`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：
"In the legal domain, LLMs can improve reliability and interpretability by generating explicit reasoning traces. However, training LLMs to produce high-quality reasoning traces remains challenging. Existing supervised fine-tuning methods struggle when reasoning traces are significantly longer than final answers, as the learning signal for the answer becomes diluted. Meanwhile, reinforcement learning methods such as Group Relative Policy Optimization also have several limitations, including costly reward design, performance plateaus, and reward hacking. We therefore propose a two-stage training framework. In Stage I, JurisCoT-SFT employs a length-normalized dual objective to balance learning signals between reasoning traces and final answers. In Stage II, Lifecycle-Aware Backtrackable Policy Optimization dynamically evaluates and retires auxiliary reward signals based on their impact on final evaluation objectives, enabling efficient reward utilization. Our framework substantially improves legal reasoning performance on relevant tasks selected from multiple legal benchmarks, including ChJGBench, LawBench, and LegalBench. Experimental results demonstrate consistent improvements in both reasoning trace quality and final answer accuracy, while our models remain competitive with larger models."
这段文本是一个自然段。可以将其分成几个小节，用``或`
`划分，以符合“长文用`
`/`
`划分小节”的规则，同时不增删改内容。
例如：
研究背景与现有挑战
In the legal domain... diluted.
Meanwhile, reinforcement learning methods... reward hacking.
两阶段训练框架
We therefore propose a two-stage training framework. In Stage I... final answers. In Stage II... efficient reward utilization.
实验结果与评估
Our framework substantially improves... larger models.
检查是否增删改内容：
小标题是新增的，规则说“不得增删改内容”，这通常指正文内容。但为了安全起见，小标题如果是总结性的，算不算增改？
“保留原文所有信息，不得增删改内容” -> 最好不添加小标题文本，或者只使用原文中存在的词作为小标题？
如果原文只有一个长段落，强行拆分并添加小标题可能会被视为“增加内容”。
那如果不加小标题，只用``包裹整个段落？规则2说“每个自然段落用`
`包裹，长文用`
`/`
`划分小节”。如果原文只有一段，可以不划分小节，或者按照句子逻辑划分但不加文字标题？不，划分小节必须有标题。
或者，把原文的每句话或部分句群作为段落，不加h2/h3？
“长文用
/
划分小节”是条件性的，如果不算长文或者不需要划分，可以直接用`
`。但为了体现排版工程师的专业度，可以加小标题，但小标题如果不在原文中，就违反了“不得增删改内容”。
因此，最安全的做法是：不添加任何原文没有的文字作为标题，直接将原文的自然段用`
`包裹，或者把原文按语义拆成几个`
`，不使用`
`，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： "In the legal domain, LLMs can improve reliability and interpretability by generating explicit reasoning traces. However, training LLMs to produce high-quality reasoning traces remains challenging. Existing supervised fine-tuning methods struggle when reasoning traces are significantly longer than final answers, as the learning signal for the answer becomes diluted. Meanwhile, reinforcement learning methods such as Group Relative Policy Optimization also have several limitations, including costly reward design, performance plateaus, and reward hacking. We therefore propose a two-stage training framework. In Stage I, JurisCoT-SFT employs a length-normalized dual objective to balance learning signals between reasoning traces and final answers. In Stage II, Lifecycle-Aware Backtrackable Policy Optimization dynamically evaluates and retires auxiliary reward signals based on their impact on final evaluation objectives, enabling efficient reward utilization. Our framework substantially improves legal reasoning performance on relevant tasks selected from multiple legal benchmarks, including ChJGBench, LawBench, and LegalBench. Experimental results demonstrate consistent improvements in both reasoning trace quality and final answer accuracy, while our models remain competitive with larger models." 这段文本是一个自然段。可以将其分成几个小节，用``或`
`划分，以符合“长文用`
`/`
`划分小节”的规则，同时不增删改内容。 例如：
研究背景与现有挑战
In the legal domain... diluted.
Meanwhile, reinforcement learning methods... reward hacking.
两阶段训练框架
We therefore propose a two-stage training framework. In Stage I... final answers. In Stage II... efficient reward utilization.
实验结果与评估
Our framework substantially improves... larger models.
检查是否增删改内容： 小标题是新增的，规则说“不得增删改内容”，这通常指正文内容。但为了安全起见，小标题如果是总结性的，算不算增改？ “保留原文所有信息，不得增删改内容” -> 最好不添加小标题文本，或者只使用原文中存在的词作为小标题？ 如果原文只有一个长段落，强行拆分并添加小标题可能会被视为“增加内容”。 那如果不加小标题，只用``包裹整个段落？规则2说“每个自然段落用`
`包裹，长文用`
`/`
`划分小节”。如果原文只有一段，可以不划分小节，或者按照句子逻辑划分但不加文字标题？不，划分小节必须有标题。 或者，把原文的每句话或部分句群作为段落，不加h2/h3？ “长文用
/
划分小节”是条件性的，如果不算长文或者不需要划分，可以直接用`
`。但为了体现排版工程师的专业度，可以加小标题，但小标题如果不在原文中，就违反了“不得增删改内容”。 因此，最安全的做法是：不添加任何原文没有的文字作为标题，直接将原文的自然段用`
`包裹，或者把原文按语义拆成几个`
`，不使用`
/
`，因为原文并没有小节标题。 等等，如果原文是一段，我直接输出：
In the legal domain, LLMs can improve reliability and interpretability by generating explicit reasoning traces. However, training LLMs to produce high-quality reasoning traces remains challenging. Existing supervised fine-tuning methods struggle when reasoning traces are significantly longer than final answers, as the learning signal for the answer becomes diluted. Meanwhile, reinforcement learning methods such as Group Relative Policy Optimization also have several limitations, including costly reward design, performance plateaus, and reward hacking. We therefore propose a two-stage training framework. In Stage I, JurisCoT-SFT employs a length-normalized dual objective to balance learning signals between reasoning traces and final answers. In Stage II, Lifecycle-Aware Backtrackable Policy Optimization dynamically evaluates and retires auxiliary reward signals based on their impact on final evaluation objectives, enabling efficient reward utilization. Our framework substantially improves legal reasoning performance on relevant tasks selected from multiple legal benchmarks, including ChJGBench, LawBench, and LegalBench. Experimental results demonstrate consistent improvements in both reasoning trace quality and final answer accuracy, while our models remain competitive with larger models.
