DiG-bench (Discovery in Games) is a new benchmark consisting of 70 games designed to evaluate how well AI systems can discover unwritten rules of an environment through exploration and interaction. Each game is a self-contained miniature world with its own laws; both the rules and the objective are hidden and must be uncovered by interacting with the world, a setup similar in spirit to the visual ARC family of tasks.
Several of the games are playable on the project website (digbench.ai) to give a hands-on sense of the tasks.
What does the benchmark measure?
The benchmark focuses on an agent’s ability to identify mechanics that determine success. By experimenting in the game world, players learn how their actions change the environment and thereby reveal the mechanics needed to win. The idea is that success on these games indicates an ability to spot salient information in novel environments and update priors accordingly.
Who built DiG-bench?
Authors come from Thinking About Thinking, University of Oxford, Princeton University, King Abdullah University of Science and Technology, Swiss AI Lab, Inria, and MIT. One of the listed authors is Juergen Schmidhuber, noted as an early, creative researcher in AI.
Design choices and notable properties
- Text-native: the games are naturally suited to language models and are generally short enough that most traces fit within the context windows of current frontier models.
- Handcrafted and mostly private: human experts authored the games, and the majority are kept private to avoid leaking training data to models.
- Beatable but challenging: every game has been solved by at least one human player, but many were reported as difficult.
- Diverse skills required: solving the full suite requires a variety of strategies and competencies.
- Optional experimentation mode: an experimental setting reduces strict step limits for people who want to explore freely.
- Reassuringly hard: the games remain challenging enough that current frontier models do not uniformly solve them.
How do AI systems perform?
The benchmark is split into seven tiers (Tier 1 easiest, Tier 7 hardest). Of the 70 games, 21 have been released publicly while the others remain withheld.
Most games have multiple levels. Available actions per step range from 2 to 34.
Reported performance highlights:
- Opus 5 and Fable 5 (when used with Claude Code) are the best-performing models overall.
- Only Opus 5 and Fable 5 managed to beat any tasks in Tier 7 (the source notes a 0.2 result for that level).
- Opus 5, GPT-5.5, and Kimi K3 were able to beat some Tier 6 tasks, particularly when given access to a harness (e.g., Claude Code).
- GLM-5.2 and Gemini 3.1 Pro beat some levels in Tier 4.
Overall, the benchmark remains hard for current models.
Why this matters
Benchmarks like DiG-bench aim to isolate a prerequisite for creativity: the ability to autonomously discover useful, undocumented structure in novel situations. The results show that some frontier models can perform notable discoveries, yet they still lag behind humans in many respects—the source, for example, gives a 20% success rate on Tier 7 as a point of comparison to human solutions in some cases.
The source’s author speculates that human parity on DiG-bench could be reached by mid-2027, and suggests that at that point mechanisms like recursive self-improvement may become more active. That projection is an opinion offered by the author of the original piece, not an empirical claim from the benchmark itself.
Where to read more and try the games
The DiG-bench repository and PDF documentation are available through the project’s materials. Playable games and the leaderboard are hosted at the official site: digbench.ai.
Summary
DiG-bench is a structured, hand-crafted testbed for measuring discovery and creative intuition in text-based interactive environments. While some leading models already show capability on easier tiers, the benchmark continues to present substantial challenges, helping to map how well systems can autonomously uncover hidden mechanics.
