Graph-Based Alternatives to LLMs for Human Simulation¶
Conference: ACL 2026 arXiv: 2511.02135 Code: GitHub Area: Graph Learning / Human Behavior Simulation Keywords: GNN, Human Simulation, Link Prediction, Heterogeneous Graph, Survey Prediction
TL;DR¶
GEMS models closed-form human behavior simulation as link prediction on heterogeneous graphs with three node types (subgroups, individuals, choices) and two bidirectional relations, matching or surpassing strong LLM baselines across three datasets and three evaluation settings while using 1000x fewer parameters.
Method¶
Key Designs¶
-
Heterogeneous Graph Construction and Link Prediction: Individuals use uniform features (non-identifiable); GNN learns node embeddings through relation-aware message passing; decoder computes \(p(c|u,q) = \text{softmax}(\text{Dot}(z_u^O, z_c^O) / \tau)\).
-
Three Evaluation Settings: Imputation (missing answers), new individual prediction, and new question prediction — covering core application scenarios.
-
LLM-to-GNN Projection Layer (Setting 3 only): Linear projection from frozen LLM hidden states to GNN embedding space for unseen questions.
Key Experimental Results¶
| Method | OpinionQA | Twin-2K | Dunning-Kruger |
|---|---|---|---|
| Few-shot FT (best LLM) | 55.98 | 66.36 | 57.21 |
| GEMS (SAGE) | 57.00 | 66.62 | 57.89 |
Highlights & Insights¶
- Core insight: closed-form human simulation is essentially a recommendation system problem; relational structure matters more than language understanding
- GEMS can be trained from scratch on domain data, avoiding LLM pretraining data leakage and bias concerns
Rating¶
- Novelty: ⭐⭐⭐⭐
- Experimental Thoroughness: ⭐⭐⭐⭐⭐
- Writing Quality: ⭐⭐⭐⭐
- Value: ⭐⭐⭐⭐