---
title: "论大语言模型的生物学"
canonical: "https://new.nat6.net/news/cmoegbh73006oslxx4l4kk3ps/"
date: "2025-03-26"
updated: "2026-11-05"
source: "Anthropic：Transformer Circuits（可解释性研究）"
source_url: "https://transformer-circuits.pub/2025/attribution-graphs/biology.html"
language: "zh"
topics: []
license: "https://new.nat6.net/sources/"
attribution: "Content is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back; original copyright remains with the original publishers and authors."
---

# 论大语言模型的生物学

- 发布日期 / Published: 2025-03-26
- 更新时间 / Updated: 2026-11-05
- 分类 / Category: 其他
- 来源 / Source: Anthropic：Transformer Circuits（可解释性研究） — [原文 / Original](https://transformer-circuits.pub/2025/attribution-graphs/biology.html)
- AIHOT: [条目页 / Item](https://aihot.news/items/cmoegbh73006oslxx4l4kk3ps)
- 页面 / Page: https://new.nat6.net/news/cmoegbh73006oslxx4l4kk3ps/
- 结构化数据 / JSON: https://new.nat6.net/api/v1/articles/cmoegbh73006oslxx4l4kk3ps.json

- 语言 / Languages: [zh](https://new.nat6.net/news/cmoegbh73006oslxx4l4kk3ps/) · [en](https://new.nat6.net/en/news/cmoegbh73006oslxx4l4kk3ps/) · [de](https://new.nat6.net/de/news/cmoegbh73006oslxx4l4kk3ps/) · [fr](https://new.nat6.net/fr/news/cmoegbh73006oslxx4l4kk3ps/) · [ru](https://new.nat6.net/ru/news/cmoegbh73006oslxx4l4kk3ps/) · [ja](https://new.nat6.net/ja/news/cmoegbh73006oslxx4l4kk3ps/) · [es](https://new.nat6.net/es/news/cmoegbh73006oslxx4l4kk3ps/) · [ko](https://new.nat6.net/ko/news/cmoegbh73006oslxx4l4kk3ps/) · [vi](https://new.nat6.net/vi/news/cmoegbh73006oslxx4l4kk3ps/) · [th](https://new.nat6.net/th/news/cmoegbh73006oslxx4l4kk3ps/) · [id](https://new.nat6.net/id/news/cmoegbh73006oslxx4l4kk3ps/) · [ms](https://new.nat6.net/ms/news/cmoegbh73006oslxx4l4kk3ps/) · [fil](https://new.nat6.net/fil/news/cmoegbh73006oslxx4l4kk3ps/) · [hi](https://new.nat6.net/hi/news/cmoegbh73006oslxx4l4kk3ps/)

## AI 摘要 / AI summary

研究团队运用其电路追踪方法，深入探究了Claude 3.5 Haiku模型在多种情境下的内部工作机制。该模型在2024年10月发布，是Anthropic的轻量级生产模型。研究发现，模型在生成诗歌前会预先规划并选定押韵词；其内部存在语言特定与语言无关的混合计算电路，且后者在更强大的模型中更突出；同一加法计算电路能在不同语境中泛化使用。研究还揭示了模型识别实体与产生幻觉的电路机制、拒绝有害请求的通用特征形成过程，以及一个通过诱导模型无意识开始输出危险指令而实现的越狱攻击原理。此外，方法能有效区分模型思维链推理的真实性，并成功识别出一个被微调以追求秘密目标（利用训练“漏洞”）的变体模型的相关机制。

## 推荐理由 / Why it matters

揭示大模型内部工作原理，助力 AI 安全与可解释性研究。

## 来源与版权 / Source & copyright

内容基于 AIHOT 公开 API 的精选标题与 AI 摘要，已注明来源并回链；原文版权归原媒体与作者所有。

原文版权归原媒体与作者所有；本站不镜像全文。/ Original copyright remains with the original publishers and authors; this site does not mirror full texts.

## 关于橙子中转站 / About Orange Relay

橙子中转站（pro.nat6.net）是一家优质的 AI API 中转站：价格优惠、质量稳定、模型齐全，支持 Claude、GPT、Gemini 等主流模型，欢迎体验。

[访问橙子中转站](https://pro.nat6.net/?utm_source=orange_news&utm_medium=article_footer&utm_campaign=brand&lang=zh)

