Source-linked AI summary
PhysicsAgentABM: Physics-Guided Generative Agent-Based Modeling
Kavana Venkatesh, Yinhan He, Jundong Li, Jiaming Cui
TL;DR
Generative ABMs struggle to combine scalable, calibrated timestep simulation with rich individual and contextual signals. PhysicsAgentABM addresses this by inferring transitions for behaviorally coherent clusters through uncertainty-aware neuro-symbolic fusion, with stochastic individual realization; across diverse domains, it reports calibrated, temporally aligned dynamics over existing approaches.
Problem
Existing generative ABMs face costly per-agent inference, limited integration of multimodal and non-stationary signals, and gaps in group-level inference and principled symbolic-neural combination.
Method
PhysicsAgentABM performs cluster-level inference using state-specialized symbolic and multimodal neural pathways with uncertainty-aware fusion, while agents stochastically realize transitions under local constraints; ANCHOR clusters agents by cross-contextual behavioral responses.
Results
Across public health, finance, and social sciences, PhysicsAgentABM consistently yields calibrated, temporally aligned dynamics while preserving heterogeneous individual realization.
Takeaways & Limitations
The framework provides a system-level foundation for generative simulation under real-world complexity by shifting inference from isolated agents to adaptive clusters.
Abstract
from arXiv · showhide
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.
1. Introduction
Complex systems require scalable models that combine mechanistic structure, data-driven dynamics, symbolic context, and uncertainty. Existing ABMs and learned approaches each leave important gaps in adaptability, calibration, or inference level.
- Complex systems evolve through state-dependent interactions shaped by local, group-level, and broader contextual signals.Accurate simulation must jointly represent mechanistic structure, data-driven dynamics, symbolic context, and uncertainty while remaining scalable and interpretable.
- Classical ABMs offer interpretable mechanistic transitions but rely on static hand-crafted rules and coarse calibration.These constraints limit adaptation to heterogeneous populations, multimodal signals, and non-stationary regimes.
- Learned transition models improve flexibility but can obscure mechanistic structure and exhibit poor uncertainty calibration.
- Existing generative ABMs perform inference at the individual-agent level despite strong group-level, institutional, and contextual forces.
- Symbolic reasoning and neural learning are combined heuristically rather than through a principled integration.
D Parellel Reasoning
PhysicsAgentABM organizes simulation around cluster-level inference using symbolic and neural pathways, then realizes transitions stochastically at the individual level. ANCHOR supplies behaviorally coherent abstraction for this architecture.
- The architecture performs uncertainty-aware fusion before stochastic agent-level realization.
- Cluster-level predictions define probabilistic transition priors, while individual agents realize transitions stochastically under local attributes and neighborhood context.This preserves individual heterogeneity while maintaining population-level coherence.
- Symbolic and neural pathways are treated as complementary sources for transition inference.The architecture combines mechanistic constraints with multimodal neural transition modeling before agent-level realization.
- PhysicsAgentABM uses hierarchical cluster-level inference with state-specialized symbolic and neural reasoning.The framework is described as a neuro-symbolic approach that shifts inference away from isolated agents.
- ANCHOR is an LLM-agent-anchored clustering method that treats abstraction as a semantic control problem for transition-faithful simulation.
2. Related Work
Related work spans LLM-based agents, classical and LLM-enhanced ABMs, and graph clustering. PhysicsAgentABM targets their limitations by aligning clustering with behavioral semantics and downstream transitions.
- LLM-based systems provide language-based reasoning and coordination through role-specialized and large-scale multi-agent frameworks.
- ABMs model interacting individuals and emergent collective behavior across health, social, and systemic-risk applications.
- LLM-enhanced ABMs enrich agent decision-making but can incur costly per-agent inference and lack principled population-level multimodal integration.
- Conventional graph clustering typically emphasizes structural objectives while ignoring node attributes, temporal dynamics, and behavioral semantics.
- ANCHOR uses LLMs as semantic controllers of abstraction, aligning cross-contextual behavioral probing with downstream state-transition modeling.
