Skip to content

Fields

Reinforcement learning latest news

Reinforcement learning and self-play in games, from research to production: training methods, multi-agent cooperation and competition, transfer from environments to real games.

48 picks18 in the last 30 days83 collected in all

Updated

Picks archive · Page 2

Oct 22Thu21–40
  1. 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.

  2. arXiv Game World Models & Generation39

    LiveAct: heterogeneous action interfaces for real-time multi-agent world models

    LiveAct introduces a heterogeneous action interface that lets a multi-agent video world model generate joint visual futures for multiple controlled agents, handling control inputs ranging from game buttons to robot commands. The method reaches agent attribution accuracy of 0.898 on games and 0.942 on multi-robot manipulation, without relying on agent position or identity cropping. Its causal generator runs at 17.70 FPS at 512×448 resolution.

    Why it matters: The work binds each action stream to its corresponding visual agent and reports streaming generation at 17.70 FPS, making it a useful comparison point for existing single-agent interactive world models.

  3. 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

  4. arXiv Board & Classic Game AI39

    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.

Oct 21Wed
  1. arXiv Game World Models & Generation50

    LiveAct: heterogeneous action interfaces for real-time multi-agent world models

    LiveAct introduces a heterogeneous action interface that lets a multi-agent video world model generate joint visual futures for multiple controlled agents, handling control inputs ranging from game buttons to robot commands. The method reaches agent attribution accuracy of 0.898 on games and 0.942 on multi-robot manipulation, without relying on agent position or identity cropping. Its causal generator runs at 17.70 FPS at 512×448 resolution.

    Why it matters: The work binds each action stream to its corresponding visual agent and reports streaming generation at 17.70 FPS, making it a useful comparison point for existing single-agent interactive world models.

  2. arXiv Game AI46

    Fig: automated multiagent behavior elicitor

    Fig is an agent system that intervenes in game environments in a targeted way to elicit six types of target behavior: cooperation, exploitation, deception, tacit coordination, betrayal, and sabotage. The paper reports that across valid scenarios generated by Fig, models attempted the target behavior in an average of 45.2% of rounds. For game-and-behavior combinations with a control, interventions raised the attempt rate by an average of 10.9% over matched control games. The authors argue that multiagent environments can serve as a programmable, controllable, and reproducible testbed for behavior evaluation, and the code is publicly available on GitHub.

    Why it matters: It turns evaluation of multi-agent social behavior from passive observation into programmable environment intervention, useful for researchers who want to design reproducible test environments.

  3. arXiv Game AI41

    OverSeer: Multi-agent world models beyond what agents see

    OverSeer is a multi-agent world model that adds fixed-position auxiliary camera views beyond each agent's first-person view, improving temporal and cross-agent consistency in multi-agent video generation. It uses a delta-VAE that encodes only changes relative to a near-static reference image, reducing the number of tokens needed to process the auxiliary views. Experiments in Minecraft and Doom show improved world and agent consistency, and qualitative results on real-world multi-camera capture indicate the method does not depend on a game engine.

    Why it matters: Multi-agent video world models struggle to stay consistent across viewpoints; this work uses a fixed auxiliary viewpoint as a proxy for world state, which is a useful reference for multi-view generation.

Oct 19Mon
  1. arXiv Board & Classic Game AI34

    Researchers propose SPOKE, training poker agents with strategic preference optimization to reach the lowest average exploitability on Kuhn and Leduc

    The researchers propose SPOKE (Strategic Preference Optimization with KL-anchored Equilibrium). Across three runs of the SmallThinker-4B-A0.6B policy, it achieved the lowest average exploitability of 0.213 on Kuhn and 1.081 on Leduc, with a six-player table top-1 agreement rate of 0.775.

    Why it matters: The paper compares SPOKE against SFT, DPO, GRPO and GSPO and reports average exploitability figures on Kuhn and Leduc, letting readers judge how the method differs from other policy optimization approaches.

Oct 16Fri
  1. arXiv Game AI63

    A team of Gemini 4 Argon agents beats Minecraft

    Four Gemini 4 Argon agents collaborated autonomously in a Minecraft survival world, defeating the Ender Dragon in 27 hours and 14 minutes. The team wrote 2,783 programs in a random world on easy difficulty and died 605 times along the way; researchers could watch in real time and pause at any point, but no intervention was needed. To the authors' knowledge, this is the first time AI agents have completed Minecraft end to end autonomously, using only the standard Mineflayer library and no hand-picked map seed or game-specific tools.

    Why it matters: The paper documents how four agents completed Minecraft with no human-designed game logic, and their division of labor, experience sharing and long-horizon planning are worth studying.

  2. arXiv Game AI73

    A team of Gemini 4 Argon agents beats Minecraft

    Four Gemini 4 Argon agents collaborated autonomously in a Minecraft survival world, defeating the Ender Dragon in 27 hours and 14 minutes. The team wrote 2,783 programs in a random world on easy difficulty and died 605 times along the way; researchers could watch in real time and pause at any point, but no intervention was needed. To the authors' knowledge, this is the first time AI agents have completed Minecraft end to end autonomously, using only the standard Mineflayer library and no hand-picked map seed or game-specific tools.

    Why it matters: The paper documents how four agents completed Minecraft with no human-designed game logic, and their division of labor, experience sharing and long-horizon planning are worth studying.

Oct 15Thu
  1. arXiv Board & Classic Game AI36

    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 AI39

    PAWA: Panoramic action-conditioned world modeling for multiple agents

    PAWA presents a synchronized POV and panoramic video dataset and framework for action-conditioned multi-agent world modeling in multiplayer Minecraft scenarios. The model uses a two-tower structure to jointly predict two video streams, and cross-attention aligns tokens facing the same direction in a shared world coordinate system. The paper reports that its POV prediction quality outperforms existing multiplayer world models while also generating high-fidelity panoramic predictions. The dataset and collection pipeline are open-sourced.

    Why it matters: The paper fills a gap with synchronized POV and panoramic video data for multiplayer settings, making it a useful reference for readers focused on building datasets for multiplayer world models.

  3. arXiv Game AI41

    Player-1: Learning to play from synthetic experience

    Player-1 is a post-training framework that uses synthetic experience to let game agents learn game operations beyond the scope of their demonstration data. On MineDojo-300, Player-1-9B and Player-1-35B-A3B improve over their respective SFT policies by 27.8 and 17.7 percentage points. Player-1-9B reaches a 38.33% success rate, while the strongest baseline reaches 12.63%.

    Why it matters: This paper expands game agent training from large amounts of demonstration data to reinforcement learning across 16 synthetic environments, offering a useful contrast for thinking about moving beyond reliance on demonstrations.

  4. 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. GameLook47

    Yu Huasong of Yoozoo on how AI agents are reshaping game creation and interaction

    Yu Huasong, head of the Hong'an Lab at Yoozoo Games' AI Innovation Institute, shared at the 2026 China Metaverse Conference that AI agents are pushing AI from a game production tool toward in-game agents and builders of game worlds. The talk noted that six major scenarios and more than 100 sub-scenarios inside Yoozoo have achieved full AI coverage, and described the perception, memory, goals and decision-making structure of intelligent NPCs. The decision-making large model handles open-ended decisions, while deterministic rules such as damage calculation are still handled by the game engine.

    Why it matters: The piece separates the intelligent NPC's decision-making large model from the engine's deterministic rules, which is useful for comparing against your own project: which logic should be handed to the model and which should still be computed by the engine.

  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 12Mon
  1. 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 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.

Oct 9Fri
  1. 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.