首页 > AI前沿 > ShadowBench: Exposing Lexical Anchoring and the Illusion of Forgetting in Large Language Models

ShadowBench: Exposing Lexical Anchoring and the Illusion of Forgetting in Large Language Models

arXiv自然语言 2026-11-12 03:19 29 阅读 查看原文
用户要求作为专业的新闻排版工程师,将一段纯文本新闻正文排版成语义化的HTML片段。 需要遵循以下规则: 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。
    
    分析正文内容:
    正文是一段关于大语言模型(LLMs)评估基准ShadowBench的学术新闻/摘要。
    包含以下信息:
    - LLMs作为事实检索接口,当前评估依赖显式实体名称。
    - 本工作展示LLM知识在词汇上锚定:模型拥有广泛事实信息,但缺乏显式名称token时难以检索。
    - 提出ShadowBench,一个抗捷径的基准,通过Dual-Trait Association (DTA)任务评估潜在实体关联(Latent Entity Association)。
    - 评估揭示所有模型规模存在普遍的"Shadow Gap",直到前沿模型GPT-5.4和Claude-Sonnet-4.6,移除词汇锚点导致性能下降超过20%。
    - 词汇依赖暴露AI安全关键漏洞。
    - 应用于Machine Unlearning,发现SOTA算法(Gradient Difference, NPO)只实现表层词汇擦除。
    - 虽然unlearned模型在直接查询上失败,但新的Latent Entity Leakage Rate (LELR)指标显示,对于流行实体,模型在内部推理轨迹中显式重构"遗忘"实体超过74%的情况,利用残余影子知识解决关联任务。
    - 相反,对于小众实体,尽管LELR为0%,模型通过隐式属性相关性保持高影子准确率(60-90%)。
    - 最终,ShadowBench表明当前遗忘范式制造了"遗忘幻觉"(Illusion of Forgetting),证明现有方法只是表层输出过滤器,尚不可靠以实现真正的参数擦除。
    
    排版设计:
    由于是一段学术摘要/新闻,可以将其分为几个小节,使用`

    `或`

    `来划分。 比如:

    词汇锚定与ShadowBench基准

    Large Language Models... without explicit name tokens.

    Shadow Gap与AI安全漏洞

    To quantify this limitation... critical vulnerability in AI safety.

    机器遗忘与“遗忘幻觉”

    Applying ShadowBench to Machine Unlearning... parametric erasure.

    或者更简单的划分,因为原文其实可以看作3-4个逻辑段落。 原文逻辑: 1. 背景与问题提出:Large Language Models (LLMs) have emerged... without explicit name tokens. 2. 提出基准与主要发现:To quantify this limitation... over 20%. 3. 安全漏洞与机器遗忘应用:Furthermore, we demonstrate... solve associative tasks. 4. 小众实体与最终结论:Conversely, for niche entities... parametric erasure. 让我们用 `

    ` 来划分小节,增加语义化。 小节1:研究背景与核心问题

    研究背景与核心问题

    Large Language Models (LLMs) have emerged as primary interfaces for factual retrieval, yet current evaluation paradigms rely almost exclusively on explicit entity names. In this work, we demonstrate that LLM knowledge is fundamentally lexically anchored: models possess extensive factual information but find it significantly difficult to retrieve without explicit name tokens.

    小节2:ShadowBench与“影子差距”

    ShadowBench与“影子差距”

    To quantify this limitation, we introduce ShadowBench, a rigorously hardened, shortcut-resistant benchmark designed to evaluate Latent Entity Association via a novel Dual-Trait Association (DTA) task. Our evaluation reveals a pervasive "Shadow Gap" across all model scales$\unicode{x2013}$up to the frontier models GPT-5.4 and Claude-Sonnet-4.6$\unicode{x2013}$where removing lexical anchors causes performance drops of over 20%.

    小节3:AI安全漏洞与机器遗忘

    AI安全漏洞与机器遗忘

    Furthermore, we demonstrate that this lexical dependency exposes a critical vulnerability in AI safety. Applying ShadowBench to Machine Unlearning, we find that state-of-the-art algorithms (e.g., Gradient Difference, NPO) achieve only superficial lexical erasure. While unlearned models fail on direct queries, our novel Latent Entity Leakage Rate (LELR) metric reveals that for popular entities, models explicitly reconstruct the "forgotten" entity in their internal reasoning traces in over 74% of cases, utilizing residual shadow knowledge to solve associative tasks.

    小节4:潜在实体泄漏与最终结论

    潜在实体泄漏与最终结论

    Conversely, for niche entities, models retain high shadow accuracy (60-90%) via implicit attribute correlation despite 0% LELR. Ultimately, ShadowBench indicates that current unlearning paradigms create an "Illusion of Forgetting", demonstrating that existing methods act as superficial output filters and are not yet reliable for achieving true parametric erasure.

    检查是否增删改内容: 原文: Large Language Models (LLMs) have emerged as primary interfaces for factual retrieval, yet current evaluation paradigms rely almost exclusively on explicit entity names. In this work, we demonstrate that LLM knowledge is fundamentally lexically anchored: models possess extensive factual information but find it significantly difficult to retrieve without explicit name tokens. -> 完全一致。 To quantify this limitation, we introduce ShadowBench, a rigorously hardened, shortcut-resistant benchmark designed to evaluate Latent Entity Association via a novel Dual-Trait Association (DTA) task. Our evaluation reveals a pervasive "Shadow Gap" across all model scales$\unicode{x2013}$up to the frontier models GPT-5.4 and Claude-Sonnet-4.6$\unicode{x2013}$where removing lexical anchors causes performance drops of over 20%. -> 完全一致。注意 `$\unicode{x2013}$` 是原文的LaTeX数学公式写法,需要原样保留,不能改成 `–` 或 `-`,因为规则要求“不得增删改内容”。 Furthermore, we demonstrate that this lexical dependency exposes a critical vulnerability in AI safety. Applying ShadowBench to Machine Unlearning, we find that state-of-the-art algorithms (e.g., Gradient Difference, NPO) achieve only superficial lexical erasure. While unlearned models fail on direct queries, our novel Latent Entity Leakage Rate (LELR) metric reveals that for popular entities, models explicitly reconstruct the "forgotten" entity in their internal reasoning traces in over 74% of cases, utilizing residual shadow knowledge to solve associative tasks. -> 完全一致。 Conversely, for niche entities, models retain high shadow accuracy (60-90%) via implicit attribute correlation despite 0% LELR. Ultimately, ShadowBench indicates that current unlearning paradigms create an "Illusion of Forgetting," demonstrating that existing methods act as superficial output filters and are not yet reliable for achieving true parametric erasure. -> 完全一致。 检查HTML标签: 使用了 `

    ` 和 `

    `,``。没有使用文档级标签。没有Markdown围栏。没有解释。 小节标题是否算“增删改内容”? 规则2:“长文用

    /

    划分小节”。这意味着添加小节标题是符合排版规则的,属于排版操作,而不是修改正文内容。为了稳妥,小节标题可以提炼自原文,或者直接使用简短的描述性标题。这里使用的描述性标题是合理的排版行为。 再核对一下原文的特殊字符: `$\unicode{x2013}$` 需要原样输出。 最终