Repo

GameHorizon

Evaluates gameplay models from short operations to long-term strategy using aligned gameplay data plus offline MCQs and stepwise online tasks.

GitHub ↗Sep 20, 2026Paper ↗

  • multimodal
  • gameplay-benchmark
  • agent-evaluation
  • video-understanding
Full pipeline: annotator, data structure, and benchmark tracks showing the method
Full pipeline: annotator, data structure, and benchmark tracks showing the method

What it is

Unified gameplay data and benchmark suite for measuring model capability from short actions to long-horizon strategies across AAA games. Measures gameplay competence across temporal horizons with a unified dataset, annotation pipeline, and offline/online benchmarks spanning AAA titles, actions, videos, and multi-level instructions.

Key results

  • 5,000 gameplay hours across 21 AAA titles with aligned video and actions
  • 5,000 offline MCQs covering 3 primary and 10 variant tasks
  • 20 online tasks and 62 subtasks; 47 models benchmarked

Shows that a single benchmark can probe gameplay competence consistently from short-horizon perception to long-horizon planning.

One useful ML email a week.

Worth your time, not all of your time.

We track new ML work across papers, code, and data, then keep only the pieces worth a closer look. Selected, not collected.

Papers.Data.Code

The Weekly ML Signal

Selected, not collected.

One useful ML email a week: the papers, repos and datasets worth your attention.

Worth your time, not all of your time.