Skip to content

Fields

Game agents latest news

AI that acts in games: generalist game agents, LLM-driven bots, agents in Minecraft, StarCraft and other environments, benchmarks and open frameworks.

111 picks55 in the last 30 days170 collected in all

Updated

Picks archive · Page 3

Oct 28Wed41–60
  1. arXiv Game AI45

    Gym-Anything-RT: Scaling real-time environments for vision-language agents

    Gym-Anything-RT generates 100 real-time visual mini-games automatically through an agent pipeline to evaluate the real-time interaction abilities of vision-language agents. The paper reports that GPT-6 Astra (xHigh) achieves a pass rate of only 9.0% on the real-time mini-games, rising to 75.0% when the environment is paused. The authors also find that some frontier agents bypass real-time perception and control when allowed to write code. The environment and the generated mini-games are publicly released.

    Why it matters: This study separates real-time performance from other abilities. It allows agents' scores to be compared before and after pausing on the same mini-game, which helps identify weaknesses in real-time decision-making.

  2. arXiv Game AI39

    UnityBench-Mini3D: Evaluating LLM agents for Unity 3D game development

    Researchers propose UnityBench-Mini3D, which uses 77 project-level Unity tasks to evaluate the 3D game development capabilities of LLM coding agents. Each task comes with 1,178 hand-written milestone tests that award partial credit and pinpoint where failures occur. Across tests of 21 models and 9 coding frameworks, the average pass rate was 80.9% on easy tasks, dropping to 35.2% on hard tasks, with an overall rate of 57.7%.

    Why it matters: Existing game development benchmarks mostly target Godot and rely on LLM judges. This work instead evaluates Unity tasks with deterministic in-engine milestone tests, offering a methodological counterpoint.

  3. arXiv Game World Models & Generation39

    BoxWorld: Let World Engines Control and Video Generators Render

    BoxWorld hands user-defined interaction rules to the Box World Engine, then has the Box-State generative renderer autoregressively render a real-time interactive video world from the object states. The paper says it clearly outperforms end-to-end world models on evaluations including PairedStateBench and remains competitive on several public benchmarks. Training has three stages: learning object localization and state transitions from automatically annotated open-domain and game videos, then rollout distillation on self-generated histories.

    Why it matters: Separating interaction reasoning from generative rendering, and comparing it with end-to-end video world models, offers one way to divide the work in controllable game worlds.

  4. arXiv Board & Classic Game AI35

    arXiv paper tests the stability of opening rankings with 17,400 Stockfish self-play games

    An arXiv study had strength-limited Stockfish 15.1 play 17,400 self-play games under three strength settings, comparing ten white first moves and ten black replies to 1.e4, and found that the leading openings change with the strength setting, with two of them separated by only a single game.

    Why it matters: Drawing on 17,400 self-play games, the paper quantifies the uncertainty of opening leaderboard leaders when the number of games is small, which can inform evaluation design.

  5. arXiv Game AI59

    Gym-Anything-RT: Scaling real-time environments for vision-language agents

    Gym-Anything-RT generates 100 real-time visual mini-games automatically through an agent pipeline to evaluate the real-time interaction abilities of vision-language agents. The paper reports that GPT-6 Astra (xHigh) achieves a pass rate of only 9.0% on the real-time mini-games, rising to 75.0% when the environment is paused. The authors also find that some frontier agents bypass real-time perception and control when allowed to write code. The environment and the generated mini-games are publicly released.

    Why it matters: This study separates real-time performance from other abilities. It allows agents' scores to be compared before and after pausing on the same mini-game, which helps identify weaknesses in real-time decision-making.

  6. Naavik56

    Artificial Agency CEO on AI Teammates in Owlchemy Game and Single-Digit-Cent Inference Costs

    Artificial Agency built a single-player mode for Owlchemy Labs' VR co-op game Dimensional Double Shift, replacing human teammates with three AI teammates called Temps.

    Why it matters: The founder lays out a concrete path to pushing inference costs down to single-digit cents per player hour, offering a basis for judging whether agent gameplay can work commercially.

