首页 > AI前沿 > LMMR: Learned Multigranular Memory Retrieval for Small Language Model Agents

LMMR: Learned Multigranular Memory Retrieval for Small Language Model Agents

arXiv自然语言 2026-11-17 17:00 9 阅读 查看原文
用户要求将一段纯文本新闻正文排版成语义化的HTML片段。 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(这里文本较短,可以不使用或者简单划分,但最好按段落处理,或者提取关键信息作为小标题,不过规则说“不得增删改内容”,所以不能随便加标题,如果原文没有标题,就不加,或者把第一句话作为标题?不,规则说“不得增删改内容”,所以最好只用

,或者如果必须用

/

,只能从原文提取,但原文没有明显的小节标题。为了安全,全部用

,或者把背景、方法、结果等作为隐含的段落。原文是一整段。我可以把它分成几个自然段落,但原文是一整段,如果分成多个

,算不算增删改?“每个自然段落用

包裹”,原文是一个自然段落,所以用一个

包裹。或者根据语义拆分,但不改变文字。为了稳妥,直接用一个

包裹整段,或者按照句子逻辑拆分成几个

,但不增删字。其实原文就是一个长段落,用一个

最安全。 3. 列表内容用

  • ,引用用
    ,代码用
    ,重要句子可用。原文没有列表、引用、代码。重要句子可以用,比如结论或核心贡献。
    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。
    
    原文:
    Small language model agents have limited space to read the memory needed for multi-hop questions. Relevant facts are often scattered across records whose surrounding text consumes much of that space. We propose LMMR, a learned multigranular memory retrieval method that selects both the next source and how much of it to read. LMMR links paragraphs and sentence fragments to their original sources and pairs views that supply the same missing support during training. Relevant distractors and repeated information provide contrasting examples. At inference, the reranker scores what each candidate adds to an anchor passage, and a selector chooses the evidence within a token budget. On 22,398 official validation questions from HotpotQA, 2WikiMultihopQA, and MuSiQue, a frozen 2B answer model reaches 61.47, 43.26, and 30.58 F1 under a 400-token budget. The gains over query expansion are 3.06, 2.14, and 4.55 points, while memory input decreases by 23.1-26.9%. LMMR also uses 23.5-27.7% fewer input tokens than a coarse retriever trained on the same questions. Against that retriever followed by LLMLingua-2 at comparable input lengths, F1 gains reach 3.99 points and are significant on HotpotQA and 2WikiMultihopQA after Holm correction. All eight small answer models evaluated on HotpotQA improve over query expansion and MMR in both F1 and exact match.
    
    排版思路:
    原文是一段学术摘要/新闻。可以将其分为几个

    ,或者保持为一个

    。为了体现“排版”,可以按逻辑分成几个

    :

    Small language model agents have limited space to read the memory needed for multi-hop questions. Relevant facts are often scattered across records whose surrounding text consumes much of that space.

    We propose LMMR, a learned multigranular memory retrieval method that selects both the next source and how much of it to read. LMMR links paragraphs and sentence fragments to their original sources and pairs views that supply the same missing support during training. Relevant distractors and repeated information provide contrasting examples.

    At inference, the reranker scores what each candidate adds to an anchor passage, and a selector chooses the evidence within a token budget.

    On 22,398 official validation questions from HotpotQA, 2WikiMultihopQA, and MuSiQue, a frozen 2B answer model reaches 61.47, 43.26, and 30.58 F1 under a 400-token budget. The gains over query expansion are 3.06, 2.14, and 4.55 points, while memory input decreases by 23.1-26.9%. LMMR also uses 23.5-27.7% fewer input tokens than a coarse retriever trained on the same questions.

    Against that retriever followed by LLMLingua-2 at comparable input lengths, F1 gains reach 3.99 points and are significant on HotpotQA and 2WikiMultihopQA after Holm correction. All eight small answer models evaluated on HotpotQA improve over query expansion and MMR in both F1 and exact match.

    检查是否增删改内容: 拆分段落不改变文字内容,只是排版。 加LMMR,重要句子/核心方法加粗,符合“重要句子可用”。 没有列表、引用、代码,所以不需要
      等。 不需要

      /

      ,因为原文没有小节标题,强行添加标题属于“增删改内容”。 再审视规则2:“每个自然段落用

      包裹,长文用

      /

      划分小节”。如果原文不是长文,或者没有明显的小节,就只用

      。 规则4:“保留原文所有信息,不得增删改内容”。拆分段落不违反,加HTML标签不违反。 确认HTML片段:

      Small language model agents have limited space to read the memory needed for multi-hop questions. Relevant facts are often scattered across records whose surrounding text consumes much of that space.

      We propose LMMR, a learned multigranular memory retrieval method that selects both the next source and how much of it to read. LMMR links paragraphs and sentence fragments to their original sources and pairs views that supply the same missing support during training. Relevant distractors and repeated information provide contrasting examples.

      At inference, the reranker scores what each candidate adds to an anchor passage, and a selector chooses the evidence within a token budget.

      On 22,398 official validation questions from HotpotQA, 2WikiMultihopQA, and MuSiQue, a frozen 2B answer model reaches 61.47, 43.26, and 30.58 F1 under a 400-token budget. The gains over query expansion are 3.06, 2.14, and 4.55 points, while memory input decreases by 23.1-26.9%. LMMR also uses 23.5-27.7% fewer input tokens than a coarse retriever trained on the same questions.

      Against that retriever followed by LLMLingua-2 at comparable input lengths, F1 gains reach 3.99 points and are significant on HotpotQA and 2WikiMultihopQA after Holm correction. All eight small answer models evaluated on HotpotQA improve over query expansion and MMR in both F1 and exact match.

      检查文字是否完全一致: Small language model agents have limited space to read the memory needed for multi-hop questions. Relevant facts are often scattered across records whose surrounding text consumes much of that space. (一致) We propose LMMR, a learned multigranular memory retrieval method that selects both the next source and how much of it to read. LMMR links paragraphs and sentence fragments to their original sources and pairs views that supply the same missing support during training. Relevant distractors and repeated information provide contrasting examples. (一致,LMMR加了strong) At inference, the reranker scores what each candidate adds to an anchor passage, and a selector chooses the evidence within a token budget. (一致) On 22,398 official validation questions from HotpotQA, 2WikiMultihopQA, and MuSiQue, a frozen 2B answer model reaches 61.47, 43.26, and 30.58 F1 under a 400-token budget. The gains over query expansion are 3.06, 2.14, and 4.55 points, while memory input decreases by 23.1-26.9%. LMMR