按年被提到
20242024 · 16 次2026
强化学习与智能体›博弈与游戏 AI·arXiv:2402.04494 · cs.LG

Amortized Planning with Large-Scale Transformers: A Case Study on Chess

大规模 Transformer 的摊销规划:以国际象棋为例
Anian Ruoss, Grégoire Delétang, Sourabh Medapati, Jordi Grau-Moya, Li Kevin Wenliang, Elliot Catt, John Reid, Cannada A. Lewis, Joel Veness, Tim Genewein · 2024

无搜索的 Transformer 蒸馏 Stockfish,达到特级大师棋力

开始阅读 ↗PDFalphaXiv 讨论
被推荐 · 15 人首次 2024-02-08 · 最近 2026-03-09发表于 neural information processing systems
15不同的人
11活跃月份
18讨论串
8被引(OpenAlex)

作者发布 ChessBench 数据集(1000 万局棋、150 亿个数据点,由 Stockfish 16 标注走法合法性与价值),用监督学习训练至多 2.7 亿参数的 Transformer 预测状态价值、动作价值和行为克隆目标。最大的模型不做任何显式搜索,在 Lichess 闪电战中达到 2895 Elo(特级大师水平)并能解出难题,但对 Stockfish 搜索算法的蒸馏仍不完美。

HN 为什么提它

HN 评论者把它当作「不做搜索也能下出特级大师水平的棋」的证据反复引用,也有人批评特级大师水平对棋类程序其实并不算强。

实证研究难度 3 / 5约 40 分钟经得起时间
读之前最好知道
1Transformer 架构
2监督学习与行为克隆(behavioral cloning)
3国际象棋引擎与 Stockfish
4Elo 评分

HN 书评

推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及
强烈推荐QuantumGood · 2024-11

“and this news hit me as "finally, evidence enough to intrigue more people that this is possible!"”

称这个无需搜索的国际象棋引擎成果终于让更多人看到这条路可行
Something weird is happening with LLMs and Chess原评论 ↗
批评matteoraso · 2024-02

“The big problem with that paper is that grandmaster level chess is actually really bad for a chess bot.”

Grand-Master Level Chess Without Search: Modeling Choices and Their Implications原评论 ↗
批评bob1029 · 2024-10

“If we find we need to rewrite the tokenizer for every problem variant, then I argue this is just ordinary programming in a very expensive disguise.”

Grandmaster-level chess without search原评论 ↗
引用janalsncm · 2024-02

“the LLM-based model that Deep Mind announced a couple of weeks ago”

提到 DeepMind 的基于 LLM 的国际象棋模型,并拿它与 NNUE 比较
Stockfish 16.1原评论 ↗
引用adw · 2024-04

“So the last frontier of human expertise in chess is search heuristics”

感叹国际象棋里人类最后的专长领域是搜索启发式,而学习模型或可取代
EURISKO Lives原评论 ↗
引用salamo · 2024-06

“Deep Mind tried [3] this.”

指出 DeepMind 已尝试过:用 Transformer 在无搜索下学下棋
AI Appears to Rapidly Be Approaching Brick Wall Where It Can't Get Smarter原评论 ↗
看全部 19 条 →
引用sorobahn · 2024-08

“I'm planning on comparing with this paper [2] which achieves GM level chess without any search.”

打算与这篇论文对比,它无需任何搜索就达到特级大师水平
Ask HN: What are you working on (August 2024)?原评论 ↗
引用famouswaffles · 2024-09

“We have a chess transformer playing world class chess (grandmaster level) - Amortized Planning with Large-Scale Transformers.”

以论文为证:已有 Transformer 能下出特级大师水平的国际象棋
Ilya Sutskever's SSI Inc raises $1B原评论 ↗
引用QuantumGood · 2024-10

“This repository provides an implementation of our paper Grandmaster-Level Chess Without Search.”

贴出这篇“无需搜索的特级大师水平国际象棋”论文的代码实现
Grandmaster-level chess without search原评论 ↗
引用thomasahle · 2024-10

“So I guess when LLMs play chess at 2800 ELO they are just taking a vote on what top grand masters would play in the given position.”

引用 LLM 下棋达 2800 Elo 的结果,调侃它们不过是在对顶级大师的走法投票
LLMs Aren't Thinking, They're Just Counting Votes原评论 ↗
引用wenc · 2024-11

“A recent discovery showing that computer chess (which was traditionally based on searching a combinatorial space, i.e. NP-hard) that is instead now being solved with transformer models, is actually playing better at the ELO level.”

指出传统靠搜索的计算机象棋,如今改用 Transformer 后 Elo 水平反而更高
Something weird is happening with LLMs and Chess原评论 ↗
引用motoboi · 2024-11

“google developed a LLM that can play chess. And play it at grandmaster level: Amortized Planning with Large-Scale Transformers”

说明 Google 做出了能以特级大师水平下棋的 LLM
Something weird is happening with LLMs and chess原评论 ↗
引用famouswaffles · 2024-11

“In Google's paper, then titled "Grandmaster level chess without search", they evaluate turbo-instruct to have a lichess Elo of 1755 (vs bots)”

引用 Google 这篇论文对 turbo-instruct 的评估:对机器人的 lichess Elo 为 1755
OK, I can partly explain the LLM chess weirdness now原评论 ↗
引用lxgr · 2024-12

“There are several neural network based engines these days, including one that does exclusively what you describe (i.e. "patterns only", no calculation at all)”

指出已有纯靠学到的模式、完全不计算的神经网络象棋引擎
Gukesh becomes the youngest chess world champion in history原评论 ↗
引用janalsncm · 2025-03

“Even in the highly constrained realm of chess, an enormous neural net will be outclassed by a small program that can run on your phone.”

拿它作例子:即使在国际象棋里,巨大的神经网络也会输给手机上的小程序
Ask HN: Any insider takes on Yann LeCun's push against current architectures?原评论 ↗
引用famouswaffles · 2025-04

“And there's this grandmaster level transformer - Amortized Planning with Large-Scale Transformers”

列出一个达到特级大师水平的下棋 Transformer
Recent AI model progress feels mostly like bullshit原评论 ↗
引用famouswaffles · 2025-04

“Then there's this grandmaster level chess transformer - Amortized Planning with Large-Scale Transformers.”

补充提到这个达到特级大师水平的国际象棋 Transformer
OpenAI o3 and o4-mini原评论 ↗
引用pu_pe · 2026-03

“There are chess engines based on transformers, even DeepMind released one [1]. It achieved ~2900 Elo.”

举 DeepMind 发布的 Transformer 象棋引擎为例,约 2900 Elo
Based on its own charter, OpenAI should surrender the race原评论 ↗
引用pu_pe · 2026-03

“See for example the chess transformer built by DeepMind a couple of years ago which I referred to in a sibling comment [1].”

举 DeepMind 的象棋 Transformer 为例,说明象棋训练已是被解决的问题
Based on its own charter, OpenAI should surrender the race原评论 ↗
被提交为 story · 1 次