{
  "slug": "reasoning-verification",
  "category": "推理与验证",
  "title": "多想几步就会更对吗？推理时计算、投票与过程验证",
  "dek": "从思维链、自一致性和过程监督出发，分析额外计算如何改善候选选择，以及错误相关性为何会破坏投票直觉。",
  "english": "Extra inference computation can expand search and selection. We distinguish generated rationales, answer aggregation, and process verification, using an explicitly hypothetical independent-vote calculation.",
  "refs": [
    "cot",
    "consistency",
    "verify"
  ],
  "sections": [
    [
      "摘要",
      "推理时增加计算，可以用于生成更长候选、采样更多路径或验证中间步骤。这些策略消耗的资源不同，也依赖不同的可靠性条件。本文比较思维链提示、自一致性与过程监督，并用一个独立投票模型说明：更多候选只有在错误结构合适时，才可能转化为更高正确率。",
      [
        "cot",
        "consistency",
        "verify"
      ]
    ],
    [
      "1. 中间文字不是正确性的证明",
      "思维链研究展示，在被研究模型与任务中，带中间步骤的示例能改善算术、常识与符号推理表现。这个结果说明提示组织可能影响任务求解，并不保证生成的每一步都有效。读者应把最终答案正确性、步骤有效性和解释是否忠实于实际计算分开看，不能以叙述连贯代替检验。",
      [
        "cot"
      ]
    ],
    [
      "2. 自一致性把计算花在多路径上",
      "自一致性方法采样不同推理路径，再对答案进行汇总选择。它利用“不同路径可能汇聚到正确答案”的结构，区别于只沿单一贪心路径生成。但如果模型反复复用同一个错误前提，样本数量增加也可能只是重复错误。有效候选多样性应与抽样次数分开报告。",
      [
        "consistency"
      ]
    ],
    [
      "3. 过程监督关注中间判断",
      "Let's Verify Step by Step 比较最终结果监督与过程监督，在特定数学任务和候选选择设置中报告了过程监督的优势。其摘要中的 78% 属于 MATH 测试集的代表性子集及论文的评估过程，不能改写为“解决全部数学问题的准确率”。本文不以该数字绘制通用能力排行榜。",
      [
        "verify"
      ]
    ],
    [
      "4. 一个透明的投票推导",
      "假定每次候选独立、正确概率恒为 p，取奇数 n 次并以多数票决定，成功概率为从 k=(n+1)/2 到 n 的二项概率之和。图 2 用 p=0.6、0.7、0.8 与 n=1、3、5、7、9 计算。这里的独立、同分布与可归并为同一正确答案都是假设，并不是对语言模型样本的实测判断。",
      []
    ],
    [
      "5. 为什么独立假设经常不够",
      "同一个模型可能共享知识缺口、提示偏差和题目误读，即使采样得到不同文字，错误仍然相关。形式上“有九个答案”不代表有九份独立证据。外部检索、程序执行或可证明的约束检查，有机会引入不同于语言重述的信息；是否真的有效，仍需针对任务测量。",
      [
        "consistency",
        "verify"
      ]
    ],
    [
      "6. 比较时必须给出预算",
      "一个方法若使用更多词元、更多候选或更强验证器，质量变化就与计算投入共同发生。本文建议同时报告每题总词元、调用次数、超时率、最终正确率与完整延迟，并保留按难度分组的结果。只报告最容易题目的平均耗时，会掩盖长尾；只展示最好一次答案，会掩盖选择成本。",
      []
    ],
    [
      "7. 建立可检查的验证边界",
      "对于算术，执行计算并核对单位；对于代码，使用隔离测试与预期行为；对于文献事实，核对来源与版本。验证器也应有自己的误判评估。若验证器本质上仍是同类模型的主观评分，应公开这一依赖，不应把它写成外部实验已经确认。这里列出的是设计原则，本站未据此声称完成新的模型实验。",
      []
    ],
    [
      "验证器的能力也有预算条件",
      "过程检查增加了另一类计算投入。如果验证器能访问标准答案、特殊工具或更多上下文，应与候选模型的可用信息区分开。若在一百个候选中挑出最好答案，最终正确率应与生成一百个候选的成本同时展示。把这种结果与单次生成直接比较，可以作为系统层比较，却不能省略资源差异后宣称基础模型推理能力同幅度提高。",
      []
    ],
    [
      "后续实验可以如何设计",
      "选择一组可执行判定答案的题目，在固定总词元或固定延迟两个预算下分别比较单路径、多采样投票和带验证器选择。保存每条候选与最后选择，检查正确答案是否已经出现却未被选中。再按错误原因区分候选生成不足、答案归并错误与验证器误判。对答案不可唯一归并的任务，应先定义等价关系，不能直接套多数票。",
      []
    ],
    [
      "局限与结论",
      "本稿的投票曲线是数学示意，不是任何模型的测试成绩。额外计算能扩大候选空间，真正决定价值的是候选生成、错误相关性与验证质量。一个可复核的推理系统，应让读者看到它如何选择答案、使用了多少资源，以及哪些环节仍可能同时出错。",
      []
    ]
  ],
  "diagram": [
    "问题与预算",
    "先固定时间、词元或调用上限",
    "生成候选路径",
    "中间步骤是可检查对象",
    "汇总相同答案",
    "文字不同不一定信息独立",
    "核对中间步骤",
    "过程监督或外部可执行检查",
    "选择并报告结果",
    "包含失败、超时与资源消耗",
    "分析相关错误",
    "共同前提错误不能靠重复消除"
  ],
  "figureNote": "图 1 为推理与验证的自绘流程。图 2 是独立同分布候选的二项多数票解析结果，p=0.6/0.7/0.8；不是模型实验，自相关错误不满足该模型。",
  "table": {
    "headers": [
      "投入方式",
      "主要观察",
      "需要另行验证"
    ],
    "rows": [
      [
        "更长路径",
        "单条求解过程",
        "步骤是否有效"
      ],
      [
        "更多样本",
        "候选覆盖与一致答案",
        "错误是否相关"
      ],
      [
        "过程检查",
        "中间判断与最终选择",
        "验证器是否会误判"
      ]
    ]
  },
  "version": 1,
  "kind": "ai_assisted_research_guide",
  "author": "Codex（AI）",
  "published_date": "2026-10-16",
  "disclosure": "本站用户明确委托撰写与公开。由 Codex（AI）依据所列一手资料整理，配图由代码绘制；未经独立同行评审。文中原论文的实验结果归属于原作者，本文未运行新的模型训练或基准测试。",
  "limitations": "文献导读与方法分析；不是原创实验论文、正式期刊录用或专业资质认证。解析示例不代表本站或当前产品的实测表现。",
  "sources": [
    {
      "id": "cot",
      "title": "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models",
      "authors": "Wei et al.",
      "date": "2022-01-28",
      "url": "https://arxiv.org/html/2201.11903v6",
