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 7

Oct 15Thu121–140
  1. arXiv Board & Classic Game AI35

    Why accuracy fails to certify theory of mind in multi-agent teams

    The paper argues that modules in multi-agent cooperation that infer actions from teammates should not be certified on accuracy alone, because the cost of errors and the correlation between errors change the acceptable accuracy threshold. The authors tested this in task allocation, Level-Based Foraging and Hanabi, finding that accuracy fails in both directions. In Hanabi, where errors compound, it approved a harmful LLM deployment.

    Why it matters: The paper shifts the certification standard for theory-of-mind modules from predictive accuracy to deployment cost and error dependence, offering a direct reference for evaluating whether collaborative AI can be released.

Oct 14Wed
  1. arXiv Board & Classic Game AI42

    What and where in identifying and quantifying conventions in humans and AI agents

    The paper proposes a convention gap metric in the card game Hanabi to compare how humans and layered agents use conventions. In two independently collected sets of human games, humans have a larger gap than every agent tier. On conventions both sides share, however, humans only match Tier 2, and their advantage is concentrated in clues that point to untouched cards. Convention use relates to players' knowledge of their cards and is most common in the mid-game. Masking several high-gap situations has a much larger effect than the sum of masking each one alone.

    Why it matters: The paper turns the convention gap into a measurable metric, which can be used to compare how differently humans and agents of various tiers develop tacit understanding in cooperative games.

  2. arXiv Game World Models & Generation35

    Paper proposes CGWMs, learning interpretable world model concept representations with temporal logic

    The paper proposes Concept-Grounded World Models (CGWMs), which learn interpretable concept representations from observations using complete knowledge expressed in temporal logic rather than black-box representations. Planning is formulated as differentiable inference, optimizing agent actions via gradient ascent to satisfy goals. In comparisons with pure-neural and neurosymbolic baselines in two game environments, CGWMs' concept grounding is more accurate, holds over long horizons as well, and plans more robustly both in and out of distribution.

    Why it matters: The paper places temporal-logic knowledge in the same framework as concept learning and planning for world models, and compares against pure-neural and neurosymbolic baselines in two game environments, making it a basis for judging the method trade-offs of this line of work.

  3. arXiv Board & Classic Game AI40

    Why accuracy fails to certify theory of mind in multi-agent teams

    The paper argues that modules in multi-agent cooperation that infer actions from teammates should not be certified on accuracy alone, because the cost of errors and the correlation between errors change the acceptable accuracy threshold. The authors tested this in task allocation, Level-Based Foraging and Hanabi, finding that accuracy fails in both directions. In Hanabi, where errors compound, it approved a harmful LLM deployment.

    Why it matters: The paper shifts the certification standard for theory-of-mind modules from predictive accuracy to deployment cost and error dependence, offering a direct reference for evaluating whether collaborative AI can be released.

Oct 13Tue
  1. arXiv Game AI47

    Creative destruction: Human-AI competition in an experimental market

    A study with three experiments (N = 1,025) finds that in a game design market, games independently designed by AI outperform those from human producers, including both amateur and professional creators. Human producers did not anticipate this advantage. Telling consumers that a game was designed by humans also does not raise their willingness to pay.

    Why it matters: Experimental markets let human creators, amateurs and professionals, compete directly with games independently designed by AI, which makes it possible to test whether the human edge in taste actually holds.

Oct 12Mon
  1. QbitAI61

    Huawei's Memento 3 clears all 25 games in ARC-AGI-3 with 56% fewer actions than humans

    Huawei's Noah's Ark Lab, with UCL and other teams, proposed Memento 3, which cleared all 25 games and 183 levels in the public ARC-AGI-3 set, scoring a perfect 100 RHAE. It does not update model weights or fine-tune for specific games; using Claude Opus 5 as its base model, it recursively self-improves through a continuously updated Markdown rulebook and an executable Python world model, completing the run in 7,518 total actions — about 56% fewer than the 17,135 taken by humans.

    Why it matters: The method does not update model weights, relying instead on an external rulebook and an executable world model for continuous iteration, offering a transferable approach to optimizing agent efficiency under unknown rules.