3. Methodology
PhysicsAgentABM models interacting populations through behaviorally coherent clusters, combining structural, behavioral, symbolic, and neural signals to infer shared transition dynamics. Individual agents then realize these transitions stochastically under local constraints, while ANCHOR refines clusters using behavioral similarity and contextual compatibility.
- Population and cluster representation: The population is represented as an interaction graph whose agents occupy application-dependent discrete states and are partitioned into behaviorally similar clusters.Clusters capture similar responses to contextual stimuli, while neighborhood statistics provide scalable interaction information.
- Cluster-level transition inference: Cluster-level symbolic and neural pathways estimate hazards over valid transitions, and uncertainty-aware epistemic fusion produces population-consistent transition priors.The symbolic pathway uses cluster context and state-specialized reasoning, while the neural pathway uses aggregated tabular, temporal, and graph inputs.
- ANCHOR clustering: ANCHOR initializes coarse clusters from GraphSAGE structural-semantic embeddings and agent attributes without invoking agent reasoning.Spectral clustering on the combined representations supplies a structural prior.
- ANCHOR clustering: ANCHOR discovers behavioral motifs from short-horizon diagnostic simulations, summarizes agents by motif-frequency profiles, and selects representative anchor agents.Reasoning-action traces are embedded and clustered, allowing behavioral patterns across situations such as home and community mobility to guide refinement.
- ANCHOR clustering: Anchor-guided contrastive refinement combines learned representations with LLM-based regime-compatibility judgments to align clusters by functional decision logic and behavioral similarity.The contrastive loss separates regimes while its alignment term matches learned similarity to anchor-defined compatibility; final representations combine structural, motif, and profile information.
- Cluster adaptation: Boundary agents are reassigned using motif-guided pull scores, while redundant clusters are merged and heterogeneous clusters are split based on motif divergence and entropy.Cluster quality is monitored with modularity, silhouette score, and motif coherence, with refinement re-invoked after degradation.
- Agent-level realization: Individual agents modulate fused cluster hazards with attributes and neighborhood interactions, then sample next states through normalized competing-risk transitions.The distribution includes temporal persistence, and synchronous realization preserves individual heterogeneity under a shared population regime.
4. Experiments
PhysicsAgentABM is evaluated across epidemiology, finance, and social attention using rolling-window forecasts and event-time metrics. It consistently improves temporal alignment, event discrimination, and calibration, with qualitative evidence of coherent population-level dynamics under volatility and policy shifts.
- Experimental Setup: Experiments cover epidemiology, finance, and social attention, evaluating structured-network agents against population-level ground-truth signals under regime shifts.The protocol uses partially observable domains with latent population dynamics and discrete agent states.
- Experimental Setup: Rolling-window evaluation uses 28-day training or calibration windows followed by 7-day forecasts, emphasizing event timing, regime coherence, and calibration.Models are retrained or recalibrated at every step, and metrics are computed only on event-bearing trajectories.
- Quantitative Evaluation: PhysicsAgentABM achieves the best EETE, ET-F1, NLL, and Brier results in every evaluated domain.The comparison spans mechanistic, neural, LLM-based, and hybrid baselines.
- Epidemiology: PhysicsAgentABM tracks epidemic peaks and post-policy declines more closely than Rule-ABM, GNN-LSTM, and LLM-MAS in rolling-window COVID-19 forecasts.Symbolic reasoning handles policy and contact constraints, neural prediction captures intra-cluster dynamics, and epistemic fusion reweights them over time.
- Finance: In finance, the model assigns 60–75% probability mass to realized daily market regimes while reducing confidence during volatile periods and maintaining smooth weekly alignment.It also recovers a regime distribution closer to the empirical structure than rule-based ABMs.
- Social Attention: With 250 agents, the model infers a smooth S-shaped attention trajectory that captures emergence, amplification, and fatigue rather than reproducing transient pageview spikes.Normalized Wikipedia pageviews provide the noisy macro-level observation signal.
- Policy Shock Case Study: During Singapore’s Circuit Breaker, symbolic inference is upweighted at policy onset, then fusion rebalances toward neural prediction as dynamics stabilize.The case study reports Brier = 0.16, a 76% reduction versus LLM baselines.
