CivBench: A Long-Horizon Benchmark for Tool-Mediated Agents in Civilization VI
The collaborative research paper behind CivBench, studying how agents monitor a game and follow through on plans.
Article · 2026Sources & confidence: Qualified confidence
Checked 4 October 2026 · AI-assisted review
The collaborative research paper behind CivBench, studying how agents monitor a game and follow through on plans.
- CivBench: A Long-Horizon Benchmark for Tool-Mediated Agents in Civilization VI
Publication identity and subject.
- CivBench — research update
Author’s announcement of NeurIPS 2026 poster acceptance.
What this check doesn’t establish
- Acceptance is recorded from the author’s announcement; this citation describes the arXiv preprint, not a proceedings version.
Confidence applies to the scope above, not every claim made by the underlying source. How confidence works
The paper introduces a tool-mediated Civilization VI benchmark and studies agents’ monitoring and follow-through. Its 23 admissible runs are a pilot, not a reliable ranking of models.
Written with Austin Tudor David Andrews, Jamie Heagerty, Harry Coppock, Jakob Nicolaus Foerster and Rui Ponte Costa. The work was accepted as a poster at the NeurIPS 2026 Evaluations and Datasets Track. For me, it is a reminder of what becomes possible when you find people who share your curiosity and build together.
The project is the research apparatus; the essay tells the story of the early experiments.
In the essays
- I Gave an AI a Civilization to Run. It Built a Nuke.
Read the essay
People
- Liam Wilkinson
Author
Projects
Ideas
- The knowing-doing gap
Subject
- The sensorium effect
Subject
Record history
Catalogue source: knowledge/works/civbench-paper.md
Generated by codex/gpt-6 on 2026-10-04