Skip to content

Formats

Research latest news

Game AI papers and technical reports worth reading: a selection on agents, environments, generation and player research.

160 picks75 in the last 30 days242 collected in all

Updated

Picks archive · Page 3

Oct 31Sat41–60
  1. arXiv Board & Classic Game AI34

    arXiv paper conditions a MIDI model on similarity labels to control section similarity in Super Mario game music

    The paper further trains a pretrained MIDI model on section pairs from Super Mario music, using similarity labels and context vectors to control how close a newly generated section is to a given section. In an evaluation spanning 157 test contexts, 6 generated outputs per label and 12 metrics, outputs asked to be similar were on average closer to their context than outputs asked to be dissimilar, and were no closer to other pieces; all 12 metrics support this result, and every main metric remains significant after multiple-comparison correction.

    Why it matters: The paper offers a similarity-label conditioning method for MIDI music generation together with results on twelve metrics, showing a concrete route to controllable generation for game music.

  2. arXiv Board & Classic Game AI39

    How a MIDI model further trained on Super Mario tracks uses similarity labels to control how closely game music matches the theme

    The paper continues training a pretrained MIDI model on pairs of segments from Super Mario tracks, using similarity labels and context vectors to control how closely a generated segment approaches a given theme. Across 157 test contexts, each label yields 6 output segments, and all 12 metrics show that outputs asked to be similar are closer to their context than to other tracks, with the four main metrics still significant after multiple-testing correction. Compared with a matched model without the labels, only high-similarity requests make outputs more similar; agreement between listening-test scores and melodic overlap is the highest among the metrics.

    Why it matters: The paper compares with and without similarity labels across 157 test contexts and 12 metrics, offering an evaluation reference for controllable music generation.

Oct 30Fri
  1. arXiv Game World Models & Generation58

    Wizard: Manipulate any world in real time via agentic control

    The paper proposes Wizard, an agentic game world model that can precisely control character actions in real time and achieve zero-shot generalization to new scenes, characters, and actions. The authors built a dataset covering 750 scenes and 178 characters, using skeleton and geometry information as structured representations to condition a video generation model, and introduced an adaptive masking mechanism to improve the consistency between the conditioning signal and the generated video. A companion agent framework explicitly maintains and updates game state, letting players freely configure virtual worlds and control character actions at fine granularity.

    Why it matters: The paper presents a dataset of 750 scenes and 178 characters and a structured agentic conditioning scheme, offering a look at an implementation path for real-time controllable world models.

  2. arXiv Game World Models & Generation58

    Algebraic World Model: Turning game environment interactions into Lean 4-provable mathematical models

    Researchers propose the Algebraic World Model (AWM), which turns interactions with unfamiliar environments into executable Python hypotheses, then organizes them into typed objects and shared rules, simplifies the structure with category-theory-guided transformations, and finally formalizes it in Lean 4 to prove reachability and impossibility.

    Why it matters: The paper presents a pipeline that turns interactions with unfamiliar game environments into Lean 4-checkable mathematical models, letting readers understand how formal proofs can be used to determine level unreachability.

  3. arXiv Game AI58

    Fight World Model: Customizable combat through visual interaction proxies

    The authors propose Fight World Model (FWM), which separates rule-constrained combat execution from video rendering: an executable engine maintains state and resolves interactions, and a deterministic compiler then encodes the results into spatiotemporally aligned interaction proxies.

    Why it matters: The paper separates rule execution from video rendering, letting readers understand how a programmable combat world model keeps hit and health states consistent.

  4. arXiv Game AI53

    Fight World Model: customizable combat through visual interaction proxies

    The researchers propose Fight World Model (FWM), separating rule-driven combat execution from video rendering, with an executable engine maintaining state and resolving interactions. A deterministic compiler encodes resolution results and attempted actions into spatiotemporally aligned interaction proxies, aligning actions with skill windows, brief hit reactions, and continuous health by character identity and time; the model is trained on the FightWorld Corpus, which covers 52 skill effects, 187 character appearance variants, and 15 arenas.

    Why it matters: The paper separates combat rule execution from video rendering and passes state through interaction proxies, showing an approach that applies new rules such as damage scaling and lifesteal without modifying the renderer.

  5. Tencent Hunyuan58

    Tencent Hunyuan releases Hy-Nomas technical report and evaluation set

    Tencent Hunyuan has released a technical report for Hy-Nomas, an open-source interactive video world model, and open-sourced its evaluation set, Hy-Nomas-Bench. The authors also used Hy-Nomas to reproduce examples from Genie 4 and attached comparison videos. The cited release notes say the model supports free interaction between text prompts and camera trajectories, and offers long-horizon generation with world memory.

    Why it matters: The authors open-sourced the evaluation set and reproduced Genie 4's examples, so readers can compare the public videos and judge the interactivity and long-horizon generation quality for themselves.

  6. arXiv Game World Models & Generation60

    Algebraic world model AWM proposes a framework from environment exploration to Lean 4 formal proofs

    The research team proposes the Algebraic World Model (AWM), which turns interaction with an unknown environment into a mathematical model that can be formally reasoned about in Lean 4. On two unsolvable game levels, six AWM runs all produced kernel-verified, closed Lean proofs ruling out every valid winning sequence, whereas Claude Opus 5.5 and GPT-6 Astra, equipped with programming tools, failed to produce a proof in twelve baseline runs.

