.85	0.43	3.46	4.71	60.61	0.47
Self-consistency	2.89	4.04	58.30	0.57	3.01	4.27	61.70	0.55	2.91	4.07	54.34	0.57
PEMAND w/o ReST	2.59	3.91	53.80	0.64	2.41	3.64	49.54	0.68	2.32	3.78	47.87	0.63

Appendix

Table 4: LLM Baselines on Puget Sound 2023 (test). Comparison of Llama-3.1-8B-Instruct , Gemma-3-27b-it , and GPT-5_4-mini under the same baseline families, with a homogeneous backbone in each column (persona generation, member agents, moderator, and judge share the same model family). Bold is best among the four LLM baselines within each model block and metric column. PEMAND w/o ReST / PEMAND w ReST : multi-agent pipeline without or with QLoRA SFT on member models for that backbone column.

\downarrow

\uparrow

indicate lower / higher is better.

Figure 3: Robustness across homogeneous LLM backbones (Puget Sound test, MAE). Grouped bars compare baseline families under Llama-3.1-8B-Instruct, Gemma-3-27b-it, and GPT-5.4-mini. PEMAND achieves the lowest MAE on every backbone; stronger models reduce error for most families, though persona conditioning is not uniformly beneficial. PEMAND w ReST is omitted for GPT-5.4-mini (QLoRA training applied only to open-weight backbones; Table 4 ).

Placeholder	Travel Domain (NHTS/Puget)	Mobility Domain (PSID)
<target decision>	Total daily trip frequency	Relocated between 2021–2023
<shared constraints>	Vehicle deficit/bottlenecks	Homeownership lock-in
<agent role>	Travel Diary Participant	Head of Household/Spouse
<output format>	Integer (e.g., 4)	Boolean (True/False)

Appendix

Table 5: Dynamic prompt instantiation across domains.

Panel of Figure 4 — Figure 4: The Theory of Planned Behavior (TPB) and the proposed HA-CoPB adaptation. Our hierarchy mandates that shared resource barriers ( $\mathcal{PBC}$ ) override social obligations ( $\mathcal{SN}$ ), which in turn override individual preferences ( $\mathcal{A}$ ).

Variable	Raw Code Example	Natural Language Mapping
Demographics
R_AGE_IMP	$34$	"I am 34 years old."
R_SEX_IMP	$1$ (Male), $2$ (Female)	"I am a man/woman." (uses "boy/girl" if Age $<18$ )
EDUC	$5$ (Grad Degree)	"I have a graduate or professional degree."
HHFAMINC	$11$ ( $>\$ 200k$ )	"My household income is $200,000 or more."
Household

Appendix

Table 6: Travel Domain Variable Translation Logic. Raw codes from the travel survey are deterministically mapped to natural language atoms before narrative synthesis.

Variable	Raw Code Example	Natural Language Mapping
Demographics
AGE	$45$	"I am a 45-year-old man/woman."
EMPLOYMENT_STATUS	$2$ (Temporarily laid off)	"Regarding employment, I am temporarily laid off."
DEGREE_TYPE	$2$ (Bachelor’s)	"My highest degree is a bachelor’s degree."
Household
RELATION_TO_HEAD	$10$ (Reference Person)	"I am the reference person (head) of this household."

Appendix

Table 7: Residential Mobility (PSID) Variable Translation Logic. To avoid leaking the move-outcome, we explicitly enforce temporal labeling (e.g., "Prior-wave snapshot") and neutralize relationship categories.

Domain	Construct	Narrative Statement	HA-CoPB Implication
Travel	Tech-Savvy	"Learning new technologies is often frustrating for me. (Reverse)"	High PBC for app-based mobility and e-shopping substitution.
Travel	Polychronic	"I prefer to do one thing at a time. (Reverse)"	Tolerance for longer transit/ride-hail trips if productive; justifies complex trip chaining.
Travel	Materialistic	"I enjoy having a lot of luxury things."	Vehicle Preference: Correlates with private vehicle usage and higher discretionary shopping generation.
Travel	Pro-Car	"I like the idea of driving as a means of travel."	Resistance to mode switching; high utility for driving.
Travel	Pro-Transit	"I like the idea of public transit as a means of travel."	Increases probability of transit usage if PBC allows.
Travel	Wait Tolerant	"The trip itself is a waste of time. (Reverse)"	Lower disutility of travel time; longer commute acceptance.

Appendix

Table 8: Attitudinal Constructs mapped to their HA-CoPB theoretical implications across both domains.

	Travel Domain (NHTS)			Mobility Domain (PSID)
Metric	Health	Price	Place	Food Afford.	Life Sat.	K6 Distress
Accuracy (Exact)	38.1%	26.2%	27.7%	71.0%	41.4%	47.9%
Accuracy ( $\pm 1$ )	85.0%	64.3%	73.4%	94.7%	88.4%	87.7%
MAE	0.80	1.32	1.07	0.34	0.71	0.69
QWK (Kappa)	0.29	0.02	0.03	0.11	0.18	0.18
Wasserstein ( $W_{1}$ )	0.47	0.86	0.28	0.12	0.26	0.33

Appendix

Table 9: Comprehensive Perception Validation Results. Performance of PEMAND across all targeted perception probes in both domains.

