Original title: GPS-Bench: A Governance Policy Benchmark for Automating Policy Analysis
Linh Le, Melanie Bui, My Chiffon Nguyen, Zachary Schlosser +1 · September 2026
Policy analysis means predicting who a proposed rule affects, how those groups react, and what happens next. Computer programs using large language models can simulate these steps, but checking whether their simulations match reality is tough. To fix this, researchers built GPS-Bench. It is a benchmark that tests policy simulations against real-world evidence, such as laws, lobbying reports, company filings, and economic data. Instead of making up generic personas, the system builds each actor directly from dated public documents. All tested methods read the same evidence-based setup and produce matching outputs. The researchers compared single-model reasoning, communicating agent groups, graph methods, and fine-tuning. They found that fine-tuning on real records predicts actor-level impacts best. Breaking the problem down into multiple interacting agents does not beat fine-tuning at predicting outcomes. However, multi-agent setups show the real mechanism: they reveal how groups form coalitions and bargain based on private evidence.
When planning a picnic, friends might argue based on what they think is in the fridge, not what's actually there. Without checking, they could double-buy items or forget essentials. Here, the researchers provide a central record of actual past picnics—what was brought and eaten. In this paper, GPS-Bench acts like that record, so each simulated person negotiates with real evidence, and their deals can be checked against history.
Policy analysts can now test whether their simulations reflect real actor behaviors and outcomes. This makes automated policy analysis more trustworthy and useful for decision-making.
The abstract does not specify universal limitations, but GPS-Bench relies on public evidence sources like legislative records and corporate filings, which may not cover all policy contexts. It is designed for controlled comparisons of specific simulation methods, not for comprehensive policy forecasting.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
GPS-Bench: A Governance Policy Benchmark for Automating Policy Analysis — Linh Le, Melanie Bui, My Chiffon Nguyen et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.