5. Cost and Scalability Analysis
PhysicsAgentABM reduces computational demands by centralizing symbolic reasoning at the cluster level while retaining lightweight, parallel individual realization. Scaling and ablation results support cluster-level inference as the basis for efficient, calibrated simulation.
- Cost Reduction: Cluster-level symbolic inference and sparse agent-level LLM usage reduce cost and latency while preserving behavioral heterogeneity and predictive fidelity.Agents otherwise realize transitions through fused cluster hazards, neighborhood states, and local memory.
- Cost Reduction: 6.7× fewer API calls reduce requests from 8,250 to 1,233 relative to a flat GABM baseline.The reduction is reported for the epidemiological setting with 1,000 agents and four clusters.
- Cost Reduction: 2.9× and 4.7× token reductions lower per-timestep usage from 2.3M to 0.79M at α=1.0 and 0.49M at α=0.6.Under GPT-4o-mini pricing, cost falls from $0.48 to $0.23 and $0.14 per timestep.
- Cost Reduction: 7.5–12.5× runtime speedups reduce wall-clock time from ∼300 s to ∼40 s and ∼24 s per timestep.Predictive ET-F1 remains fixed at 0.81 across the reported configurations.
- Scalability: As population size increases, ET-F1 remains stable or improves slightly while per-timestep cost and latency scale near linearly with N.The reported scaling reflects cluster-governed symbolic reasoning and lightweight, parallel individual realization.
- Architecture Ablation: Removing clustering causes severe temporal misalignment and prohibitive latency, while disabling epistemic fusion produces overconfident neural-only or regime-fragile symbolic-only variants.State specialization and meta-coordination are also necessary for performance and efficiency.
- ANCHOR Ablation: Removing motifs or contrastive alignment degrades behavioral coherence, and disabling boundary optimization fragments clusters near regime transitions.The full ANCHOR model is Pareto-dominant across separability, structural coherence, and contextual behavioral similarity.
6. Conclusion
PhysicsAgentABM shifts generative ABM toward population-level inference through ANCHOR’s multi-stage behavioral clustering and cluster-level transition modeling. The framework aims to preserve heterogeneous individual realization while producing calibrated, temporally aligned dynamics.
- PhysicsAgentABM shifts generative ABM from isolated agent simulation to population-level inference, while preserving heterogeneous individual realization.
- Cluster-level inference supports calibrated dynamics while individual agents stochastically realize transitions under local constraints.
- After Singapore’s Circuit Breaker, PhysicsAgentABM captures infection decline and accelerated recovery in temporal alignment with ground truth.
- ANCHOR combines structural clustering, behavioral motif extraction, semantic separation, and agent-driven contrastive refinement in a four-stage pipeline.
B.2. Additional Ablations
ANCHOR’s ablations show that behavioral clustering captures context-dependent adaptation beyond structural connectivity. Its stages progressively refine coarse graph groupings into compact clusters aligned with distinct epidemiological transition regimes.
- ANCHOR separates agents by context-sensitive adaptation patterns that purely structural clustering does not capture.
- Clusters display distinct response speed, volatility, avoidance, isolation, compliance sensitivity, and social reactivity across home, work, and community contexts.
- ANCHOR’s four stages progress from graph-structured groupings to behavioral motifs, contrastive refinement, and final hybrid clusters.
- The final hybrid clusters are compact, well-separated, and aligned with distinct epidemiological state-transition regimes.
C. PhysicsAgentABM Calibration Analysis
PhysicsAgentABM produces calibrated uncertainty across epidemiology, finance, and social diffusion despite domain-specific noise and regime shifts. Calibration remains reliable while its tightness varies with signal-to-noise structure and regime persistence.
- PhysicsAgentABM’s uncertainty-aware hierarchical design avoids overconfident instance-level predictions under noisy or shifting dynamics.
- Epistemic fusion reallocates weight toward symbolic reasoning when regime uncertainty is high.
- Epidemiology exhibits near-ideal calibration, whereas finance and social diffusion show greater dispersion under faster regime turnover and higher exogenous noise.
- PhysicsAgentABM preserves monotonic reliability across all evaluated domains, indicating stable uncertainty ranking despite varying calibration tightness.