    Why it matters: Reaching a kernel-verified unreachability conclusion on levels that cannot be completed shows how formal reasoning can be applied to modeling game environments.

  7. arXiv Game World Models & Generation58

    AWM: an algebraic world model that turns unfamiliar game environments into Lean-provable mathematical models

    Researchers propose the Algebraic World Model (AWM), which turns interactions with unfamiliar environments into explicit mathematical models that support formal reasoning, and proves reachability and unreachability in Lean 4.

    Why it matters: The paper presents a pipeline for turning unfamiliar game environments into checkable mathematical models and uses Lean to prove levels unreachable, helping readers understand how formal reasoning can enter game agents.

  8. arXiv Game World Models & Generation58

    Wizard: Manipulate any world in real time via agentic control

    The paper proposes Wizard, an agentic game world model that can precisely control character actions in real time and achieves zero-shot generalization to new scenes, characters, and actions. The researchers built a dataset covering 805 scenes and 178 characters, using skeleton and geometry information as structured proxies to condition video generation, and supporting character-aware adaptive masking and world-anchored spatial conditions through explicit tracking and updating of game state.

    Why it matters: The paper provides a dataset of 805 scenes and 178 characters along with the WBench evaluation, offering a concrete reference for comparing existing world models on real-time controllable generation.

Oct 29Thu
  1. arXiv Game AI40

    TaskFrame proposes hierarchical planner and controller framework connecting task-level planning to frame-level keyboard-and-mouse control

    Guangzhi Wang proposes TaskFrame, a hierarchical planner and controller framework that connects task-level planning with frame-level keyboard and mouse control. A vision-language task planner breaks high-level goals into ordered atomic instructions, optimized with supervised learning and dependency-aware GRPO; the hidden state of a special goal token that terminates each atomic instruction is cached as a continuous goal embedding and reused in subsequent control steps.

    Why it matters: The paper presents a three-stage training pipeline for the planner and controller along with a cached goal embedding approach, so readers can judge how the hierarchical structure connects vision-language reasoning with keyboard-and-mouse control.

  2. arXiv Board & Classic Game AI36

    RetroChessBench: A benchmark for reconstructing and proving chess game histories

    The authors propose RetroChessBench, a retrospective chess benchmark with 12 tasks covering fixed-length construction, unique final positions and exhaustive answer sets, together with reference answers, grading rubrics and a terminal-reward environment.

    Why it matters: The paper presents 12 retrospective chess tasks with grading rubrics and documents the specific failure types of frontier models on pure rule-based reasoning, serving as a reference for evaluating reasoning ability.

Oct 28Wed
  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 World Models & Generation37

    GameVerse: a minute-scale gameplay dataset for long-horizon interactive world modeling

    GameVerse is a minute-scale gameplay video dataset for long-horizon interactive world modeling, aimed at easing the problem of overly short clips in existing data. It contains 69.7K clips, 1655 hours and 349.6 million frames, of which 1358 hours are continuous two-minute sequences. Each clip comes with a camera trajectory that stays coherent across minutes, and captions aligned to event boundaries that distinguish static scenes from dynamic foreground events. Experiments show that training on it extends the temporal span of interactive world models.

    Why it matters: Addresses the data bottleneck of short game video clips and rough subtitles by offering a minute-scale dataset and an event-level annotation approach, which can be compared with how long-horizon world models build their data.

  3. 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.

  4. 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.

  5. 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.

  6. 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.

Oct 27Tue
  1. arXiv Game World Models & Generation37

    GameVerse releases minute-scale gameplay video dataset for long-horizon interactive world modeling

    Researchers released GameVerse, a minute-scale gameplay video dataset for long-horizon interactive world modeling, containing 69.7K videos, 1,655 hours and 349.6 million frames, covering first-person and third-person gameplay. Of this, 1,358 hours are continuous two-minute sequences, an order of magnitude longer than previous datasets; each video comes with a camera trajectory that stays consistent across minutes, plus annotations aligned to event boundaries that distinguish static scenes from dynamic foreground events. The paper states that training on this dataset can significantly extend the time horizon of interactive world models.

    Why it matters: The dataset extends usable sequences to the two-minute range and compares it with existing public datasets on scale and annotation, making it easier to judge the training-data gap for long-horizon world models.

  2. arXiv Game World Models & Generation54

    GameVerse releases minute-scale gameplay video dataset for long-horizon interactive world modeling

    Researchers released GameVerse, a minute-scale gameplay video dataset for long-horizon interactive world modeling, containing 69.7K videos, 1,655 hours and 349.6 million frames, covering first-person and third-person gameplay. Of this, 1,358 hours are continuous two-minute sequences, an order of magnitude longer than previous datasets; each video comes with a camera trajectory that stays consistent across minutes, plus annotations aligned to event boundaries that distinguish static scenes from dynamic foreground events. The paper states that training on this dataset can significantly extend the time horizon of interactive world models.

    Why it matters: The dataset extends usable sequences to the two-minute range and compares it with existing public datasets on scale and annotation, making it easier to judge the training-data gap for long-horizon world models.