ResearchBenchmarksPuzzle Solving on Sokoban (test)Follow3.74Average Rewardturn-PPO1.85762.34632.8353.3237Dec 18, 2025Evaluation ResultsMethodMethodLinksAverage Rewardturn-PPOBackbone=Qwen2.5-7BBackbone=Qwen2.5-7B2025.123.74Token-PPOBackbone=Qwen2.5-7BBackbone=Qwen2.5-7B2025.122.9turn-PPOBackbone=Qwen2.5-3BBackbone=Qwen2.5-3B2025.122.29Token-PPOBackbone=Qwen2.5-3BBackbone=Qwen2.5-3B2025.121.93