首页 > AI前沿 > Proximal Targets: Training Looped Refiners on Problems with Many Correct Answers

Proximal Targets: Training Looped Refiners on Problems with Many Correct Answers

arXiv机器学习 2026-11-17 21:07 8 阅读 查看原文
用户需要我将一篇新闻正文排版为语义化的HTML片段。让我仔细分析内容,并应用规则: 1. 只输出HTML片段本身,不要Markdown围栏、不要解释 2. 每个自然段落用

包裹,长文用

/

划分小节 3. 列表内容用
  • ,引用用
    ,代码用
    ,重要句子可用
    4. 保留原文所有信息,不得增删改内容
    5. 不要输出//等文档级标签
    
    现在看正文内容。这是一篇关于机器学习/神经网络的技术论文摘要,描述了"proximal target"方法在解决Sudoku和图着色等 puzzle 中的应用。
    
    让我分析结构:
    - 第一句介绍了"Looped refiners"(循环优化器)和"deep supervision"(深度监督)的方法
    - 第二句描述了他们证明的结论:当puzzle有多个正确答案时,目标设定了损失下限
    - 第三句提出"proximal target"方法
    - 第四句描述实验结果
    - 第五句是与GRAM的对比
    
    这看起来是一个连续的技术段落,没有明显的列表或引用。我需要决定是否划分为小节。内容虽然较长,但逻辑连贯,是一个摘要。
    
    让我考虑是否需要

    /

    。用户说"长文用

    /

    划分小节"。这篇文字算是一个摘要,可以分成几个小节: 1. 方法与问题(Looped refiners和深度监督的问题) 2. Proximal target方法 3. 实验结果 4. 与GRAM的对比 或者我可以将其保持为段落,根据内容自然划分。让我仔细考虑。 实际上,这是一篇紧凑的论文摘要,内容连贯。我可以将整个内容作为一个段落,或者按逻辑划分为几个小节。 考虑到这是"新闻正文",而且是技术内容,我认为可以划分为几个小节使结构更清晰。 让我重新阅读并规划: 现有方法与问题: "Looped refiners such as HRM and TRM apply one small network repeatedly to a draft answer, and deep supervision trains every pass toward one stored answer per puzzle. We prove that when a puzzle has many correct answers this target sets a loss floor that no model capacity removes, and that the loss-minimising pass blends the answers into an invalid grid." Proximal Target方法: "We propose the proximal target: train each pass toward the correct answer nearest to its own input, found by scanning a list of answers, by a search solver started at the draft, or within a band of quality set by a knob; correct answers become fixed points, and a pass only has to make fewer errors than its input has for the loop to converge, which its training loss bounds." 实验结果: "With the same 1.7M-parameter network and 5 seeds, the proximal target solves 98.8% of Sudoku puzzles with 2-64 answers (standard target: 2.5%), 93.8% of sparse puzzles with more than 10,000 answers without any list of answers (0.0%), and 90.9% of 200-node 3-colourings after training on graphs of at most 50 nodes (36.9%), and different random starts end at different answers." 与GRAM的对比: "Against GRAM, a stochastic-latent reasoner, a single run is correct far more often (98.8% vs. 57.7%) and twenty runs find more of the answers (53.6% vs. 37.4%), while GRAM is slightly ahead on its own 10-node colouring benchmark; in Scrabble the knob becomes a calibrated strength dial." 这样划分很合理。我可以用

    或

    来划分小节。考虑到这是摘要性质,用

    可能比较合适,或者用

    。 用户说"长文用

    /

    划分小节",所以我可以用

    来划分。让我用

    。 现在,对于重要句子,我可以用来强调。比如: - "proximal target"可以加粗 - 关键数字和结果 但要注意规则4:"保留原文所有信息,不得增删改内容"。加粗是格式,不改变内容,所以是可以的。 让我构建HTML: ```html

    现有方法与问题

    Looped refiners such as HRM and TRM apply one small network repeatedly to a draft answer, and deep supervision trains every pass toward one stored answer per puzzle. We prove that when a puzzle has many correct answers this target sets a loss floor that no model capacity removes, and that the loss-minimising pass blends the answers into an invalid grid.

    Proximal Target 方法

    We propose the proximal target: train each pass toward the correct answer nearest to its own input, found by scanning a list of answers, by a search solver started at the draft, or within a band of quality set by a knob; correct answers become fixed points, and a pass only has to make fewer errors than its input has for the loop to converge, which its training loss bounds.

    实验结果

    With the same 1.7M-parameter network and 5 seeds, the proximal target solves 98.8% of Sudoku puzzles with 2-64 answers (standard target: 2.5%), 93.8% of sparse puzzles with more than 10,000 answers without any list of answers (0.0%), and 90.9% of 200-node 3-colourings after training on graphs of at most 50 nodes (36.9%), and different random starts end at different answers.

    与 GRAM 的对比

    Against GRAM, a stochastic-latent reasoner, a single run is correct far more often (98.8% vs. 57.7%) and twenty runs find more of the answers (53.6% vs. 37.4%), while GRAM is slightly ahead on its own 10-node colouring benchmark; in Scrabble the knob becomes a calibrated strength dial.

    ``` 等一下,我需要注意标题不要引入原文没有的信息。规则4说"不得增删改内容"。标题是我添加的小节划分,用户说"长文用

    /

    划分小节",所以添加标题是允许的,但标题应该反映小节内容。 不过,我担心添加中文标题可能被视为"增"。但用户明确要求"长文用

    /

    划分小节",所以这是排版要求的一部分。标题应该简洁描述小节内容。 让我重新考虑。也许标题应该用原文中的关键词,避免引入新概念。比如: - "现有方法与问题" → 这概括了内容 - "Proximal Target 方法" → 原文有"proximal target" - "实验结果" → 概括 - "与 GRAM 的对比" → 原文有GRAM 这些都是合理的小节标题。但为了更贴近原文,我可以用更简短的标题。 实际上,再想想,也许我不需要划分这么细。用户说"长文用

    /

    划分小节",这是可选的,不是强制的。这篇文字虽然长,但也可以作为两三个段落来处理。 不过,考虑到内容有明确的段落结构(提出问题、提出方法、实验结果、对比),划分小节会更好。让我保持上面的结构,但调整标题使其更准确。 或者,我可以只用

    而不用

    ,或者只用

    。用户说"

    /

    ",我可以选择其中一个。对于小节,用

    可能更合适,因为它是文档内的小节。但用

    也可以。 让我决定用

    ,因为它是主要的小节标题。 现在关于加粗。我加粗了一些关键术语和数字。这符合"重要句子可用"的规则。 让我再检查一下是否保留了所有原文信息: 1. "Looped refiners such as HRM and TRM apply one small network repeatedly to a draft answer, and deep supervision trains every pass toward one stored answer per puzzle." ✓ 2. "We prove that when a