这样完全符合“不得增删改内容”。 或者，把几个句子分成几个``？ “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。 重要句子可用``，比如： We therefore propose a two-stage training framework. 让我们再审视规则：“长文用
/划分小节”。如果原文没有小节，就不需要划分。如果非要划分，又不能增加文字，这是不可能的。所以不使用`
/
`是合理的。
为了体现``的使用，可以把核心结论加粗：
In the legal domain, LLMs can improve reliability and interpretability by generating explicit reasoning traces. However, training LLMs to produce high-quality reasoning traces remains challenging. Existing supervised fine-tuning methods struggle when reasoning traces are significantly longer than final answers, as the learning signal for the answer becomes diluted. Meanwhile, reinforcement learning methods such as Group Relative Policy Optimization also have several limitations, including costly reward design, performance plateaus, and reward hacking. We therefore propose a two-stage training framework. In Stage I, JurisCoT-SFT employs a length-normalized dual objective to balance learning signals between reasoning traces and final answers. In Stage II, Lifecycle-Aware Backtrackable Policy Optimization dynamically evaluates and retires auxiliary reward signals based on their impact on final evaluation objectives, enabling efficient reward utilization. Our framework substantially improves legal reasoning performance on relevant tasks selected from multiple legal benchmarks, including ChJGBench, LawBench, and LegalBench. Experimental results demonstrate consistent improvements in both reasoning trace quality and final answer accuracy, while our models remain competitive with larger models.
检查是否有没有涵盖的规则： 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 -> 原文只有一个自然段，用`
`包裹。 3. 列表、引用、代码等 -> 原文没有这些内容，不需要强行添加。重要句子用`` -> 已添加。 4. 保留原文所有信息，不得增删改内容 -> 满足，没有添加任何标题文字，只加了``标签。 5. 不要输出文档级标签 -> 满足。 再确认一下，原文中是否有可以作为列表的内容？ “including costly reward design, performance plateaus, and reward hacking.” 这是列举，但原文是逗号分隔的句子，改成`
`会改变内容结构，最好保持原样，因为规则说“不得增删改内容”。 “including ChJGBench, LawBench, and LegalBench.” 同理。 所以最终的HTML就是：
In the legal domain, LLMs can improve reliability and interpretability by generating explicit reasoning traces. However, training LLMs to produce high-quality reasoning traces remains challenging. Existing supervised fine-tuning methods struggle when reasoning traces are significantly longer than final answers, as the learning signal for the answer becomes diluted. Meanwhile, reinforcement learning methods such as Group Relative Policy Optimization also have several limitations, including costly reward design, performance plateaus, and reward hacking. <
In the legal domain, LLMs can improve reliability and interpretability by generating explicit reasoning traces. However, training LLMs to produce high-quality reasoning traces remains challenging. Existing supervised fine-tuning methods struggle when reasoning traces are significantly longer than final answers, as the learning signal for the answer becomes diluted. Meanwhile, reinforcement learning methods such as Group Relative Policy Optimization also have several limitations, including costly reward design, performance plateaus, and reward hacking. We therefore propose a two-stage training framework. In Stage I, JurisCoT-SFT employs a length-normalized dual objective to balance learning signals between reasoning traces and final answers. In Stage II, Lifecycle-Aware Backtrackable Policy Optimization dynamically evaluates and retires auxiliary reward signals based on their impact on final evaluation objectives, enabling efficient reward utilization. Our framework substantially improves legal reasoning performance on relevant tasks selected from multiple legal benchmarks, including ChJGBench, LawBench, and LegalBench. Experimental results demonstrate consistent improvements in both reasoning trace quality and final answer accuracy, while our models remain competitive with larger models.
检查是否有没有涵盖的规则： 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 -> 原文只有一个自然段，用`
`包裹。 3. 列表、引用、代码等 -> 原文没有这些内容，不需要强行添加。重要句子用`` -> 已添加。 4. 保留原文所有信息，不得增删改内容 -> 满足，没有添加任何标题文字，只加了``标签。 5. 不要输出文档级标签 -> 满足。 再确认一下，原文中是否有可以作为列表的内容？ “including costly reward design, performance plateaus, and reward hacking.” 这是列举，但原文是逗号分隔的句子，改成`
-
`会改变内容结构，最好保持原样，因为规则说“不得增删改内容”。 “including ChJGBench, LawBench, and LegalBench.” 同理。 所以最终的HTML就是：
In the legal domain, LLMs can improve reliability and interpretability by generating explicit reasoning traces. However, training LLMs to produce high-quality reasoning traces remains challenging. Existing supervised fine-tuning methods struggle when reasoning traces are significantly longer than final answers, as the learning signal for the answer becomes diluted. Meanwhile, reinforcement learning methods such as Group Relative Policy Optimization also have several limitations, including costly reward design, performance plateaus, and reward hacking. <