Skip to content

Back to the catalog

GameHorizon Suite: Multi-Horizon Data and Evaluation in Gameplay
arxiv.orgpaper

GameHorizon Suite: Multi-Horizon Data and Evaluation in Gameplay

GameHorizon Suite is a data and evaluation package for measuring AI gameplay capabilities across multiple temporal horizons, combining visual understanding, instruction following, goal planning, and action control. It includes GameHorizon-Annotator (an automated multi-horizon instruction annotation pipeline), GameHorizon-Data (5,000 hours of AAA gameplay video, player actions, and instructions from 21 games and 100 expert players), and GameHorizon-Bench (offline and stepwise online evaluation tracks). The authors evaluate 47 models across more than one million invocations, exposing a task-difficulty hierarchy and sizable capability gaps between models.

TL;DR A 5,000-hour, 21-game AAA gameplay dataset plus offline/online benchmark suite for evaluating multi-horizon AI gameplay capabilities across 47 models.
Document

What it claims

GameHorizon is positioned against three named gaps in prior gameplay datasets/benchmarks: they cover a narrow range of games, lack language instructions, or rely on high-variance online rollouts. The suite's fix is to pair a large instruction-annotated dataset with an evaluation design that has both a reproducible offline track and an online track used specifically to check that offline scores track real gameplay ability.

How it works

The offline track is the reproducibility layer: thousands of standardized questions organized into three primary tasks plus a series of diagnostic variants — this is finer structure than "offline track" alone conveys. The online track is not just a second score; its stated job is to test whether offline scores reflect actual gameplay capability and to localize failures to specific steps within long-horizon gameplay. GameHorizon-Data is described as the first large-scale AAA dataset with temporally aligned video, player actions, and multi-horizon instructions, built via the GameHorizon-Annotator pipeline (5,000 hours, 21 games, 100 expert players).

Evidence

The abstract reports only the scale of the study — 47 models, >1M invocations — yielding a "meaningful hierarchy of task difficulty" and "pronounced differences in model capabilities." No per-model numbers, no baselines, and no leaderboard appear on this page.

How to use it

The authors state they "will release our dataset, annotator, and benchmark." Treat this as a future release: no weights, no license, no download link, no run commands are given here. If you want the annotation pipeline or the offline question set, you would need to wait for or locate the actual artifact release.

Limits

This is an abstract-only page. The concrete task definitions, the metric per task, the actual difficulty ranking, which of the 47 models lead, and any license/usage terms are all absent — you cannot judge whether the benchmark is hard for a specific model class without the paper body or the release. The claimed novelty ("first large-scale AAA gameplay dataset") and the offline-online agreement claim are asserted, not shown, in the text available here.

Organisations named here

  • GameHorizon-Bench
  • GameHorizon-Data
  • GameHorizon Suite

Technologies named here

  • GameHorizon-Annotator