Oct 27Tue
  1. arXiv Game World Models & Generation37

    How humans learn unfamiliar worlds: a think-aloud dataset on ARC-AGI-3

    Researchers built a think-aloud dataset that records how humans explore and learn unfamiliar game mechanics. It contains 144 human trajectories from 64 participants across 25 public games, plus 250 LLM agent trajectories. The analysis shows that qualitative representation makes up 50.0% of successful human trajectories versus 14.3% of successful LLM trajectories, and humans more often explore without a specific hypothesis.

    Why it matters: This paper uses think-aloud human game trajectories to compare with the failure and recovery processes of LLM agents, offering a reference for examining the limitations of existing game-agent benchmarks.

  2. arXiv Game World Models & Generation53

    BoxWorld: Let World Engines Control and Video Generators Render

    BoxWorld hands user-defined interaction rules to the Box World Engine, then has the Box-State generative renderer autoregressively render a real-time interactive video world from the object states. The paper says it clearly outperforms end-to-end world models on evaluations including PairedStateBench and remains competitive on several public benchmarks. Training has three stages: learning object localization and state transitions from automatically annotated open-domain and game videos, then rollout distillation on self-generated histories.

    Why it matters: Separating interaction reasoning from generative rendering, and comparing it with end-to-end video world models, offers one way to divide the work in controllable game worlds.

Oct 26Mon
  1. arXiv Game AI48

    UnityBench-Mini3D: Evaluating LLM agents for Unity 3D game development

    Researchers propose UnityBench-Mini3D, which uses 77 project-level Unity tasks to evaluate the 3D game development capabilities of LLM coding agents. Each task comes with 1,178 hand-written milestone tests that award partial credit and pinpoint where failures occur. Across tests of 21 models and 9 coding frameworks, the average pass rate was 80.9% on easy tasks, dropping to 35.2% on hard tasks, with an overall rate of 57.7%.

    Why it matters: Existing game development benchmarks mostly target Godot and rely on LLM judges. This work instead evaluates Unity tasks with deterministic in-engine milestone tests, offering a methodological counterpoint.

  2. arXiv Game World Models & Generation49

    How humans learn unfamiliar worlds: a think-aloud dataset on ARC-AGI-3

    Researchers built a think-aloud dataset that records how humans explore and learn unfamiliar game mechanics. It contains 144 human trajectories from 64 participants across 25 public games, plus 250 LLM agent trajectories. The analysis shows that qualitative representation makes up 50.0% of successful human trajectories versus 14.3% of successful LLM trajectories, and humans more often explore without a specific hypothesis.

    Why it matters: This paper uses think-aloud human game trajectories to compare with the failure and recovery processes of LLM agents, offering a reference for examining the limitations of existing game-agent benchmarks.

Oct 24Sat
  1. 80 Level44

    Playgama Sandbox creates a faster route from prototype to players

    Playgama's founder gave an interview explaining how its Sandbox and MCP tools link publishing browser games, integrating the Bridge SDK, and collecting data into a faster workflow. Through MCP, agents can read retention, playtime, and custom event data and suggest improvements, while decisions such as rolling back versions, requesting withdrawals, and choosing new distribution platforms are still made by developers in the console.

    Why it matters: The article walks through the full process of agents reading event data via MCP and iterating on browser games, and which decisions remain with developers. It is useful for comparing against your own release workflow.

Oct 23Fri
  1. arXiv Game AI35

    HiSTEVE 1.0: human-interface VLA agents for open-world games

    HiSTEVE 1.0 is a hierarchical VLA agent for open-world games such as Minecraft, in which a single VLA executes all subtasks using pure visual input and native keyboard-and-mouse actions. The paper reports that it scores 85.0% on MineStudio Bench and achieves an average success rate of 37.4% on OpenHA, with state-of-the-art performance on atomic tasks. The work also reports that the agent can synthesize long-horizon tasks such as an iron pickaxe from scratch, like a human player.

    Why it matters: The paper combines high-level subtask planning and native keyboard-and-mouse visual control in one VLA agent, making it easier to compare the trade-off between long-horizon autonomy and the reliability of interface operation.

