Scaling Multi-Agent Systems with Prospect-State Propagation

1 Introduction

• Representative-agent models eliminate distributional heterogeneity, while rule-based agent models fail to capture psychological processes that drive macroeconomic dynamics.Source: §1 · pp. 1–2
• Repeated semantic compression reduces behaviorally distinct agents, attenuating macroeconomic fluctuations and potentially erasing local distress mechanisms through which crises emerge.The paper names this gradual behavioral homogenization action convergence.Source: §1 · p. 2
• Localized risks can accumulate among a small subset of agents before aggregate indicators reveal them, so compressing those microstates can dilute early crisis signals.The paper gives highly indebted agents with average wages as an example in the buildup to the subprime mortgage crisis.Source: §1 · p. 2
• PspMAS decouples each agent’s microstate into a compact Prospect State and an expressive Semantic State, combining prospect-theory variables with LLM capabilities.The Prospect State tracks reference points, loss aversion, and probability weighting; prospect theory describes evaluation of gains and losses relative to a reference point.Source: §1 · pp. 1–2
• The paper presents heterogeneity preservation and a scalable LLM-based multi-agent simulation solution as contributions.Source: §1 · pp. 2–3

2.2 PspMAS

• PspMAS pairs a Markovian Prospect State, updated by lightweight propagation, with a Semantic State refreshed periodically by an LLM.The branches preserve compact psychological traces and support expressive natural-language reasoning.Source: §2 · p. 3, §2.2 · pp. 3–4

2.1 Preliminary

• Agents perceive, reason, plan, and decide through LLMs, interacting with one another and the environment to construct the macroeconomy.The simulation represents N heterogeneous agents whose actions influence later perceptions and decisions.Source: §2.1 · p. 3

2.2 PspMAS

• Each agent chooses whether to work and what fraction of available funds to spend, while market clearing aggregates labor supply and intended consumption into updated economic conditions.The environment records realized outcomes, adjusts prices and wages to imbalances, and feeds macro conditions back into subsequent agent states.Source: §2.2 · pp. 3–5
• Prospect-State propagation uses prospect-theory features, including reference dependence, loss aversion, and probability weighting, to update psychological state from immediate outcomes.The Markovian propagator uses lightweight, parallelizable rules, and projects the updated state onto a predefined valid domain.Source: §2.2 · pp. 3–4
• The Semantic State summarizes recent events and reflections every K steps, using the current Prospect State, accounting variables, recent events, and its prior summary.The two branches operate in parallel: the Prospect State evolves continuously at low cost, while periodic LLM refreshes provide richer behavioral patterns.Source: §2.2 · p. 4
• Independent, lightweight state updates and periodic Semantic-State refreshes reduce costly LLM calls and support simulations with thousands of agents under practical computational budgets.The implementation uses Qwen3-family models, with update interval K = 5 and a trajectory window L = 12.Source: §2.2 · p. 5, §2.3 · p. 6
• Heterogeneity measurement fuses similarities in recent action trajectories with similarities in path-dependent Prospect States, then uses the fused matrix’s spectrum to assess effective diversity.The Prospect Matrix distinguishes psychological states even when agents’ recent actions appear similar; low effective diversity corresponds to highly similar behavior and Prospect States.Source: §2.2 · p. 5

3 Experiments

• The experiments ask why increasing agent count fails to scale effectively in a state-of-the-art economic MAS and whether PspMAS preserves heterogeneity at scale.Source: §3 · p. 6

3.1 Experimental Setting

• The closed-economy simulation has agents make labor-supply and consumption decisions at each step, with GDP, inflation, and unemployment in the macro state.The comparisons include SaMAS, summary and reflection systems, and PspMAS; they assess SaMAS and PspMAS at N = 100 and N = 500 using VR and normalized fused diversity under the same horizon and comparable token budgets.Source: §3.1 · p. 6, §3.2 · p. 6

3.2 RQ1: Scaling Analysis of SaMAS and PspMAS

• A fivefold increase in SaMAS population, from N = 100 to N = 500, raises VR by only 0.2 points while Dnorm falls by 9.9 points.PspMAS instead increases VR from 82.9% to 85.3% while retaining Dnorm of 61.9% at N = 100 and 56.9% at N = 500; at N = 500 it exceeds SaMAS by 3.3 points in VR and 23.6 points in Dnorm.Source: §3.2 · pp. 6–7

3.3 RQ2: Heterogeneity Preservation at Scale

• As agent count varies, SUMMARY SYSTEM saturates near N ≈ 200, while PspMAS maintains high fused diversity even at N = 500.The reported saturation is attributed to repeated semantic compression homogenizing recent action trajectories and prospect-state traces; REFLECTION SYSTEM shows sublinear scaling.Source: §3.3 · p. 7
• Scaling from Qwen3-8B to Qwen3-32B raises Dnorm from 32.1% to 36.5% for REFLECTION SYSTEM and from 58.2% to 62.5% for PspMAS.Source: §3.3 · p. 7
• Across simulation horizons 50 and 200, PspMAS maintains stable fused diversity, preserving behavioral trajectories and path-dependent prospect-state differences.SUMMARY SYSTEM degrades rapidly as compression accumulates information loss, while REFLECTION SYSTEM degrades more gradually but still experiences cumulative drift.Source: §3.3 · p. 7

4 Related Work

• Traditional DSGE and VAR models analyze fluctuations and policy transmission top-down, whereas agent-based modeling simulates interacting micro-level individuals to reproduce macroeconomic phenomena.LLM-driven agents add scenario-based interaction capabilities to agent-based modeling.Source: §4 · p. 7
• Periodic semantic summarization of long economic histories risks attenuating behavioral trajectories and progressively homogenizing agents.Source: §4 · p. 7

N SUMMARY REFLECTION PspMAS

• Recent LLM-driven agent efforts have also been applied to macroeconomic forecasting and reported intricate emergent macro-level phenomena.Source: N SUMMARY REFLECTION PspMAS · p. 8

• PspMAS incorporates prospect theory into economic-state updating in a macroeconomic simulation multi-agent system.Source: §5 · p. 8
• Each agent’s microstate is decoupled into a Prospect State that injects heterogeneity and a Semantic State that generates human-like actions.Source: §5 · p. 8
• Together, the two components reduce LLM inference-token consumption, make effective scaling possible, and achieve strong simulation performance.Source: §5 · p. 8
• The work offers insights into addressing LLM-driven multi-agent-system challenges through classic behavioral-economics theories.Source: §5 · p. 8
• Future work could extend PspMAS to financial-market regulation, climate-policy negotiation, and organizational-behavior simulation.Source: §5 · p. 8

• The simulations use a simplified closed economy, limiting representation of open-economy effects, institutional constraints, and real-world policy frictions.Source: Limitations · p. 8
• LLM-based decisions may inherit underlying-model biases; the authors call for robustness checks across model families, prompting strategies, and calibration settings.Source: Limitations · p. 9
• Cultural, political, and ideological value orientations cannot currently be completely disentangled, challenging applications seeking value-neutral policy insights.Source: Limitations · p. 9