A few hours of BFS coding cracks MazeBench; scoring discrepancy flagged
FakePsyho · x · 2026-09-02
Independent developer FakePsyho shares his approach to MazeBench: with source access, he wrote an offline solver — a room solver that's basically an optimized BFS with no major tricks, plus simple glue logic chaining room solutions into a world solver that stays on the frontier of the puzzle without re-exploring solved rooms. A few hours (mostly code optimization) yielded a solver that grabbed 31 gems.
He was surprised to find no levels designed adversarially against solvers from a state-search perspective — which could have made AI-with-harness runs much harder. He also flags a scoring discrepancy: the human leaderboard max is 233 rooms and 80 gems, while official scores assume 100 total gems.
Related event: Developer Cracks MazeBench with a Simple BFS Solver(3 posts)→
More from Research
- Startup Mostik bridges AI models via their weights, tops ARC-AGI 3 at 1/20 the cost — nordicinst · 2026-09-03
- ByteDance's looped language models match 12B rivals at 1.4B size, with Bengio as co-author — peterjliu · 2026-09-03
- OpenAI's CoT monitorability hit: paper authors double down on 'fragile' AI safety window — GaryMarcus · 2026-09-03
- PufferLib author says retuning brings ~3x end-to-end speedup, up to 10x in some envs — yacineMTB · 2026-09-03
- Harvard Proposes Agentic Data Cracking, Cuts Unstructured QA Cost 53% on FanOutQA — Harvard · 2026-09-03
- Student runs explainable bone-lesion X-ray screener for £5/month — xrY- · 2026-09-03