TAG
#game-ai
0 篇评论 · 15 篇随笔
相关随笔
Tudor et al.: The scoreboard was one query away, and the agent never opened it — Fukai Reads
An arXiv preprint (submitted 2 September 2026) by seven authors from Oxford and elsewhere. They wired 76 tool endpoints into Sid Meier's Civilization VI and had language-model agents play whole games of 300+ turns. Agents queried victory progress only once every 30-75 turns (the supplied playbook recommended every 20), and in 7 of 20 losses that were foreseeable they never checked it in the final 20 turns. Commitments the agents wrote down for themselves were carried out within ten turns only 48.2%-65.8% of the time.
Inside Toru Iwatani's Philosophy — Build a Maze, Then Hide That It Is One
A study of Toru Iwatani, the designer who planned Pac-Man (1980), built only from his own lectures and interviews. He made a maze game and then drew the maze walls as thin outlines so players would not notice it was a maze. He says to think by subtraction, he gave the same account of difficulty pacing in 1986 and in 2015, and in 2011 he scolded the industry for a flood of very simple games. A man who built the outer limit of simplicity criticising simplicity: this is an attempt to line that apparent contradiction up and read it.
Kelidari et al.: A Card-Game Agent Is Only as Strong as the Yardstick You Build First — Fukai Reads
An arXiv preprint by Nima Kelidari and two co-authors, under submission to AIIDE 2026. Using Gin Rummy and a hand-written fixed expert as an immovable yardstick, they run more than a hundred controlled experiments on what makes a lightweight reinforcement learning agent strong. Win-rate against the expert is 15.0% for PPO, 22.5% for TRPO and 34.2±2.1% with every working ingredient stacked; swapping network shapes leaves win-rates overlapping, while a search that can see the hidden cards reaches 85% against 26% for one that cannot.
Battleday et al.: Measuring AI Discovery With 70 Games That Never Explain Their Rules — Fukai Reads
An arXiv preprint by Ruairidh M. Battleday and fifteen co-authors. On DiG-bench — 70 text-string games with both rules and win conditions hidden, across seven tiers, 21 released publicly — the strongest single model beat 50 games and all models pooled beat 57, while all 70 were beaten by at least one human on a first attempt. Handed the ground-truth rules, the same model jumps from 18 games to 69, and agentic harnesses did not improve on the basic one.
Cai et al.: Bringing the Authoritative Server into Learned World Models — Fukai Reads
A paper on multiplayer world models by Cai and eight colleagues at Alaya Lab, Peking University and Institute of Science Tokyo. It ports the authoritative-server contract of online games into a learned model, splitting it into a Logic Engine that advances a typed shared state and a Rendering Engine that draws each camera from it. On matched multiplayer Snake it reaches 0.764 state recovery against 0.128 for the best video-based baseline, with cross-view disagreement of 0.000 by construction, and advances 1,024 player entities for 10,000 ticks. arXiv preprint, submitted 6 August 2026, not yet peer-reviewed.
Tarun Kumar S: What Happens When You Tell a Human-Move Predictor the Last 20 Moves and the Clock — Fukai Reads
A paper by Tarun Kumar S of Peargent Labs on Otter, a chess AI that predicts human moves. Where earlier models treated each position independently, Otter conditions on the last 20 moves and on remaining clock time, reaching 55.23% top-1 accuracy with 15.3M parameters — 1.98 points above Maia 2. Of the +7.62 point gain over a board-only baseline, history contributes +5.24 and the clock +2.38. An arXiv preprint posted 5 August 2026, not yet peer reviewed.
Wang et al.: Making a Puzzle Solver the Teacher for Every Single Move — Fukai Reads
A game-AI paper by Yu Wang and colleagues. Where long-horizon puzzles reward only the final win, they convert the drop in a solver's remaining-distance-to-goal into a per-move score and mix it into training. Averaged over Sokoban, Minesweeper and Rush Hour, success rises from 16.6% to 62.1%, and on unseen difficulty from 5.9% to 28.4%. Querying the solver costs about 73 parts per million of training wall clock.
Nath et al.: Training Game AI When Streaming Dirties the Video — Fukai Reads
A paper by a Microsoft team (Nath et al.) on imitation-learning agents for streamed video games. It proposes streaming augmentations that artificially manufacture the temporally connected noise of cloud gaming and mix it into training. Even from five demonstrations, completion rises by up to ~40%, and performance loss under network lag drops from 49.82% to 7.45%.
Ying et al.: Measuring AI's General Intelligence Through Every 'Human Game' — Fukai Reads
A preprint from a team at MIT, Harvard and others that measures AI's general intelligence through games humans made. Rebuilding 100 popular App Store and Steam titles with an LLM and having seven frontier vision-language models play them, the best reached only 8.5 against a human median of 100, falling far short on memory, planning and inferring rules.
Sestini et al.: Making AAA Game NPCs Feel Authentic with Reinforcement Learning — Fukai Reads
A vision paper from the research team at Electronic Arts. It tests whether AAA game NPCs can be improved with reinforcement learning, through two real cases — goalkeeper positioning in EA SPORTS FC 25 and infantry locomotion in Battlefield 6 — and lays out seven requirements RL must meet in production. Its conclusion: RL is a tool to augment, not replace, existing game AI.
Xu 等人:生成式 AI 成为“玩法之芯”的游戏是什么样子——Fukai 解读 AI 原生游戏调查
由 Zhiyue Xu 等 6 人撰写的调查论文(arXiv 预印本),研究生成式 AI 本身成为核心循环的“AI 原生游戏”。论文以“去掉 AI 玩法是否还能成立”这一反事实标准来定义,并将实际存在的 53 部作品按游戏类型(G)与主导 AI 作用(N)两个维度分类,结果显示作品明显偏向叙事类,而用于裁定规则的用法仍然稀薄。
Özkan：让生成关卡的AI和攻略关卡的AI一起成长 — Fukai 解读
Miraç Buğra Özkan的一篇论文，让关卡生成与关卡攻略通过强化学习同时习得。在Unity中让蜂鸟(攻略方)与浮岛(生成方)一边观察彼此的成绩一边学习，在100种未知布局上达到约90.2%的攻略成功率。
Jiang 等：仅凭语言能否生成「可玩的游戏」——Fukai 读 OpenGame
香港中文大学 Yilei Jiang 等人研究的论文，介绍了一种从自然语言出发自动生成完整可玩2D网页游戏的智能体 OpenGame。通过可复用骨架与「活的调试手册」抑制集成错误，在150个课题中达到最高水准。然而谜题类游戏依然是最难处理的类型。
Li 等人：LLM 能「玩并通关」2D游戏吗 — Fukai 解读 GVGAI-LLM
Li 等人（NYU等）提出的 GVGAI-LLM 论文。该基准测试让语言模型游玩118款2D游戏，以测量推理能力与空间感知。将盘面翻译为ASCII地图后以零样本方式求解，GPT-4o-mini 在540关中的477关胜率为0%，整体胜率仅10.27%，未能达到经典搜索算法的水平。本文按「问题·方法·发现·应用场景·局限」的顺序逐一解析。
Kar：验证生成关卡的实时可通行性——Fukai 的论文导读
King's College London 的 Rishabh Kar 发表的关于 PCG（程序化内容生成）的论文。提出 Momentum 系统——在不暂停游戏的前提下，于同一游戏循环中实时验证生成关卡的可通行性。两个自主智能体在玩家前方行进，分别通过空中几何检测和地面 NavMesh 检测提前勘察路径。评估结果以从代码推导出的结构性估算呈现。