  2. arXiv Game AI43

    Esports World Model dataset (EWM)

    EWM is a dataset of human matches across StarCraft II, Dota 2, and Counter-Strike 2, comprising 3,053 matches and maps and about 4,880 hours of player-perspective footage. The duration of the CS2 portion is an estimate.

    Why it matters: The dataset links human decisions, shared state, and each player's private field of view. Researchers studying controllable multiplayer world models can draw directly on its data organization.

Oct 11Sun
  1. QbitAI53

    Kaiming He's team releases VISTA, a multimodal visual Harness framework

    In the paper "VISTA: A Visual Harness for Reasoning in an Interactive World," Kaiming He's team proposes VISTA, a vision-native Harness framework that stores the raw frames returned by the environment along with animated intermediate frames outside the context, indexed by turn number and frame number, so that off-the-shelf multimodal models can use an inspect tool during inference to call up, crop, zoom in on, and compare them on demand, without training a new model.

    Why it matters: Compared with existing approaches that turn the environment into a text grid, VISTA stores the raw images outside the context and looks back at them on demand, and it provides comparative data across multiple benchmarks.

Oct 10Sat
  1. arXiv Board & Classic Game AI42

    MARL-BattleGrounds: A JAX-native benchmark for heterogeneous competitive multi-agent RL

    MARL-BattleGrounds is a JAX-based multi-agent reinforcement learning benchmark offering a Team Deathmatch mode with two teams of five heterogeneous agents each. It includes 52 maps, 8 endgame generalization scenarios, and a monthly round-robin Elo leaderboard. On a single RTX 5090, 1,024 parallel matches reach 50,000 5v5 game steps per second.

    Why it matters: It fills the gap between SMAX and Dota 2 with a competitive multi-agent benchmark that is cheap to reproduce, making it easier to compare different algorithms.

  2. arXiv Game AI49

    MC-EvoBench: Auditing experience-driven self-improvement in Minecraft agents with 19 evolution chains

    The paper builds a controlled Minecraft benchmark, MC-EvoBench, which audits experience-driven self-improvement in agents using 19 registered evolution chains (12 main evaluation chains and 7 calibration or probe chains). It requires the target world to be reset and allows only declared agent artifacts to cross the source-target boundary.

    Why it matters: The paper uses paired probes to separate capability changes from leftover resources and uneven source exposure, making it a useful reference design for controlled evaluation of self-improving agents.

Oct 9Fri
  1. arXiv Game AI43

    LiveSynth: A Streaming Neural Synthesizer for Instrument Cloning and Text-to-Instrument

    LiveSynth is a streaming neural synthesizer that supports zero-shot instrument cloning and text-to-instrument generation, and can be played interactively in real time. It uses a feedback-free causal Transformer conditioned on MIDI events and CLAP timbre embeddings to output continuous audio latents at 100 Hz, which a causal variational autoencoder then decodes into audio; frame-level timbre conditioning enables continuous timbre morphing, and MIDI conditioning dropout allows autonomous continuation from a performance prefix.

    Why it matters: The paper sets out the architecture and real-time performance data of a streaming neural synthesizer, useful for understanding low-latency generation for playable instruments.

  2. arXiv Game AI40

    Mine Odyssey proposes an agentic spatial intelligence benchmark with 180 tasks, rebuilding 30 real locations in Minecraft

    Yuxuan Cao and others propose Mine Odyssey, an agentic spatial intelligence benchmark that rebuilds 30 real locations across 20 countries and regions on five continents in Minecraft, with 180 natural language tasks, 20 of them outdoor scenes and 10 indoor scenes.

    Why it matters: The paper rebuilds 30 real locations in Minecraft and sets 180 tasks, making it possible to compare the spatial intelligence gaps across eight models.

  3. arXiv Game AI56

    MultiWorldBench: can independently controlled views describe one shared world

    The authors propose MultiWorldBench, a diagnostic Minecraft benchmark for testing whether independently controlled views in a multi-player world model stay consistent with one single persistent shared world; it contains 495 case configurations, seven task suites and ten capabilities.