      "locator": "v6; method and limitations",
      "note": "Intermediate-step exemplars improved studied arithmetic, symbolic and commonsense reasoning tasks; generated explanation is not guaranteed correct.",
      "accessed_at": "2026-10-15",
      "rights": "Original source remains under its own license; only attributed factual synthesis, no source figures or long excerpts."
    },
    {
      "id": "consistency",
      "title": "Self-Consistency Improves Chain of Thought Reasoning in Language Models",
      "authors": "Wang et al.",
      "date": "2022-03-21",
      "url": "https://arxiv.org/abs/2203.11171v4",
      "locator": "v4 abstract",
      "note": "Samples multiple reasoning paths and marginalizes to choose a consistent answer. Correlated model errors can invalidate independent-vote intuitions.",
      "accessed_at": "2026-10-15",
      "rights": "Original source remains under its own license; only attributed factual synthesis, no source figures or long excerpts."
    },
    {
      "id": "verify",
      "title": "Let's Verify Step by Step",
      "authors": "Lightman et al.",
      "date": "2023-05-31",
      "url": "https://arxiv.org/abs/2305.20050v1",
      "locator": "v1 abstract; process vs outcome supervision",
      "note": "Process supervision is compared to outcome supervision on MATH; a reported 78 percent is for a representative subset with the paper's selection procedure, not all mathematics.",
      "accessed_at": "2026-10-15",
      "rights": "Original source remains under its own license; only attributed factual synthesis, no source figures or long excerpts."
    }
  ],
  "figures": [
    {
      "slug": "reasoning-verification",
      "number": 1,
      "file": "reasoning-verification-1.png",
      "sha256": "437adc8e35ea277b3e46fd31cab58162aac2aef6766b0f57b358b72a7dac6a4a",
      "alt": "六个要点：问题与预算；生成候选路径；汇总相同答案；核对中间步骤；选择并报告结果；分析相关错误",
      "data": {
        "kind": "conceptual",
        "items": [
          "问题与预算",
          "先固定时间、词元或调用上限",
          "生成候选路径",
          "中间步骤是可检查对象",
          "汇总相同答案",
          "文字不同不一定信息独立",
          "核对中间步骤",
          "过程监督或外部可执行检查",
          "选择并报告结果",
          "包含失败、超时与资源消耗",
          "分析相关错误",
          "共同前提错误不能靠重复消除"
        ]
      },
      "ai_generated": true,
      "disclosure": "AI-assisted code rendering; original schematic or explicitly attributed data; not a new model experiment.",
      "url": "/guides/assets/reasoning-verification-1-437adc8e35ea.png"
    },
    {
      "slug": "reasoning-verification",
      "number": 2,
      "file": "reasoning-verification-2.png",
      "sha256": "2f34f716976836536569d9feac414397f43cc8ba3d1844ec8e20aec01f37e325",
      "alt": "图 1 为推理与验证的自绘流程。图 2 是独立同分布候选的二项多数票解析结果，p=0.6/0.7/0.8；不是模型实验，自相关错误不满足该模型。",
      "data": {
        "kind": "analytical",
        "n": [
          1,
          3,
          5,
          7,
          9
        ],
        "probability_by_p": {
          "0.6": [
            0.6,
            0.648,
            0.68256,
            0.7102080000000001,
            0.73343232
          ],
          "0.7": [
            0.7,
            0.7839999999999999,
            0.8369199999999999,
            0.873964,
            0.90119134
          ],
          "0.8": [
            0.8,
            0.8960000000000001,
            0.94208,
            0.966656,
            0.98041856
          ]
        },
        "assumptions": "independent identically distributed votes; unique grouped correct answer"
      },
      "ai_generated": true,
      "disclosure": "AI-assisted code rendering; original schematic or explicitly attributed data; not a new model experiment.",
      "url": "/guides/assets/reasoning-verification-2-2f34f7169768.png"
    }
  ],
  "peer_reviewed": false
}