Recursive Game Creator: an experience-oriented agentic harness hits SOTA 77.89 on GameCraft-Bench
hkuhk · hf · 2026-10-08
HKU researchers present Recursive Game Creator, an experience-oriented harness that advances agentic game development from rough prototypes to entertaining games, arguing that program correctness doesn't ensure player enjoyment.
The recursive loop involves four roles:
- Designer: translates user instructions and Reviewer feedback into detailed plans;
- Builder: implements plans into candidate games;
- Player: a coding-native player that creates and executes reusable policies via programmatic interfaces for efficient, diverse gameplay trajectory collection, avoiding GUI-based collection bias;
- Reviewer: evaluates games with trajectory-based metrics that induce player preferences, combining visual evidence and textual preferences, accepting the better version each round.
Results: SOTA 77.89 on GameCraft-Bench; 53.2% strict success rate on GameASG-Bench (+34.1% over same-model baseline); highest runtime-check pass rate at 93.4%; user study shows longer playtime and higher ratings. Code coming soon.
More from coding & agent
- DeLM's decentralized multi-agent system runs 2.49x faster, but MAS evals pick wildly different metrics — jyangballin · 2026-10-08
- Musk endorses Grok bot + Cursor cloud agents combo as a major productivity boost — elonmusk · 2026-10-08
- Dev: in the AI coding era, tabs-vs-spaces and style nitpicking never mattered — facontidavide · 2026-10-08
- RunningTab: environment-side task ledger consistently beats in-model tracking across 3 benchmarks and 3 LLMs — RexDouglass · 2026-10-08
- The top skill now: running overnight agents reliably without getting reward-hacked — zack_overflow · 2026-10-08
- Relic turns multi-agent collaboration into persistent organizational protocols, hitting 79.1% on CooperBench — youjiaxuan · 2026-10-08