Figure 5: Perception probe distributional fidelity. Normalized histograms comparing human ground truth (dark) vs. synthesized persona responses (green) for all six holdout probes. Top row: Travel domain (NHTS). Bottom row: Mobility domain (PSID).

	Travel (Puget Sound)			Mobility (PSID)
Dimension	Human	LLM (Before)	LLM (After)	Human	LLM (Before)	LLM (After)
	QWK	QWK	QWK	QWK	QWK	QWK
Persona Consistency	0.446	0.133	0.450	0.847	0.139	0.490
Constraint Adherence	0.626	0.188	0.533	0.517	0.162	0.471
Interaction Validity	0.491	0.272	0.527	0.500	0.236	0.538
Decision Quality	0.845	0.506	0.749	0.729	0.460	0.648

Appendix

Table 10: Moderator calibration inter-rater reliability. Average pairwise QWK by rubric dimension. Travel (Puget Sound) includes LLM–Human QWK before and after role-conditioned SFT; Mobility (PSID) reports human agreement only.

Variable	Description
HHVEHCNT*	Number of vehicles in household
HHFAMINC*	Household income category
HOMEOWN*	Home ownership status
URBRUR	Urban vs rural household location
RAIL	Rail availability in household area
TRAVDAY	Travel day type (weekday vs weekend)

Appendix

Table 11: Variable names and descriptions for NHTS 2017. Variables marked with “*” appear in both NHTS and Puget Sound.

Variable	Description
HHSIZE	Household size
NUM_CHILDREN_FU	Number of children in the family unit
AGE_YOUNGEST_CHILD	Age of the youngest child in the household
MARITAL_STATUS*	Marital status of the reference person
HOMEOWN*	Home ownership or housing tenure status
STATE_FIPS*	State location code

Appendix

Table 12: Processed PSID variables used for residential mobility prediction. Variables marked with * are treated as categorical variables in the preprocessing pipeline.

Human Behavior Simulation

Human behavior simulation models how individuals and groups decide, interact, move, and adapt in social, economic, educational, and physical settings.

See all 239 papers

HouseholdBench: Evaluating Large Language Models as Predictors of Household Economic Behavior
Oct 6, 2026Jin Huang, Diego Ferreras Garrucho, Yutong Xie +4LLM Evaluation LLM Fine-Tuning
University of Michigan, Ann Arbor · London School of Economics · MobLab Inc +1
Hierarchical Multi-Persona Induction from User Behavioral Logs: Learning Evidence-Grounded and Truthful Personas
Apr 28, 2026Nayoung Choi, Haeyu Jeong, Changbong Kim +2Personality Modeling in Language Models User Modeling
Department of Computer Science, Emory University, Atlanta, GA, USA · Naver Corporation, South Korea
Learning Dynamic User Personas from Implicit Interaction Streams via Iterative Refinement
Jul 29, 2026Haifeng WuLLM Personalization User Modeling
PayPal

Oct 6, 2026J/K move · Enter open · S save

HouseholdBench: Evaluating Large Language Models as Predictors of Household Economic Behavior

Jin Huang, Diego Ferreras Garrucho, Yutong Xie, Walter M. Yuan+3

University of Michigan, Ann Arbor · London School of Economics · MobLab Inc+1

LLM Evaluation LLM Fine-Tuning Forecasting Benchmarks Human Behavior Prediction

Open paper AskPDF arXiv

Large language models (LLMs) have the potential to meet a key goal in economics: a quantitative model of household decision making, across a variety of settings. Yet existing evaluations cover few surveys and outcomes, and do not study how households adjust to changing economic conditions. We introduce a new evaluation, HouseholdBench, which unites 6 U.S. household surveys and 32 prediction tasks spanning numeric, categorical and probabilistic outcomes, related to consumption, income, labor, expectations, and housing. Using past behavior, demographics and macroeconomic conditions, the tasks test whether LLMs predict behavior, including how households adjust to changes in various policies. We evaluate 13 proprietary and open-weight LLMs against a no-change baseline and a gradient-boosted tree model. Most LLMs outperform the no-change baseline, including for policy response tasks -- with the best model lowering error for numeric outcomes by 12.2%. Across most tasks, gradient-boosted trees rank first; leading proprietary LLMs approach their performance, but open-weight models lag. LLMs exhibit systematic over- and underprediction across different tasks. We identify methods that enable a 4 billion parameter open-weight model to match proprietary models' performance: fine-tuning and aggregating 16 predictions per observation. Improvements generalize to policy-response tasks, which are excluded from fine-tuning. We release our datasets, code, and leaderboard on our website: https://jn-huang.github.io/householdbench

Abstract

Human Behavior Simulation

Explore similar work

HouseholdBench: Evaluating Large Language Models as Predictors of Household Economic Behavior

Hierarchical Multi-Persona Induction from User Behavioral Logs: Learning Evidence-Grounded and Truthful Personas

Learning Dynamic User Personas from Implicit Interaction Streams via Iterative Refinement