Recursive Game Creator reports 34.1% jump on GameASG-Bench

A four-agent harness (Designer, Builder, Player, Reviewer) recursively builds and play-tests prototype games, and reports a 34.1% gain in strict task success on GameASG-Bench over its baseline plus a 93.4% runtime-check pass rate the authors call the highest among tested methods. The pitch, laid out in an arxiv paper from Jiajun Chen, Haoyu Wu, Mingda Jia and Xihui Liu, is that agentic coding research has fixated on whether generated games compile and run, when the thing that matters for a game is whether anyone wants to play it.

So the harness swaps in a Player that is "a code-based agent that generates and executes reusable gameplay strategies through programmatic interfaces for collecting varied gameplay data," feeding the resulting trajectories to a Reviewer that mixes trajectory-based metrics, visual analysis and player preference data into its score. "Rather than merely ensuring code functionality, it prioritizes player enjoyment," the paper writes. On GameCraft-Bench the system records an overall score of 77.89, and a user study reports longer playtime and higher satisfaction ratings, though the abstract gives no sample size.

The code-based Player is a deliberate shortcut. "The coding-native Player component efficiently mitigates evaluation bias by avoiding slower GUI-based data collection methods," per the paper. Faster than a human, cheaper than visual-interface agents, and arguably further from how a real player plays. The abstract does not name what the baseline being beaten here is, nor which underlying language model drives any of the four roles.