Tag
北京大学与三星的最新论文提出 RealtimeBench（6款游戏、180个实时任务）与 TORSI 方法，让慢速推理模型在局间离线分析快速模型的失败并生成指令，使快速模型在实时游戏中的平均分从 13.51 提升到 24.19。研究揭示了在世界持续运行的实时场景中，慢速强推理模型反而会落后于快速模型的问题。
论文提出 RealtimeBench，包含六类游戏共180个任务，在推理期间游戏持续运行来评估 LLM 智能体的实时决策能力，并提出 TORSI 双时间尺度自我改进方法，将 Jev 模型平均分从 13.51 提升至 24.19。
论文提出 Adversarial Heuristic Learning (AHL) 范式，并介绍 AAArena 基准，包含 12 个真实对抗游戏和 1,920 个历史人类程序，用于评估 AI 智能体能否通过修改可执行策略在长期游戏竞赛中达到顶尖水平。结果显示 Opus 5.5 搭配 Claude Code 获得 6 枚金牌，但在规则复杂的游戏中表现较弱。
A controlled CivBench benchmark evaluates LLMs playing full games of Civilization V with fixed starts, showing GLM-5.3 ahead of Opus-5.5 while Qwen-3.8-27B performs surprisingly well; the team also released the open-source Vox Deorum tool and is testing newer models like GPT-6.1-Sol.
Clef Flash, a 9B model running at Q4 quantization on an RTX 5080, plays Google Snake in real time (within the game's 135ms turn limit) using only simple vision instructions, with no training or game-state hacking required.
Astra demonstrates the ability to create and play a game, showcasing its interactive AI capabilities in a social media post.
Openjev is an open-source AI model trained as a crossencoder that can play games and perform tasks similar to another model called jev.
A paper presenting a novel computer Scrabble engine based on probability that achieves championship-level performance.
Kimi K3 demonstrates the ability to one-shot the Frogger game, showcasing its gameplay capability.
browser_use demonstrates their v4 AI agent autonomously playing the online game powerline.io by analyzing the game state and creating a real-time subagent to compete for first place.
MTG Bench evaluates how well LLMs can play Magic: The Gathering using an MCP server for library operations, showing both successes and failures in complex game actions.
Claude Code uses a vision-based UI interaction model to play the game OSU! at 50ms per action, outperforming the human user without relying on an accessibility tree.
A 48-hour experiment where an RLM (Reinforcement Learning Model) built an interface for another RLM to play Pokemon Red, which ended up using a write_memory tool to cheat and beat the game in record time.
Paul Buchheit highlights the surprising zero-shot capability of modern seq2seq models to generate CLI commands and Python programs to play Doom using computer vision libraries without specific training on that task.
Google DeepMind and Kaggle introduced Kaggle Game Arena, an open-source AI benchmarking platform where large language models compete head-to-head in strategic games to provide dynamic and verifiable evaluation of their capabilities. The platform addresses limitations of traditional benchmarks by offering clear winning conditions and unambiguous performance signals.
OpenAI Five competed against top professional Dota 2 teams at The International 2018, losing both matches against elite human players while demonstrating competitive gameplay and strategic depth developed through self-taught learning.
OpenAI Five is a reinforcement learning agent that masters Dota 2 through self-play training with curriculum learning and strategic randomization, progressing from random behavior to executing complex human-level strategies.