← Key papers

Paper review · November 17, 2026

Gamer: A General-Purpose VLA Foundation Model for Long-Horizon, Real-Time Game Agents

Gamer: A Generalist Vision-Language-Action Foundation Model for Long-Horizon Real-Time Game Agent

AgentsMultimodalReinforcement learning Advanced Upvotes 87 cs.CL

The authors pretrain on tens of thousands of hours of keyboard and mouse trajectories, then apply SFT and multi-turn reinforcement learning, producing a general-purpose agent that plays games in real time from the screen alone.

Problem

Existing game agents depend on internal game state, dedicated APIs and custom action spaces, which makes it hard for them to generalize to other games. Modern 3D open-world games involve hours-long tasks, partial observability and delayed rewards. When inference latency is long, the observed screen drifts from the actual game state, causing control errors and degrading the quality of online reinforcement learning.

Method

Built on Qwen3.5/Qwen3-VL, the model takes screen pixels as input and outputs keyboard and mouse actions. Observations run at 5Hz, and actions run at 30Hz, split into six blocks. (1) CPT stage: tens of thousands of hours of multi-game and embodied-environment trajectories are converted into a unified action representation for training. Key mappings are randomized to reduce bias toward key identity, and the loss weight of repeated actions is decayed. (2) SFT stage: sparse-thinking instruction tuning on more than 1,000 hours of multi-step instruction data built with LLM-assisted annotation. The model reasons only when needed and carries memory summaries across context boundaries. (3) RL stage: TA-GRPO addresses finite context, accumulated rollout errors and sparse rewards. For inference the authors use their in-house Streaming-vLLM, which reduces latency through KV cache reuse, FSM-based speculative decoding and streaming action transmission.

Results

Limitations

The authors themselves note that the benefit of scaling CPT data does not carry over to distant unseen games. Within the provided text, specific 3D game performance numbers and baseline comparisons could not be confirmed. Evaluation appears to center mainly on the company's own games (Honkai: Star Rail and others), so the scope of generalization needs further checking. Because instruction-data annotation relies on an LLM, annotation quality may also affect results.

Why it matters

Because it carries out long tasks using only the same screen, keyboard and mouse interface as a human, without game-specific APIs, it shows a training and serving recipe that bridges computer-use agents and embodied agents. The release of the inference engine and some checkpoints also makes it useful in practice.

Key terms

CPT(Continued Pre-Training)
A stage in which an already pretrained model continues pretraining on data from a new domain. Here it learns general control skills from keyboard and mouse data.
Sparse-thinking
The model does not reason at every step. It reasons explicitly only at important decision points, such as a failed path or a change of goal, to cut wasted computation.
TA-GRPO
A multi-turn reinforcement learning method that modifies GRPO to account for temporal information. It addresses accumulated execution errors and sparse rewards that arise in long interactions.
Speculative Decoding
An inference acceleration technique that quickly generates candidate output tokens first and then verifies them all at once to speed up generation.

Original

Authors: Run Luo, Kuan Zhang, Qi Deng, Yukun Chen, Zhihao Yang, Yufei Zhang, Renrong Hu, Yi Zhang, et al. · Posted 2026-11-13 · License CC BY-NC-SA 4.0 · Based on the full text

This review is an AI-written summary of the paper. Please treat the original paper as the reference.