Scaling Multi-Agent Systems with Prospect-State Propagation

AI summary

Overview

• Research area: LLM-based multi-agent systems (MAS) for macroeconomic simulation, combining agent-based modeling with behavioral economics (prospect theory).
• Technical level: Intermediate (familiarity with LLM agents and basic economics helps, but the core argument is conceptual).
• Scope: The paper diagnoses a heterogeneity-collapse failure mode in scaling LLM multi-agent economic simulations and proposes a compact "Prospect State" branch as a fix.

What This Paper Is About

LLM-driven economic simulations reduce token costs by periodically summarizing or "reflecting" on agent histories, but this compression slowly makes agents behave alike. The authors show that as a result, adding more agents fails to improve simulation realism, and they propose PspMAS (Prospect-State Propagation for Multi-Agent Systems), which keeps each agent's micro state split into a tiny psychological vector updated cheaply and a richer textual state updated only every K steps. The goal is to scale to large agent populations without losing the behavioral diversity that drives macroeconomic phenomena such as the subprime mortgage crisis.

Key Contributions

1. The paper identifies and names a heterogeneity-loss problem in LLM-based MAS for economic simulation: periodic semantic compression gradually reduces the effective number of behaviorally distinct agents (the authors call the resulting homogenization "action convergence").
2. It incorporates prospect theory into agent state updating, defining a compact Prospect-State Vector of reference point, loss aversion, and probability weighting, updated by a Markovian, lightweight, parallelizable propagator.
3. It introduces a scalable LLM-based multi-agent simulation framework (PspMAS) that decouples each agent's micro state into a Prospect State and a Semantic State, plus a spectral-entropy metric for quantifying population heterogeneity.
4. It provides a scaling analysis against the SaMAS baseline showing that nominal agent count is not equivalent to effective system scale.

Main Findings

• Scaling fails in the strong baseline (SaMAS): Going from N = 100 to N = 500 improved Volatility Realism (VR) by only 0.2 points (81.8% to 82.0%), while normalized fused diversity D_norm dropped 9.9 points (43.2% to 33.3%). The paper calls this a clear failure of the expected scaling law.
• PspMAS scales better: VR rose from 82.9% to 85.3% while diversity stayed much higher (61.9% at N = 100 and 56.9% at N = 500). At N = 500, PspMAS exceeded SaMAS by 3.3 points in VR and 23.6 points in D_norm.
• Diversity versus agent count: D_norm (%) for Summary / Reflection / PspMAS was 68.2 / 71.3 / 73.8 at N = 20; 40.3 / 55.6 / 67.5 at N = 50; 21.2 / 33.5 / 59.6 at N = 200; and 20.1 / 30.5 / 56.5 at N = 500. The Summary System saturates near N = 200, the Reflection System scales sublinearly, and PspMAS stays high at N = 500.
• Diversity versus model size: Scaling from Qwen3-8B to Qwen3-32B raised D_norm from 32.1% to 36.5% for the Reflection System and from 58.2% to 62.5% for PspMAS.
• Diversity versus simulation horizon: At horizons 50 and 200, D_norm (%) was 28.5 and 18.5 for the Summary System, 42.5 and 30.2 for the Reflection System, and 61.2 and 57.8 for PspMAS. The authors describe PspMAS diversity as remarkably stable over long horizons.
• RQ2 headline claim: At N = 500, PspMAS maintains D_norm > 0.56 while the Summary System saturates near N = 200 at roughly 0.21.
• Note on reported numbers: The paper reports D_norm of 56.9% for PspMAS at N = 500 in the SaMAS comparison table but 56.5% at N = 500 in the agent-count table; the paper does not comment on this difference.

Methodology in Plain English

Each simulated agent makes two decisions per step: whether to work (0 or 1) and what fraction of available funds to spend (between 0 and 1). Instead of storing a long written profile, PspMAS splits the agent's micro state into three parts: a three-dimensional Prospect-State Vector (reference point, loss-aversion coefficient, probability-weighting parameter), a textual Semantic State, and ordinary economic bookkeeping variables such as wealth, debt, wage, tax paid, consumption, and employment status.

The Prospect State updates every step through deterministic arithmetic rules inspired by prospect theory: the reference point drifts toward wealth and toward a cohort average subject to social comparison; loss aversion moves toward a population mean during crises plus a small noise term; and the probability-weighting parameter rises with recent experience intensity and drops under shocks. A projection step keeps values inside a predefined valid domain. The Semantic State is refreshed only every K steps by an LLM that sees the current Prospect State, accounting variables, recent raw events, and the previous summary.

