PhysicsAgentABM: Physics-Guided Generative Agent-Based Modeling

Large language model (LLM)-based multi-agent systems enable expressive agent reasoning but are expensive to scale and poorly calibrated for timestep-aligned state-transition simulation, while classical agent-based models (ABMs) offer interpretability but struggle to integrate rich individual-level signals and non-stationary behaviors. We propose PhysicsAgentABM, which shifts inference to behaviorally coherent agent clusters: state-specialized symbolic agents encode mechanistic transition priors, a multimodal neural transition model captures temporal and interaction dynamics, and uncertainty-aware epistemic fusion yields calibrated cluster-level transition distributions. Individual agents then stochastically realize transitions under local constraints, decoupling population inference from entity-level variability. We further introduce ANCHOR, an LLM agent-driven clustering strategy based on cross-contextual behavioral responses and a novel contrastive loss, reducing LLM calls by up to 6-8 times. Experiments across public health, finance, and social sciences show consistent gains in event-time accuracy and calibration over mechanistic, neural, and LLM baselines. By re-architecting generative ABM around population-level inference with uncertainty-aware neuro-symbolic fusion, PhysicsAgentABM establishes a new paradigm for scalable and calibrated simulation with LLMs.

💡 Research Summary

The paper addresses two fundamental shortcomings of current generative agent‑based modeling (GABM): (1) the high computational cost and poor calibration of large language model (LLM)‑driven multi‑agent simulations that reason at the individual level, and (2) the limited ability of classical agent‑based models (ABMs) to incorporate rich, multimodal individual signals and non‑stationary dynamics. To bridge this gap, the authors propose PhysicsAgentABM, a hierarchical neuro‑symbolic framework that shifts inference from individual agents to behaviorally coherent clusters.

In PhysicsAgentABM, the population is represented as an interaction graph G = (V, E) with an adjacency matrix A. Agents are partitioned into M clusters {Ck} using a novel LLM‑driven clustering method called ANCHOR. ANCHOR proceeds in three stages: (i) structural‑semantic initialization via GraphSAGE embeddings and spectral clustering to obtain coarse clusters; (ii) behavioral motif discovery through short‑horizon domain‑specific mini‑simulations that record reasoning‑action traces, which are embedded and clustered to define motif profiles; (iii) anchor‑guided contrastive refinement, where each coarse cluster selects an “anchor” agent whose motif profile is most representative. The anchor evaluates compatibility scores qij for other agents using LLM reasoning, and a joint loss combining contrastive learning and KL divergence aligns learned similarities with these scores, producing refined clusters that respect both topology and behavior.

Within each cluster, two complementary pathways estimate transition hazards for all admissible state changes: a symbolic pathway and a neural pathway. The symbolic pathway employs state‑specialized symbolic agents that encode mechanistic priors (e.g., epidemiological equations, policy rules) and produce hazard estimates λsym along with epistemic uncertainty usym. The neural pathway uses a multimodal encoder (tabular, temporal, and graph inputs) to generate neural hazards λneu and uncertainty uneu. Both hazard estimates are fused via uncertainty‑aware epistemic fusion:

 λfused = (c_sym · λsym + c_neu · λneu) / (c_sym + c_neu)

where confidence weights (c_sym, c_neu) are learned by a lightweight MLP that calibrates each source’s reliability. The fused hazards constitute a calibrated, population‑consistent transition prior for the cluster.

Individual agents then realize state transitions stochastically by modulating the cluster‑level prior with their own attributes, local neighborhood statistics from A, and personal memory. This decouples costly population‑level inference from per‑agent variability while preserving heterogeneity.

The authors evaluate PhysicsAgentABM on three public datasets spanning public health (infectious disease spread), finance (stock market regimes), and social science (opinion diffusion). Baselines include mechanistic ABMs, pure neural state‑space models, and existing LLM‑based GABMs. Metrics comprise event‑time accuracy (e.g., mean absolute error of outbreak peaks) and calibration quality (Expected Calibration Error). PhysicsAgentABM consistently outperforms baselines, achieving 5–12 % improvements in timing accuracy and 30–45 % reductions in calibration error. Notably, the advantage grows for long‑horizon simulations where temporal drift is severe. The ANCHOR clustering reduces LLM calls by an average factor of 7× while maintaining or improving cluster cohesion, demonstrating substantial computational savings.

Key contributions are: (1) a cluster‑level neuro‑symbolic inference architecture that integrates mechanistic knowledge and data‑driven dynamics with explicit uncertainty modeling; (2) ANCHOR, the first LLM‑anchored clustering method that treats abstraction as a semantic control problem, enabling transition‑faithful clustering; (3) a calibrated, timestep‑aligned paradigm that separates population inference from agent realization, improving scalability and robustness.

Limitations include the need to pre‑specify the number of clusters, sensitivity of anchor selection and LLM prompt design, and potential dominance of the neural pathway in highly non‑linear interaction regimes. Future work is suggested on automatic cluster‑number determination, prompt optimization, and incorporation of reinforcement‑learning‑based policy updates to enhance adaptability in dynamic environments.