Researchers simulate LLM-agent societies on a laptop
A single researcher ran 1M-agent simulations for 'a few dollars' using a new statistical method.
Independent researcher Igor Itkin introduced *Poor Man’s Agentic Modeling*, a technique that replaces expensive LLM agents with lightweight statistical surrogates trained on just a few hundred to thousand queries to models like DeepSeek. The method leverages an [interaction order × memory] taxonomy to predict surrogate error trends without running the full simulation, enabling accurate macro-scale analysis (e.g., phase behavior, scaling laws) at a fraction of the cost.
Itkin validated the approach on eight LLM simulations, including a reimplementation of the EconAgent macroeconomy, showing that the predicted error trends matched observed outcomes cell-by-cell. Crucially, the method refuted two prior predictions about saturating responses, matching their curvature quantitatively with zero free parameters. Code, data, and pre-registered systematic reviews are available via arXiv and Zenodo, with demos hosted on platforms like Hugging Face Spaces and Replicate.
- Replaces each LLM agent with a low-parameter surrogate trained on ~100–1,000 queries to models like DeepSeek, cutting compute costs by 100–1000x
- Validated across 8 simulations (including EconAgent), predicting macro-level behaviors like phase transitions and scaling laws with zero free parameters
- Full toolkit (code, data, demos) available on arXiv, Zenodo, and platforms like Hugging Face Spaces
Why It Matters
Unlocks large-scale LLM-agent research for teams without cloud-scale budgets, democratizing AI simulation while preserving scientific rigor.