Curious Now

Story

MALLES: A Multi-agent LLMs-based Economic Sandbox with Consumer Preference Alignment

ComputingMath & Economics

Key takeaway

Researchers developed a simulated economic system using AI agents to better understand how consumer preferences shape real-world markets.

Read the paper

Quick Explainer

MALLES is a novel economic simulation platform that leverages large language models (LLMs) to simulate realistic consumer behaviors across different product categories. The key innovations include: 1) Economically aligning LLMs through post-training on transaction data to enable effective knowledge transfer, 2) Employing multi-agent discussions to distribute the cognitive load when processing complex product information, and 3) Implementing stabilizing mechanisms like mean-field approximations and attention-based input reconstruction to model the dynamic interplay between product environments and customer populations. These techniques allow MALLES to generate high-fidelity consumer decisions that closely match real-world patterns, even in data-sparse product categories.

Deep Dive

MALLES: A Multi-agent LLMs-based Economic Sandbox with Consumer Preference Alignment

Overview

The paper introduces MALLES, a Multi-Agent Large Language Model-based Economic Sandbox that leverages cross-category transaction data and collaborative multi-agent reasoning to simulate heterogeneous consumer behaviors with high fidelity. Key innovations include:

  • Using extensive transaction records to economically align LLMs via post-training, enabling effective knowledge transfer across categories and mitigating data sparsity.
  • Employing a multi-agent discussion mechanism to distribute cognitive load when processing long-context product information.
  • Implementing mean-field stabilization to model the dynamic interplay between product environments and customer populations.
  • Incorporating input augmentation with attention control to reconstruct partial observations and reduce biases from hidden variables.

Methodology

Problem Formulation

The simulation environment has real decision makers (retail customers, wholesalers, manufacturers, etc.) indexed by i. Simulated decisions are â_i, actual behaviors are a_i, observable inputs are X_i^obs, hidden factors are X_i^hid, and personality/preference parameters are ρ_i. The true decision function is a_i = D(X_i^obs, X_i^hid, ρ_i), and the agent decides via â_i = D̂(X_i^obs, Z_i, ρ̂_i; θ), where Z_i approximates X_i^hid and θ are trainable parameters.

Retail and Wholesale Simulation

  • Retail simulation trains LLMs to be economically aware via post-training on transaction data, mitigating per-category data sparsity.
  • Wholesale decisions involve multi-agent dialogue between wholesalers, marketing personnel, and manufacturers to derive profit-oriented purchasing formulas.
  • Symbolic regression is integrated to discover compact mathematical expressions of decision patterns.

Stabilization and Calibration

  • Style parameters ρ_i and specialized submodules inject commercial insights (e.g. discount sensitivity, loss aversion) to align agent decisions with human preferences.
  • Consistency regularization and multi-sampling stabilize decisions.
  • A mean-field mechanism iteratively approximates the real distribution of customer responses.
  • Calibration functions f and reweighting w(X, Z) post-process outputs or adjust prompts/parameters.
  • Conflict feedback and bottleneck detection enhance robustness.

Data & Experimental Setup

  • The dataset contains product information, transaction records, and dialogue logs from real-world retail and wholesale interactions.
  • Experiments evaluate MALLES in retail scenarios, comparing to existing LLM-based economic simulators.
  • Metrics include purchase decision hit rate, prediction quantity error, and decision stability.

Results

  • MALLES achieves significant improvements in product selection accuracy, purchase quantity prediction, and simulation stability compared to existing baselines.
  • Ablation studies confirm the importance of post-training alignment, multi-agent discussion, and mean-field mechanisms.
  • Cross-category training demonstrates the ability to generalize to data-sparse categories and new products.

Limitations and Ethical Considerations

  • Further investigation into cross-modal alignment, tighter coupling between micro and macro dynamics, and ethical guidelines for personalized behavioral prediction are needed.

Source