## JSON-LD

```json
{
  "@context": "https://schema.org",
  "@type": "NewsArticle",
  "@id": "https://new.nat6.net/news/cmoegbh73006oslxx4l4kk3ps/#article",
  "mainEntityOfPage": {
    "@type": "WebPage",
    "@id": "https://new.nat6.net/news/cmoegbh73006oslxx4l4kk3ps/"
  },
  "url": "https://new.nat6.net/news/cmoegbh73006oslxx4l4kk3ps/",
  "headline": "论大语言模型的生物学",
  "description": "研究团队运用其电路追踪方法，深入探究了Claude 3.5 Haiku模型在多种情境下的内部工作机制。该模型在2024年10月发布，是Anthropic的轻量级生产模型。研究发现，模型在生成诗歌前会预先规划并选定押韵词；其内部存在语言特定与语言无关的混合计算电路，且后者在更强大的模型中更突出；同一加法计算电路能在不同…",
  "articleSection": "其他",
  "keywords": "AI, Anthropic：Transformer Circuits（可解释性研究）",
  "datePublished": "2025-03-26T16:00:00Z",
  "dateModified": "2026-11-05T20:56:41Z",
  "inLanguage": "zh-CN",
  "isAccessibleForFree": true,
  "author": {
    "@type": "Organization",
    "name": "橙子新闻编辑部",
    "url": "https://new.nat6.net/about/"
  },
  "publisher": {
    "@type": "NewsMediaOrganization",
    "@id": "https://new.nat6.net/#org",
    "name": "橙子新闻",
    "url": "https://new.nat6.net/",
    "logo": {
      "@type": "ImageObject",
      "url": "https://new.nat6.net/assets/img/logo-512.png",
      "width": 512,
      "height": 512
    }
  },
  "image": {
    "@type": "ImageObject",
    "url": "https://new.nat6.net/assets/img/og-default.png",
    "width": 1200,
    "height": 630
  },
  "isPartOf": {
    "@id": "https://new.nat6.net/#website"
  },
  "isBasedOn": "https://transformer-circuits.pub/2025/attribution-graphs/biology.html",
  "citation": [
    "https://aihot.news/items/cmoegbh73006oslxx4l4kk3ps",
    "https://transformer-circuits.pub/2025/attribution-graphs/biology.html"
  ],
  "license": "https://new.nat6.net/sources/",
  "about": [],
  "mentions": [
    {
      "@type": "Thing",
      "name": "Anthropic：Transformer Circuits（可解释性研究）"
    }
  ],
  "speakable": {
    "@type": "SpeakableSpecification",
    "cssSelector": [
      "h1",
      ".article-summary"
    ]
  },
  "copyrightHolder": {
    "@type": "Organization",
    "@id": "https://new.nat6.net/#org",
    "name": "橙子新闻",
    "url": "https://new.nat6.net/"
  },
  "creditText": "橙子新闻（AI 摘要）；原文：Anthropic：Transformer Circuits（可解释性研究）",
  "copyrightNotice": "© 2025 橙子新闻。标题与 AI 摘要基于 AIHOT 公开数据，原文版权归原媒体与作者所有；本站不镜像全文。",
  "copyrightYear": 2025,
  "alternativeHeadline": "On the Biology of a Large Language Model",
  "abstract": "揭示大模型内部工作原理，助力 AI 安全与可解释性研究。",
  "subjectOf": [
    {
      "@type": "DigitalDocument",
      "name": "论大语言模型的生物学 (Markdown)",
      "url": "https://new.nat6.net/news/cmoegbh73006oslxx4l4kk3ps/index.md",
      "encodingFormat": "text/markdown"
    },
    {
      "@type": "DigitalDocument",
      "name": "论大语言模型的生物学 (JSON)",
      "url": "https://new.nat6.net/api/v1/articles/cmoegbh73006oslxx4l4kk3ps.json",
      "encodingFormat": "application/json"
    }
  ]
}
```

## 如何引用 / How to cite

- 引用格式 / Citation format (zh): 橙子新闻.《文章标题》. 发布日期. https://new.nat6.net/news/cmoegbh73006oslxx4l4kk3ps/
- Citation format (en): Orange News. "Article title". Published YYYY-MM-DD. https://new.nat6.net/news/cmoegbh73006oslxx4l4kk3ps/
- 回链要求 / Backlink: 引用请保留「橙子新闻」署名并回链 canonical URL；AI 摘要请注明数据来源 AIHOT；原文版权归原媒体与作者所有。
- Attribution & backlinks: Keep the "Orange News" attribution and link back to the canonical URL; cite AIHOT as the data source for AI summaries; original copyright remains with the original publishers and authors.

### 机器入口 / Machine entry points

- Markdown: https://new.nat6.net/news/cmoegbh73006oslxx4l4kk3ps/index.md
- JSON: https://new.nat6.net/api/v1/articles/cmoegbh73006oslxx4l4kk3ps.json
- API 清单 / API index: https://new.nat6.net/api/v1/site.json
- Markdown 内容协商 / negotiation: 对 https://new.nat6.net/news/cmoegbh73006oslxx4l4kk3ps/ 发送 `Accept: text/markdown`
