首页 > AI前沿 > Reasoning Supervision and Capability Retention in LoRA Domain Adaptation

Reasoning Supervision and Capability Retention in LoRA Domain Adaptation

arXiv自然语言 2026-11-13 00:45 21 阅读 查看原文
用户要求将给定的英文纯文本新闻正文,按照特定的排版规则,输出语义化的HTML片段。 规则回顾: 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:
    正文共有几个自然段落?
    段落1:Fine-tuning foundation models on financial data has a wide range of applications, including identifying event types from news, monitoring risk and compliance language, and answering questions over filings and earnings calls. In this setting, the main challenge is to improve task performance without degrading the generalization capabilities of the model.
    段落2:In this paper, we present an extensive set of ablations for fine-tuning Qwen3-8B with low-rank adaptation (LoRA). We compare three training regimes: (1) response-only LoRA, where the supervised response is the label alone; (2) single-adapter reasoning, where reasoning traces from a teacher model are included alongside the labels in the training loss; and (3) dual-adapter composite training, where separate LoRA adapters are trained on traces and labels.
    段落3:Response-only LoRA improves in-domain performance and leaves general capabilities near those of the frozen model. The in-domain change is largest for calibration and argmax accuracy.
    段落4:Under single-adapter reasoning, the in-domain metrics remain behind those of the response-only regime, while MMLU-Pro can rise slightly and the other retention benchmarks stay within ordinary variation.
    段落5:Dual-adapter composite training remains on the same side of this comparison and yields no further gain.
    段落6:Finally, we compress the teacher traces by summarizing them before training, and show that shorter summaries move calibration toward the response-only regime without reaching it.
    
    需要划分小节吗?正文较短,可以不加`

    `/`

    `,或者为了体现结构加一个简单的`

    `。不过规则说“长文用

    /

    划分小节”,这篇不算长文,直接全部用`

    `包裹即可。或者加个`

    `作为标题?原文没有标题,只给了正文。所以不加标题,直接段落。 段落2中有三个训练机制,可以用`
    • `来呈现,或者保持原样。原文是:"(1) response-only LoRA, where the supervised response is the label alone; (2) single-adapter reasoning, where reasoning traces from a teacher model are included alongside the labels in the training loss; and (3) dual-adapter composite training, where separate LoRA adapters are trained on traces and labels." 为了符合列表规则,可以将其转换为列表,但必须“不得增删改内容”。如果把原文的(1)(2)(3)改成`
      • `,算不算增删改?把句子结构稍微调整以适应列表,或者直接在段落里保留。为了安全,保持段落原样,或者将这三个机制用`
          `展示。 如果改成列表:

          In this paper, we present an extensive set of ablations for fine-tuning Qwen3-8B with low-rank adaptation (LoRA). We compare three training regimes:

          • (1) response-only LoRA, where the supervised response is the label alone;
          • (2) single-adapter reasoning, where reasoning traces from a teacher model are included alongside the labels in the training loss;
          • (3) dual-adapter composite training, where separate LoRA adapters are trained on traces and labels.
          这算是语义化排版,没有增删改内容,只是改变了排版方式。不过,如果严格保留原句的标点,原句是:"...regimes: (1) ...; (2) ...; and (3) ..."。如果拆成列表,最后的标点需要调整。为了严格遵守“不得增删改内容”,最好直接在`

          `中保留原句,或者把整个文本按自然段落直接包在`

          `里。 规则3说“列表内容用

          • ”,如果原文没有明显的列表格式(比如每行一个),强行拆可能违反“不得增删改内容”的约束。因此,最安全的做法是所有自然段落都用`

            `包裹,不强行拆成列表。 重要句子可用``。比如第一段的“the main challenge is to improve task performance without degrading the generalization capabilities of the model.”可以加粗。或者不加,规则说“可用”,非必须。加一点体现排版能力。 让我们检查每个自然段: 段1:

            Fine-tuning foundation models on financial data has a wide range of applications, including identifying event types from news, monitoring risk and compliance language, and answering questions over filings and earnings calls. In this setting, the main challenge is to improve task performance without degrading the generalization capabilities of the model.

            段2:

            In this paper, we present an extensive set of ablations for fine-tuning Qwen3-8B with low-rank adaptation (LoRA). We compare three training regimes: (1) response-only LoRA, where the supervised response is the label alone; (2) single-adapter reasoning, where reasoning traces from a teacher model are included alongside the labels in the training loss; and (3) dual-adapter composite training, where separate LoRA adapters are trained on traces and labels.

            段3:

            Response-only LoRA improves in-domain performance and leaves general capabilities near those of the frozen model. The in-domain change is largest for calibration and argmax accuracy.

            段4:

            Under single-adapter reasoning, the in-domain metrics remain behind those of the response-only regime, while MMLU-Pro can rise slightly and the other retention benchmarks stay within ordinary variation.

            段5:

            Dual-adapter composite training remains on the same side of this comparison and yields no further gain.

            段6:

            Finally, we compress the teacher traces by summarizing them before training, and show that shorter summaries move calibration toward the response-only regime without reaching it.

            检查是否满足所有规则: 1. 只输出HTML片段本身,不要Markdown围栏、不要解释。 -> 满足。 2. 每个自然段落用

            包裹,长文用

            /

            划分小节。 -> 满足,没有长文,直接用

            。 3. 列表内容用

            • ,引用用
              ,代码用
              ,重要句子可用。 -> 使用了。没有列表、引用、代码,所以不需要用。
              4. 保留原文所有信息,不得增删改内容。 -> 满足,除了加标签包裹,文字完全一致。
              5. 不要输出//等文档级标签。 -> 满足。
              
              再审视一下加粗的部分,确保没有改变原文内容。
              "improve task performance without degrading the generalization capabilities of the model" -> 加粗,文字没变。
              "Response-only LoRA" -> 加粗,文字没变。
              
              最终输出的HTML片段:
              

              Fine-tuning foundation models on financial data has a wide range of applications, including identifying event types from news, monitoring risk and compliance language, and answering questions over filings and earnings calls. In this setting, the main challenge is to improve task performance without degrading the generalization capabilities of the model.

              In this paper, we present an extensive set of ablations for fine-tuning Qwen3-8B with low-rank adaptation (LoRA). We compare three training regimes: (1) response-only LoRA, where the supervised response is the label alone; (2) single-adapter reasoning, where reasoning traces from a teacher model are included alongside the labels in the training loss; and (3) dual-adapter composite training, where separate LoRA adapters are trained on traces and labels.</