After all agents act, a market-clearing function aggregates intended consumption and labor supply, rations goods proportionally if demand exceeds production, adjusts prices and wages with the imbalance, collects progressive taxes and redistributes revenue equally, and sets the interest rate annually using a Taylor rule.

To measure diversity, the authors build three population matrices: a Trajectory Matrix from the last L = 12 steps of actions using an RBF kernel, a Prospect Matrix from normalized Prospect-State Vectors using an RBF kernel, and a fused Heterogeneity Matrix combining the two with a balance coefficient. Eigenvalues of the fused matrix give spectral entropy, whose exponential is the effective diversity D_eff, divided by N to give D_norm. Collapse toward similar behavior makes the fused matrix low-rank and D_eff small.

Implementation used Qwen3-family models, temperature 0.7 for decision calls and 0.2 for summarization, update interval K = 5, propagator coefficients α_r = 0.1, β_r = 0.05, η_λ = 0.001, η_γ = 0.005, ζ = 0.1, σ_λ = 0.02, threshold E_thr = 10, trajectory window L = 12, and fusion coefficient α = 0.5. Scaling comparisons used Qwen3-32B with the same evaluation protocol, horizon, and comparable token budgets. The paper states that the environment is a closed economy with GDP, inflation, and unemployment in the macro state; it does not give the full definition or computation of Volatility Realism in the provided content.

Why This Matters

• Impact on research: The paper reframes scaling of LLM multi-agent systems as a heterogeneity-preservation problem rather than a raw agent-count problem, and argues that token-efficient state compression can systematically erase the micro-level signals from which macro-level crises emerge. It also imports a classical behavioral-economics theory into agent state design.
• Real-world applications (as suggested or motivated in the paper):
  • Financial market regulation, including earlier detection of localized distress such as the rising personal mortgage debt that preceded the subprime mortgage crisis.
  • Climate policy negotiation.
  • Organizational behavior simulation.
  • Macroeconomic forecasting and policy analysis using LLM-driven agents.
• Industry relevance: Because the Prospect State avoids long textual descriptions and the Semantic State is refreshed only every K steps, the authors report substantially reduced LLM inference token consumption, which matters for anyone running large-scale simulations under real compute budgets. Providers and users of LLM simulation tooling could adopt the two-branch state design as a drop-in replacement for summary- or reflection-based memory.

Future Directions

1. Extend PspMAS beyond the simplified closed economy to capture open-economy effects, institutional constraints, and real-world policy frictions.
2. Test robustness across different model families, prompting strategies, and calibration settings, given that LLM decisions may inherit biases from the underlying models.
3. Address the difficulty of disentangling cultural, political, and ideological biases absorbed by LLMs from internet, academic, and policy text, which the authors flag as a serious challenge for "value-neutral" policy insights.
4. Apply the framework to other socio-economic domains where bounded rationality and heterogeneous decision-making are central, such as financial market regulation, climate policy negotiation, and organizational behavior simulation.

Target Audience

Researchers and practitioners working on LLM-based multi-agent systems, agent-based macroeconomic modeling, and social simulation; engineers concerned with inference cost and scaling of agent populations; and behavioral economists interested in how prospect theory can be operationalized inside computational agents. Readers without background in agent-based modeling or LLM agent architectures will need supplementary reading on market clearing, reflection-based memory, and prospect theory.

Authors’ abstract

Current LLM-based multi-agent systems (MAS) periodically compress intermediate states to reduce inference-time token consumption, thereby attempting to incorporate more agents. However, naive scaling strategies face challenges. For example, in economic simulations, large-scale MAS typically discard semantically rich economic states, i.e., agent behavioral trajectories, which are key drivers of macroeconomic fluctuations. In this paper, we reveal a phenomenon in which agent heterogeneity gradually decreases during simulation, and propose Prospect-State Propagation for Multi-Agent Systems (PspMAS). Inspired by prospect theory, PspMAS decouples each agent's micro state into a compact Prospect State and an expressive Semantic State. The former records psychological traces through a lightweight, parallelizable propagator and continuously injects heterogeneity into the system. The latter leverages the strong perception, reasoning, planning, and decision-making abilities of LLMs. These two components work complementarily, providing a scalable LLM-based multi-agent simulation solution.