List your company
ResearchFormerly WizWand

Reinforcement Learning on Maze Gymnasium

0.97Mean Best RewardA2C
0.276840.4567950.636750.816705May 8, 2026May 15, 2026May 22, 2026May 29, 2026Jun 5, 2026Jun 12, 2026Jun 20, 2026
Updated 4mo ago

Evaluation results

MethodLinks
A2CEpisodes per run=8,000, Independent runs=102026.050.97
TRPOEpisodes per run=8,000, Independent runs=102026.050.97
ProPS+Episodes per run=8,000, Language Model backbone=GPT-4o, Independent runs=102026.050.97
R2POEpisodes per run=4,000, Language Model backbone=gpt-oss:20b, Independent runs=102026.050.97
ProPSEpisodes per run=8,000, Language Model backbone=GPT-4o, Independent runs=102026.050.55
ARTS*Base Model=Qwen 4B, Test-time trained=true2026.060.53
Human BestMethod Type=Human Baseline2026.060.525
ARTSBase Model=o32026.060.512
LinearMethod Type=Prior Works2026.060.468
MLEvolveMethod Type=Prior Works2026.060.4535
AIRAMethod Type=Prior Works2026.060.3642
ARTSBase Model=Qwen 4B2026.060.3035