Signal Scaling Multi-Agent Systems with Prospect-State Propagation

Summary

The paper points out that LLM-based multi-agent systems, in periodically compressing state to cut token consumption, tend to discard meaningful information such as agent behavioral trajectories in economic simulations. The authors observe that agent heterogeneity gradually fades as a simulation runs longer. To address this they propose PspMAS, inspired by prospect theory, which splits each agent's micro state into a lightweight prospect state and an LLM-based semantic state. The prospect state uses a lightweight, parallelizable propagator to record psychological traces and continually reinject heterogeneity into the system. The semantic state draws on the LLM's perception, reasoning, planning, and decision-making abilities to preserve expressive richness. Working together, the two components provide a scalable approach to multi-agent simulation.