D. Symbolic Reasoning Pathway
The symbolic reasoning pathway uses hierarchical MetaAgents and StateAgents to infer cluster-level transition hazards under uncertainty, while EntityAgents realize those probabilistic dynamics individually. Its protocol conditions on current and historical information without using future observations and preserves uncertainty in outputs.
- Each domain uses one MetaAgent per cluster and multiple StateAgents coordinated through structured tool calls.
- MetaAgents model population-level state dynamics and estimate stochastic hazards for admissible transitions at the cluster level.
- Entity Agents stochastically realize population dynamics using interaction, environmental context, and probabilistic transition structure.
- Inter-cluster signals modulate transition hazards or spillover risk as soft coupling signals without assumed deterministic causality.
- The pathway conditions inference only on information available up to the current timestep and prohibits future observations, labels, or outcomes.
- State Agents provide transition probability or hazard estimates with mechanistic reasoning and uncertainty qualifiers.
F.1. Task Overview: Population-Level State Inference under Partial Observability
PhysicsAgentABM formulates learning as population-level inference of latent regime dynamics from indirect, noisy signals and interaction structure, rather than pointwise prediction of individual actions or observables. The shared abstraction spans epidemiological, financial, and social systems while preserving domain-specific state meanings.
- Shared task abstraction: The task recovers and forecasts temporal evolution of latent state distributions from heterogeneous, partially observable signals generated by interacting agents.Inputs include aggregate behavioral measurements, contextual exogenous variables, and weighted interaction networks.
- Latent state representation: Latent states are discrete behavioral regimes that evolve stochastically, with system state represented as cluster-level distributions rather than individual labels.Latent agent states are never directly observed during training or evaluation.
- Modeling objective: The objective is to recover coherent population regimes, track their temporal evolution, and anticipate regime transitions with calibrated uncertainty.The formulation explicitly excludes predicting individual actions, asset prices, engagement counts, or pointwise accuracy on observed signals.
- Domain-specific objectives: Epidemiology targets population disease progression and compartment-transition timing, finance targets latent market-sentiment regimes, and social attention targets emergence, saturation, and decay of engagement states.Finance and social-attention tasks explicitly avoid forecasting prices, returns, pageviews, or raw activity quantities.
- Supervision and observability: Ground-truth construction depends on observability: epidemiological clinical records support deterministic or semi-deterministic reconstruction, whereas financial and social latent states are inherently unobservable individually.The supplied passage truncates the description of structured inference for the latter domains.
F.2. Dataset Overview
The datasets share a population-level experimental structure: heterogeneous agents interact on explicit networks, produce daily observations, and evolve through discrete latent regimes under partial observability. Their evaluation includes both gradual and abrupt transitions to test temporal alignment, detection, and uncertainty calibration.
- Dataset structure: All datasets support cluster-level inference of latent regime dynamics from heterogeneous agent populations, explicit interaction networks, and daily observations.Supervision is defined at the cluster–time level rather than for individual agents.
- Temporal coverage and dynamics: All datasets use daily temporal resolution and include gradual transitions alongside abrupt shifts from interventions, market shocks, or viral attention events.These conditions assess temporal alignment, transition detection, and uncertainty calibration.
- Shared experimental framing: Each dataset contains a fixed heterogeneous agent population, an explicit interaction graph, and discrete latent states evolving over time.Agents have static attributes and dynamic latent states; domain-specific identities differ across epidemiology, finance, and social systems.
- Cross-domain evaluation: The unified construction supports consistent evaluation across epidemiological, financial, and social systems despite differences in domain semantics and data sources.Domain differences arise from how nodes, edges, weights, and latent states are interpreted.
- Population and network construction: Interaction graphs are weighted, remain fixed over the evaluation window unless stated otherwise, and encode contact, influence, or correlation patterns.Networks are designed for structural heterogeneity, contextual relevance, and population-scale coherence.
- State supervision: Epidemiological latent states use clinical records for semi-deterministic reconstruction, including S,E,I,R,D assignments based on infection, recovery, and death dates.The supplied passage truncates the incubation-period specification after describing exposure assignment before infection.