GASim: A Graph-Accelerated Hybrid Framework for Social Simulation

GASim: A Graph-Accelerated Hybrid Framework for Social Simulation

Introduction

The paper "GASim: A Graph-Accelerated Hybrid Framework for Social Simulation" (2605.07692) addresses core scalability bottlenecks in LLM-driven multi-agent social simulations by introducing a novel, graph-based acceleration paradigm. It departs from earlier hybrid frameworks which combined LLM-based agents for opinion leaders with classic ABM-driven agents for the general population, but were hampered by high latency from memory retrieval and sequential model execution. GASim advances large-scale social simulation through a unified approach encompassing graph-optimized memory for LLM-based core agents, parallel graph-neural updating for ordinary agents, and dynamic, entropy-based agent role assignment.

Hybrid Scheme Acceleration via Graph Methods

Graph-Optimized Memory (GOM) for Core Agents

GASim replaces expensive LLM-in-the-loop memory retrieval for core agents with GOM, a sparse and structured memory architecture. Each core agent maintains a graph whose nodes represent historical neighbor messages, weighted by semantic similarity and opinion polarity. The retrieval process is cast as a convex optimization problem, balancing query relevance, inter-memory coherence, and retrieval diversity. The update, anchored in the normalized Laplacian of the memory graph, is efficiently approximated by iterative graph propagation, shifting complexity from LLM forward passes or matrix inversions to lightweight, sparse tensor operations. Empirically, GOM outperforms previous LLM-centric and vector-based memory retrieval models, achieving 71.56% accuracy (LLM-as-judge) on the LoCoMo benchmark and outperforming direct memory selection baselines by substantial margins on single-hop, multi-hop, and temporal queries.

Graph Message Passing (GMP) for Ordinary Agents

Ordinary agents' sequential ABM opinion updates are superseded by GMP, which models the crowd as a graph whose features aggregate static agent attributes (from BERT-encoded profiles) and dynamic, neighborhood-driven opinion states. GMP employs a two-layer GAT to parallelize opinion evolution in a single forward pass, capturing higher-order interaction patterns, non-linear dependencies, and homophily/echo-chamber effects in a scalable way. This generalizes traditional ABMs through fine-grained real-world data alignment and graph-edge-aware reasoning.

Entropy-Driven Grouping (EDG)

Prior hybrid frameworks used static, degree-based heuristics for identifying core agents, missing the emergent, context-dependent nature of opinion leaders. GASim instead relies on EDG, which computes the local information entropy of each agent's neighborhood opinion distribution at each timestep. Agents embedded in information-diverse neighborhoods are selected as cores. EDG is shown to dynamically correlate with network centrality, capturing temporal emergence of influential nodes and adaptively partitioning the social graph as event narratives shift.

Empirical Evaluation

Scalability and Cost Efficiency

GASim demonstrates ~10x reduction in end-to-end runtime (9.94× overall; 16.39× for core, 27.49× for ordinary agents) compared to HiSim (the previous hybrid baseline), simulating 10,000 nodes over 30 steps on real social datasets with total token consumption reduced to 20% of HiSim and 0.4% of a full-LLM system. As agent populations scale, GASim's cost advantage increases due to the graph acceleration's sublinear scaling properties.

Public Opinion Trend Alignment

GASim's trend alignment with real social media public opinion curves exhibits substantial improvements over both classic ABMs (HK, Lorenz, RA) and prior LLM-based methods (SOD, HiSim). For three diverse real-world topics, the framework achieves lowest ABias (mean error <1%), highest or second-best Pearson correlation, and superior Fréchet distance for geometric similarity. Qualitative curve visualizations show that only GASim consistently tracks real opinion evolution during complex event surges, outperforming all baselines, which either drift toward initial priors or overfit to one-sided shifts.

Ablation Analysis

Module removals confirm that GOM, GMP, and EDG each yield unique, non-redundant fidelity and efficiency contributions. Notably, disabling GOM increases bias and decreases correlation (–30%), disabling GMP sharply degrades all performance metrics, and removing EDG increases error variance by 47.3%, destabilizing trend consistency.

Memory Architecture Robustness

Head-to-head on the LoCoMo memory retrieval benchmark, GOM achieves state-of-the-art across all categories except open-domain questions (where reliance on canonical LLM world knowledge outperforms graph-local methods). However, GASim favors diverse, context-grounded memory retrieval over regurgitating generic information, which preserves simulation heterogeneity—a crucial factor for veridical social modeling.

Theoretical and Practical Implications

By formalizing agent memory retrieval as convex optimization over signed, opinion-weighted graphs and introducing fast Laplacian-regularized message aggregation, GASim provides a theoretically principled and computationally efficient architecture for large-scale, heterogeneous-agent simulation. The modularity of GOM, GMP, and EDG enables flexible extension to further agent heterogeneity, temporal network shifts, and explicit opinion leader tracking. Practically, the framework supports high-fidelity, low-latency simulation—essential for both retrospective analysis of complex social phenomena and potential real-time scenario planning.

Limitations and Future Directions

Two main limitations persist: (1) reliance on LLM-generated textual opinions, which can introduce synthetic or biased patterns due to LLM pretraining, and (2) the exclusion of multimodal (image/video) signals, which are known to influence public opinion but are not yet integrated in the simulation loop. Possible future research directions include integrating multimodal data for hybrid opinion modeling, improving adversarial robustness against LLM bias propagation, online co-evolution of agent roles and topologies, and adaptation for intervention and counterfactual policy simulation.

Conclusion

GASim articulates a mathematically rigorous, empirically validated, and highly efficient hybrid framework for social simulation. By leveraging graph-optimized memory, parallel graph-neural updating, and entropy-driven dynamic role assignment, GASim substantially advances the scalability, accuracy, and interpretability envelope for agent-based social analysis. This approach establishes a new baseline for large-scale simulation methodologies and provides a foundation for future work on multi-agent modeling and interactive, high-fidelity digital social ecosystems.