NUS-Tencent Player-1 Lifts MineDojo-300 Success to 38.33%

On MineDojo-300, Player-1-9B completes 38.33% of 300 Minecraft tasks, up from 10.54% after supervised fine-tuning alone and well past the 12.63% held by GameWAM, the strongest external baseline the authors evaluated. The number of tasks the 9B model solves in at least four of eight trials climbs from 20 to 136.

The paper, posted on Hugging Face and credited to researchers at the National University of Singapore, Tencent IEG, and Nanyang Technological University, lays out a training recipe called Player-1. It starts from Qwen3.5-9B and Qwen3.5-35B-A3B backbones, fine-tunes them on 6,044 curated gameplay trajectories covering 295,097 interaction turns, and then runs 200 PPO steps across sixteen synthetic 2D and 3D environments the group calls Player-1-Gym. The environments cover five control families: aiming, combat, driving, navigation, and UI interaction.

"We demonstrate that learning from synthetic experience substantially improves game control beyond supervised initialization, with gains that generalize beyond the RL training suite," the abstract states. The 35B-A3B variant finishes at 41.04% on MineDojo-300. On ViZDoom-10, outside the RL training suite, Player-1-9B lifts its scaled return from 0.4361 after SFT to 0.6922. On GameWorld's 170 browser-game tasks across 34 games, progress rises from 31.43% to 34.45%.

The policy also runs on a commercial game the system has never seen. The authors drive Player-1 inside Watch Dogs 2, which they call "a game unseen during both SFT and RL", and show on-foot navigation, car approach, vehicle entry, and city driving from screenshot observations and keyboard-mouse controls.

Even at 38.33%, the 9B agent still fails roughly two of every three MineDojo tasks. Release terms for Player-1-Gym, the 6,044-trajectory corpus, and the model weights are not stated in the paper, and training compute and per-turn inference latency go unreported. It arrives on a day when several agent-training papers were already moving through our agents feed.