💬 NLP

Graph-Based Alternatives to LLMs for Human Simulation

This paper introduces Graph-basEd Models for Human Simulation (GEMS), a graph neural network approach that formulates close-ended human simulation tasks as link prediction to match or surpass large language models in performance while utilizing three orders of magnitude fewer parameters.

Original authors: Joseph Suh, Suhong Moon, Serina Chang

Published 2026-04-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Joseph Suh, Suhong Moon, Serina Chang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict how a specific person will vote in an election, what they'll buy at a store, or how they'll answer a survey question. For a long time, the tech world's go-to solution has been Large Language Models (LLMs)—the same powerful AI brains that write poetry, code, and chat with you.

The thinking was: "To predict human behavior, we need a model that understands human language perfectly."

But a team of researchers from UC Berkeley asked a simple question: "Do we really need a super-computer brain for this, or can we use a simpler, smarter map?"

Their answer is GEMS (Graph-basEd Models for Human Simulation). Here is the paper explained in simple terms, using some everyday analogies.

1. The Old Way: The "Encyclopedia" Approach (LLMs)

Think of an LLM as a giant, encyclopedic librarian who has read almost every book ever written.

How it works: When you ask, "What would a 30-year-old teacher from Ohio think about this tax law?" the librarian searches their massive memory, tries to recall similar people, and writes a long, reasoned answer.
The Problem: This librarian is expensive (it costs a fortune to run), slow, and sometimes hallucinates (makes things up). Also, because they learned from the whole internet, they might accidentally bring in biases or "leak" private data they shouldn't know.

2. The New Way: The "Social Network Map" (GEMS)

The researchers realized that human choices aren't just about what we say; they are about who we are and who we connect with.

Instead of a librarian, imagine a giant, living map (a graph).

The Nodes (The Dots): On this map, you have dots for People, dots for Groups (like "Teenagers" or "Dog Owners"), and dots for Choices (like "Vote Yes" or "Buy Brand A").
The Edges (The Lines): Lines connect people to the groups they belong to, and lines connect people to the choices they've made in the past.

The Magic of the Map:
If you want to predict what a new person will choose, you don't need to read a book about them. You just look at the map:

Where do they stand? (They are connected to the "Young" group).
Who else is in that group?
What choices did those people make?

The model learns that "People in Group X usually pick Option Y." It's like predicting a movie you'll like not by reading a review, but by seeing that your friends with similar tastes loved it.

3. The Three Scenarios (The "Test Drive")

The researchers tested this map against the "Librarian" (LLM) in three different situations:

Scenario A: The Missing Puzzle Piece (Imputation)
- The Task: We have a survey, but someone forgot to answer Question 5. What did they pick?
- The Result: The Map (GEMS) guessed correctly just as often as the Librarian, but it did it 1,000 times faster and used 1,000 times less computer power. It didn't even need to read the text of the question; it just looked at the patterns of connections.
Scenario B: The Stranger (New Individuals)
- The Task: A brand new person walks in. We know their age and job, but we've never seen them before. What will they choose?
- The Result: The Map looked at the "groups" this new person belongs to and predicted their choice. It matched the Librarian's accuracy perfectly.
Scenario C: The New Question (New Questions)
- The Task: We have a brand new survey question that no one has ever seen. How will people answer?
- The Result: This is the hardest part. Here, the Map needed a tiny "translator" to understand the text of the new question. Even with this extra step, the Map performed just as well as the Librarian, but it was still much cheaper and faster.

4. Why This Matters (The "Aha!" Moment)

💰 It's Cheap and Fast
Running the Librarian (LLM) is like hiring a team of 100 PhDs to solve a simple math problem. The Map (GEMS) is like using a well-organized spreadsheet. The researchers found that GEMS uses 100 times less computing power and has 1,000 times fewer parameters (the "brain cells" of the AI) than the LLMs, yet gets the same results.

🔍 It's Transparent (No Magic)
With the Librarian, you often don't know why it gave an answer. It's a "black box."
With the Map, you can literally see the connections. You can say, "Ah, the model predicted this because the person is connected to the 'High Income' group, and 80% of that group chose Option A." This makes it much easier to trust and check for bias.

🛡️ It's Safer
LLMs are trained on the entire internet, which means they might have "memorized" private data or learned harmful stereotypes. GEMS is trained from scratch only on the specific data you give it. It doesn't carry the baggage of the whole internet.

The Bottom Line

The paper argues that for many tasks where we just need to predict a choice from a list (like surveys, voting, or test answers), we don't need a giant, expensive AI brain.

We just need a smart map that understands how people and choices connect. It's a reminder that sometimes, the simplest tool—the one that looks at relationships and patterns—is the most powerful one of all.

In short: You don't need a supercomputer to predict human behavior; you just need to understand the social network.

1. Problem Definition

The paper addresses the task of close-ended human simulation, where the goal is to predict an individual's response from a fixed set of options (e.g., survey answers, multiple-choice test responses, or voting choices). While Large Language Models (LLMs) have become the dominant approach for these tasks, they suffer from significant drawbacks:

High Computational Cost: Training and inference require massive resources.
Opacity: Pretraining processes are opaque, leading to concerns about data leakage (evaluation data appearing in training sets) and the amplification of social biases.
Questionable Necessity: It is unclear if the complex generative capabilities of LLMs are necessary for tasks that are fundamentally structured as selecting from a discrete set of options.

