首页 > AI前沿 > Reasoning over Non-Parametric Memory in Multilingual LLMs: A Cognitive-Based Analysis of Relational Knowledge

Reasoning over Non-Parametric Memory in Multilingual LLMs: A Cognitive-Based Analysis of Relational Knowledge

arXiv自然语言 2026-11-05 13:00 4 阅读 查看原文
仔细阅读用户的需求,任务是将一段关于 RAG 和模型关系推理的纯英文文本转换成结构化的 HTML 片段。需要遵守几个非常严格的约束。 第一步,先对正文进行语义切分。通读这段文字,可以分为四个逻辑部分: 1. 第一部分是背景和任务介绍:“Retrieval-augmented generation (RAG) assumes... rather than retrieved.” 2. 第二部分是基准测试 Rel-Bloom 的构造:“Rel-Bloom is a controlled multilingual diagnostic... (Factual, Counterfactual, Neutralized).” 3. 第三部分是具体的实验结果,包含三个发现:“Across 43 models, six languages... can improve higher-level performance in some model regimes.” 4. 第四部分是总结和资源链接:“Greater model capability... available at: https://...” 根据规则,每个自然段落用 `

` 包裹,且小节需要用 `

` 或 `

` 划分。我可以用 `

` 来表示大章节标题,或者把它作为单篇论文的引言处理。原文没有现成的小标题,直接用 `

` 加上合适的概括性文本,然后下面跟 `

` 似乎是最合理的排版。不过,仔细查看原文,前三部分其实是在同一段里的连续叙述。如果不拆分可能会违反“长文用

/

划分小节”的规则。因此,可以提取小标题插入。 例如: `

Introduction and Context

` 接第1段; `

Rel-Bloom Benchmark Construction

` 接第2段; `

Key Findings Across Models and Languages

` 接第3段; `

Conclusion and Resources

` 接第4段。 不过,原文是一整段连续的描述,人为插入标题算不算“增删改内容”?规则4明确说“不得增删改内容”。那最好不新增标题文本。干脆不写标题,或者只把每一个句号组成的长句分开,用 `

` 包裹,但这样也不满足 h2/h3。等等,原文虽然没有显式标题,但逻辑分层明显,如果加标题就违反了“保留原文所有信息,不得增删改内容”的死规定。所以绝对不能自己脑补小标题文本。 换个思路,直接将原本那一大段拆成四个 `

`,因为原文确实也是通过句号和逻辑词(First, Second, Third)进行分隔的。 接下来,识别特定的 HTML 标签应用: - 重要句子/强调:Relevance-augmented generation (RAG) 是核心概念,还有 `Rel-Bloom`,以及最后的结论句 “Greater model capability therefore does not by itself ensure reliable reasoning over supplied relational evidence.” 可以用 `` 加粗。 - 列表内容:原文有列举条件的,比如 “the same source relations are instantiated at five Bloom-inspired levels (Remember, Understand, Apply, Analyze, Evaluate) and under three veracity conditions (Factual, Counterfactual, Neutralized).” 这里并没有明确的列表排版,但可以用 `

  • ` 结构化吗?不行,原文是完整句子。不过后面有明确的三个发现:“First, accuracy drops sharply... Second, Evaluate items remain... Third, item- and relation-level difficulty rankings...” 这三个发现非常符合列表的特征!可以把这三点拆成 `
    • ...
    `,这完全保留了原有信息,并且没有改变句子原意,只是换成了 HTML 列表结构。 - 链接:最后的 “https://huggingface.co/spaces/QCRI/non-parametric-memory-relational-reasoning.” 可以做成超链接 `...`。 现在开始逐句构建: 第1句:`Retrieval-augmented generation (RAG) assumes that language models can reason over relations stated in their context rather than fall back on associations memorized during pretraining. We study this evidence-conditioned relational reasoning in isolation, with the facts supplied directly in the prompt rather than retrieved.` 第2句:`Rel-Bloom is a controlled multilingual diagnostic of 450 multiple-choice questions built from 30 Wikidata relations; the same source relations are instantiated at five Bloom-inspired levels (Remember, Understand, Apply, Analyze, Evaluate) and under three veracity conditions (Factual, Counterfactual, Neutralized).` 第3句及之后(关于三个发现): `Across 43 models, six languages, and two multilingual settings (226,800 predictions), three findings stand out.` 这句作为引导句放在段落里。 紧接着: `First, accuracy drops sharply from Apply to Analyze in every evaluated open-weight model series, reaching 47.2 percentage points for the Large Granite bucket, but not for either Gemini model.` `Second, Evaluate items remain a bottleneck even for models near ceiling on the other task types: Gemini~3.1 Pro averages 99.5\% on the other four levels but 60.7\% on Evaluate.` `Third, item- and relation-level difficulty rankings are consistently preserved across languages, indicating a difficulty structure partly shared across languages within this benchmark.` 这部分应该用 `
    • ` 包裹起来。 接下来的结论句: `Across knowledge conditions, Counterfactual contexts are generally harder than Factual ones, while removing entity identities can improve higher-level performance in some model regimes. Greater model capability therefore does not by itself ensure reliable reasoning over supplied relational evidence. Relevant resources are available at: https://huggingface.co/spaces/QCRI/non-parametric-memory-relational-reasoning.` 再仔细对照规则: 1. 只输出 HTML 片段本身,不要 Markdown 围栏,不要解释。 (非常关键,最后只能输出纯 HTML) 2. 每个自然段落用 `

      ` 包裹,长文用 `

      /

      ` 划分小节。因为没有现成小标题,强行加会违反不增改规则,所以直接全用 `

      ` 包裹合理。 3. 列表用 `

      • `,引用 `
        `,代码 `
        `,重要句子 ``。前面规划的三点发现非常适合 `
        • `。 4. 保留原文所有信息,不得增删改内容。原文中第一句和第二句是两段?“Retrieval-augmented generation (RAG) assumes... retrieved.” 是一段;“Rel-Bloom is...” 是另一段;“Across 43 models...” 是一段。拆成三个 `

          ` 完全契合原文的物理段落分布(虽然给的是一整个长文本,但明显有三个逻辑段)。 5. 不要输出 `//` 等文档级标签。 再检查一遍 HTML 细节。 LaTeX 标记 `\textsc{Rel-Bloom}` 应该去掉多余的命令,保留纯文本 `Rel-Bloom`,或者保留原本的标记?HTML 里没有对应的行内样式标记,直接写成 `Rel-Bloom` 并用 `` 强调。 原文中包含反斜杠:`99.5\%` 和 `60.7\%`,需要原样保留。 URL:`https://huggingface.co/spaces/QCRI/non-parametric-memory-relational-reasoning` 作为 `href`,文本也保持一致。 组装最终 HTML: ```html

          Retrieval-augmented generation (RAG) assumes that language models can reason over relations stated in their context rather than fall back on associations memorized during pretraining. We study this evidence-conditioned relational reasoning in isolation, with the facts supplied directly in the prompt rather than retrieved.

          Rel-Bloom is a controlled multilingual diagnostic of 450 multiple-choice questions built from 30 Wikidata relations; the same source relations are instantiated at five Bloom-inspired levels (Remember, Understand, Apply, Analyze, Evaluate) and under three veracity conditions (Factual, Counterfactual, Neutralized).