    Why it matters: With 495 configurations, the paper compares three generative world models against the reference Engine GT on ten capabilities, and the score gaps show where the current weaknesses lie.

  4. arXiv Game AI49

    Can AI agents learn their way to the top? AAArena evaluates heuristic learning in a long-running game competition

    The paper proposes the AAArena benchmark, using 12 real adversarial games and 1,920 archived human programs to evaluate agents' learning ability in long-running competitions. Among the evaluated model and tool configurations, Opus5.5 with Claude Code took 6 gold medals, and no configuration could take the remaining 6 human ladders.

    Why it matters: AAArena organizes 12 adversarial games and 1,920 human programs into a competition-style evaluation, which can be used to observe the limits of how agents improve their strategies from limited samples.

  5. arXiv Game AI45

    AgentGarten: a code world framework for continually evolving agents

    The paper proposes the AgentGarten framework, which couples simulators and game engines to a shared neural renderer to build real-time interactive virtual environments. The simulation backend maintains persistent world state and executes program-defined interaction rules, while the renderer generates visual observations from structured conditions exported through a unified interface; the neural renderer is adapted from a pretrained video model to take geometry-conditioned input, is distilled with the proposed Adversarial Forcing, and has its inference optimized for real-time interaction.

    Why it matters: The paper wires simulators and game engines into a single neural renderer, and reports learning-efficiency results comparing an agent's 4 rounds of experience with millions of rounds of reinforcement learning.

  6. arXiv Game AI38

    AgentGarten: Code worlds for evolving agents

    AgentGarten combines simulators, game engines, and a shared neural renderer to build real-time interactive environments, in which agents perceive the world through visual observations and interact with it. The paper proposes Adversarial Forcing, which distills a pretrained video model into a geometry-conditioned neural renderer and lets the encoding of historical observations also be updated by losses from subsequent predictions. The experiments report that agents learn from only 4 rounds of experience, whereas the traditional reinforcement learning baseline requires training on the order of millions of steps.

    Why it matters: The paper writes worlds as code and hands them to a shared neural renderer to generate the visuals. Comparing its learning efficiency with traditional reinforcement learning can help you judge whether this approach to expanding environments suits your project.

Oct 8Thu
  1. arXiv Game World Models & Generation46

    Memento 3 Proposes Reflective Rulebooks to Let Frozen LLM Agents Continuously Revise World Models

    The paper proposes Memento 3, which lets a frozen large language model agent continuously learn an explicit world model through external memory: the agent maintains a natural-language rulebook recording revisable hypotheses about environment dynamics while leaving unknown parts unspecified, then compiles it into executable code for prediction and planning, using a loop of observation, reflection, rule revision, compilation, and verification to improve both the rulebook and the code via prediction error.

    Why it matters: With the underlying large model frozen, continuously updating the world model via a revisable rulebook plus code verification offers a transferable methodological path for agent learning.

  2. arXiv Game AI39

    MultiWorldBench: Do independently controlled views describe one shared world?

    MultiWorldBench uses 495 Minecraft case configurations to evaluate whether multiplayer world models let independently controlled views keep a shared, persistent world.

    Why it matters: The benchmark uses Minecraft to test whether multiplayer views share one world, and provides data on the gap between multiplayer world models and engine references.

  3. arXiv Game AI49

    MultiWorldBench: Can independently controlled views describe one shared world

    The paper proposes MultiWorldBench, a Minecraft diagnostic benchmark for consistency in multiplayer world models, comprising 495 case configurations, 7 task suites and 10 capabilities.

    Why it matters: Which dimensions should consistency in multiplayer world models be measured along? This benchmark offers a reusable evaluation framework and pinpoints the specific gaps in existing systems.

  4. arXiv Game AI39

    MultiWorldBench: Do independently controlled views describe one shared world?

    The authors present MultiWorldBench, which uses 495 Minecraft test cases to examine whether multiplayer world models can keep independently controlled views within one shared, persistent world.

    Why it matters: The benchmark breaks shared-world consistency into ten separately tested capabilities and uses engine ground truth as a reference, providing a quantified measure of the gap between multiplayer world models and real game engines.