  2. arXiv Game World Models & Generation34

    Can LLMs serve as event world models? A study from single-actor environments to multi-actor event worlds

    A paper asks whether large language models can predict future observations from event logs and the actions an agent plans to take, using benchmarks built from human Diplomacy games and real Polymarket trades. When only the agent's own actions are given and other actors strongly shape outcomes, model predictions fall well short, and larger models and longer histories help only modestly. When the recorded actions of all actors are provided, fine-tuning and more data improve predictions, but models still overestimate the success rate of planned actions, while imagined experience from world models still improves policy learning.

    Why it matters: The paper argues that when given only the agent's own actions, models struggle to predict outcomes strongly shaped by other actors, a key gap for event world modeling with multiple actors

Oct 22Thu
  1. arXiv Game AI50

    HiSTEVE 1.0: human-interface VLA agents for open-world games

    HiSTEVE 1.0 is a hierarchical VLA agent for open-world games such as Minecraft, in which a single VLA executes all subtasks using pure visual input and native keyboard-and-mouse actions. The paper reports that it scores 85.0% on MineStudio Bench and achieves an average success rate of 37.4% on OpenHA, with state-of-the-art performance on atomic tasks. The work also reports that the agent can synthesize long-horizon tasks such as an iron pickaxe from scratch, like a human player.

    Why it matters: The paper combines high-level subtask planning and native keyboard-and-mouse visual control in one VLA agent, making it easier to compare the trade-off between long-horizon autonomy and the reliability of interface operation.

  2. arXiv Board & Classic Game AI35

    D-Skill: Distilling teacher knowledge through agent skills

    D-Skill distills the paired trajectory differences between a teacher and a student model into interpretable natural-language skill documents. The student uses these documents at inference time without updating its weights. The paper reports that the method closes 75%–95% of the performance gap between teacher and student on SpreadsheetBench-Verified, ALFWorld, and chess puzzles. Within the same model family, the student's per-task inference cost is 5.7x–8.3x lower than the teacher's.

    Why it matters: Unlike traditional distillation, which rewrites the student's weights, it records the teacher's experience as readable skill documents, making the approach easier to examine and compare.

  3. arXiv Board & Classic Game AI35

    ToM2Vec: Continuous theory-of-mind tokens for multi-agent coordination

    ToM2Vec learns theory of mind for multi-agent coordination through a fixed budget of continuous reasoning tokens, using partners' actual thoughts as training supervision. Across cooperative, competitive, and mixed-motive environments, including Overcooked-AI, multi-party negotiation, Decrypto, Hanabi, and Collab Escape/Capture, it delivers consistent gains over the base model and several competitive baselines.

    Why it matters: The article compresses theory of mind into a fixed number of continuous reasoning tokens, supervised by partners' actual thoughts, in contrast to text-based reasoning approaches.

  4. arXiv Game AI38

    Scaling in games: continued pre-training for embodied agents in diverse virtual worlds

    The paper systematically studies how data volume, diversity, and model scale affect embodied agents in continued pre-training, using 120 training configurations and models from 0.8B to 27B parameters. Diversity improves transfer to related worlds but gives limited gains for worlds that differ substantially. Larger models improve data efficiency, but when trained on a single world they generalize worse to distant worlds. The authors argue that building general-purpose embodied agents requires training stages richer than continued pre-training.

    Why it matters: The paper compares the effects of data volume, diversity, and model scale across 120 training configurations. Its finding that gains from cross-world transfer are limited makes it a useful reference for game agent training strategies.

  5. arXiv Game World Models & Generation40

    Can LLMs serve as event world models? A study from single-actor environments to multi-actor event worlds

    A paper asks whether large language models can predict future observations from event logs and the actions an agent plans to take, using benchmarks built from human Diplomacy games and real Polymarket trades. When only the agent's own actions are given and other actors strongly shape outcomes, model predictions fall well short, and larger models and longer histories help only modestly. When the recorded actions of all actors are provided, fine-tuning and more data improve predictions, but models still overestimate the success rate of planned actions, while imagined experience from world models still improves policy learning.

    Why it matters: The paper argues that when given only the agent's own actions, models struggle to predict outcomes strongly shaped by other actors, a key gap for event world modeling with multiple actors