The authors ask: Can a smaller, more transparent model class match or surpass LLMs in close-ended simulation tasks?

2. Methodology: GEMS

The authors propose GEMS (Graph-basEd Models for Human Simulation), a framework that reformulates close-ended simulation as a link prediction problem on a heterogeneous graph.

Graph Representation

The task is modeled as a graph $\mathcal{G}$ with three node types:

Individuals ( $U$ ): Representing specific people.
Subgroups ( $S$ ): Groups of individuals sharing features (e.g., demographics like age, gender, political affiliation).
Choices ( $C$ ): The answer options for specific questions. These are structured as disjoint sets for each question.

Edges:

Membership: Connects individuals to their relevant subgroups.
Response: Connects individuals to the choices they selected in observed data.

Architecture

GEMS consists of three main components:

Graph Encoder: Uses a Heterogeneous Graph Neural Network (GNN) (e.g., RGCN, GAT, GraphSAGE) to perform relation-aware message passing. It learns embeddings for individuals, subgroups, and choices based on the relational structure of observed responses.
- Note: In Settings 1 and 2, the model relies solely on the relational structure and learnable feature tables, without using textual inputs for choices.
Link Prediction Decoder: Predicts the probability of a link between an individual and a choice option using a dot-product followed by a softmax classifier.
LLM-to-GNN Projection (For Setting 3): To handle new questions (where choice nodes have no edges yet), a lightweight linear projection layer is trained to map frozen LLM hidden states (text features of the choices) into the GNN embedding space. This allows the GNN to generalize to unseen text without fine-tuning the LLM.

Training Objective

The model is trained end-to-end using self-supervised link prediction. A subset of response edges is masked, and the model is supervised to reconstruct these missing links. This approach requires no explicit negative sampling, as the softmax normalization over all options for a given question implicitly handles negative examples.

3. Experimental Setup

The authors evaluated GEMS across three datasets and three distinct simulation settings:

Datasets:

OPINIONQA: Public opinion polls (76K individuals, 500 questions).
TWIN-2K: Psychological/economic/personality measures (2K individuals, 150 questions).
DUNNING-KRUGER: Grammar and logical reasoning tests (3K individuals, 20 questions).

Evaluation Settings:

Missing Responses (Imputation): Predicting answers for known individuals on known questions where the response was masked.
New Individuals: Predicting answers for entirely new individuals (no prior response history) based on their features.
New Questions: Predicting answers for new questions (unseen during training) for known individuals.

Baselines:
GEMS was compared against strong LLM-based methods, including:

Zero-shot and Few-shot prompting.
Agentic Chain-of-Thought (CoT) prompting.
Supervised Fine-Tuning (SFT) and Few-shot Fine-Tuning.
Classical baselines (XGBoost, Matrix Factorization).

4. Key Results

Across all datasets and settings, GEMS demonstrated competitive or superior performance to LLMs:

Accuracy: GEMS matched or outperformed the strongest LLM baselines (specifically Few-shot Fine-Tuning).
- In Settings 1 & 2, GEMS achieved high accuracy without using any textual features, relying entirely on the relational structure of the graph.
- In Setting 3 (New Questions), GEMS with the LLM-to-GNN projection achieved performance comparable to fine-tuned LLMs.
Efficiency:
- Parameters: GEMS uses approximately $10^3$ (1,000) times fewer parameters than LLMs (e.g., ~420K trainable parameters vs. 7B+ for LLMs).
- Compute: GEMS requires up to $10^2$ (100) times less compute (GPU hours). For example, fine-tuning a 70B model on OPINIONQA would take ~500 GPU hours, whereas GEMS trains in less than an hour.
Ensembling: Due to its low cost, the authors could ensemble 11 differently initialized GEMS models, further boosting accuracy, a strategy prohibitively expensive with LLMs.
Comparison to Classical Models: GEMS outperformed XGBoost and Matrix Factorization, highlighting the value of learning complex relational structures via message passing.

5. Significance and Contributions

Paradigm Shift: The paper challenges the assumption that LLMs are necessary for human simulation. It demonstrates that for close-ended tasks, relational structure is often a more powerful signal than language generation capabilities.
Transparency and Trust: Unlike LLMs trained on opaque, potentially contaminated datasets, GEMS is trained from scratch on task-specific data. This mitigates data leakage concerns and reduces the risk of inheriting pretraining biases. The model's mechanics (dot-product similarity) are more interpretable than black-box LLM generation.
Scalability: The extreme efficiency of GEMS lowers the barrier to entry for researchers with limited compute resources and enables scaling to massive datasets that would be infeasible for LLM fine-tuning.
Insights into Human Behavior: Analysis of GEMS embeddings revealed that the model naturally learned latent dimensions corresponding to political ideology and class, while also capturing significant heterogeneity within demographic subgroups—addressing a known failure mode of LLMs that often stereotype groups.

Conclusion

The authors conclude that GEMS offers a compelling, efficient, and transparent alternative to LLMs for close-ended human simulation. By leveraging graph neural networks to model the relational dependencies between individuals, subgroups, and choices, GEMS achieves state-of-the-art performance with a fraction of the computational cost, suggesting that graph-based modeling should be a primary tool in the human simulation toolkit.