，这里面的内容分几个部分？首先第一部分是介绍问题和ReFrain的提出，然后是实验相关的，还有最后的代码链接。首先先处理，首先第一段：
Adding abstention to a task-adapted multimodal model can discard correct answers, including capabilities acquired during adaptation. We formulate abstention repair as maximizing correct abstention on unanswerable questions under a budget on newly lost correct answers.
然后接下来介绍ReFrain的部分，这里可以把ReFrain用？不对，原文的内容不能改，只是重要句子可以用strong，首先We propose ReFrain (Regression-Free Abstention), which combines a linear probe on first-token logits, threshold calibration on held-out image clusters, and a deterministic veto that replaces rejected outputs with the displayed "none of the above" option while preserving all other responses.
然后下一句是"Regression-free"的含义："Regression-free" denotes zero loss on the calibration set at the zero-budget target.
然后是实验设置的部分：All abstention methods in our main comparison start from the same answering-SFT checkpoint; we separately track lost correct answers, repaired errors, and retained SFT gains.
然后是Qwen2.5-VL-3B的实验结果，这里可以分？不对，原文是连贯的，哦对，然后接下来的实验数据部分，首先On Qwen2.5-VL-3B, ReFrain at zero budget retains all 57 SFT-gained answers with no losses and improves correct abstention by 7.30 percentage points.
然后下一句：At a 0.5% calibration budget, it reaches 74.05% correct abstention with 8 losses and retains 55 gains; CLIP-UP EmbLoRA reaches 64.65% with 62 losses, 41 repairs, and 40 retained gains.
然后下一句：The same probe at its default threshold loses 298 correct answers.
然后下一句关于独立验证的：With models and thresholds frozen, confirmation on 228 independent question clusters yields 66.05% correct abstention and no losses at the 0.5% point, exceeding EmbLoRA by 32.72 percentage points.
然后下一句：Ordinary refusal SFT achieves higher correct abstention at greater regression cost.
最后是代码链接：Code and reproduction materials: https://github.com/ModeLiberaL/ReFrain
等下，有没有需要用h2或者h3的？这里原文没有小节，所以不用？然后检查有没有漏内容，所有内容都要保留，不能增删改。哦对，还有有没有列表、引用、代码？这里没有列表，没有引用，代码的话那个链接不算，就正常。然后重要的地方，比如ReFrain是重要的，用strong没问题。然后链接要正常加href对吧？对，因为是HTML，所以把链接做成a标签。等下再核对原文： 原文每一句： 1. Adding abstention to a task-adapted multimodal model can discard correct answers, including capabilities acquired during adaptation. We formulate abstention repair as maximizing correct abstention on unanswerable questions under a budget on newly lost correct answers. → 这部分可以放一个？ 对，然后： 2. We propose ReFrain (Regression-Free Abstention), which combines a linear probe on first-token logits, threshold calibration on held-out image clusters, and a deterministic veto that replaces rejected outputs with the displayed "none of the above" option while preserving all other responses. → 这里ReFrain是核心方法，加strong，所以
We propose ReFrain (Regression-Free Abstention), which combines a linear probe on first-token logits, threshold calibration on held-out image clusters, and a deterministic veto that replaces rejected outputs with the displayed "none of the above" option while preserving all other responses.
3. "Regression-free" denotes zero loss on the calibration set at the zero-budget target. →"Regression-free" denotes zero loss on the calibration set at the zero-budget target.
4. All abstention methods in our main comparison start from the same answering-SFT checkpoint; we separately track lost correct answers, repaired errors, and retained SFT gains. →All abstention methods in our main comparison start from the same answering-SFT checkpoint; we separately track lost correct answers, repaired errors, and retained SFT gains.
5. On Qwen2.5-VL-3B, ReFrain at zero budget retains all 57 SFT-gained answers with no losses and improves correct abstention by 7.30 percentage points. →On Qwen2.5-VL-3B, ReFrain at zero budget retains all 57 SFT-gained answers with no losses and improves correct abstention by 7.30 percentage points.
哦，这里这个关键数据可以加strong？ 然后6. At a 0.5% calibration budget, it reaches 74.05% correct abstention with 8 losses and retains 55 gains; CLIP-UP EmbLoRA reaches 64.65% with 62 losses, 41 repairs, and 40 retained gains. →At a 0.5% calibration budget, it reaches 74.05% correct abstention with 8 losses and retains 55 gains; CLIP-UP EmbLoRA reaches 64.65% with 62 losses, 41 repairs, and 40 retained gains.
7. The same probe at its default threshold loses 298 correct answers. →The same probe at its default threshold loses 298 correct answers.
8. With models and thresholds frozen, confirmation on 228 independent question clusters yields 66.05% correct abstention and no losses at the 0.5% point, exceeding EmbLoRA by 32.72 percentage points. →With models and thresholds frozen, confirmation on 228 independent question clusters yields 66.05% correct abstention and no losses at the 0.5% point, exceeding EmbLoRA by 32.72 percentage points.
9. Ordinary refusal SFT achieves higher correct abstention at greater regression cost. →Ordinary refusal SFT achieves higher correct abstention at greater regression cost.
10. Code and reproduction materials: https://github.com/ModeLiberaL/ReFrain. →Code and reproduction materials: https://github.com/ModeLiberaL/ReFrain
等下，有没有问题？有没有增删改内容？比如加的strong只是突出，没有改内容，a标签的链接也是原文的，没问题。然后有没有小节需要加h2h3？原文没有分小节，所以不用。然后所有段落都用p包裹，对的。那现在整理成连贯的HTML片段就行，不要其他东西。哦对，还要注意不要有文档级标签，就片段本身。那现在组合起来就是：Adding abstention to a task-adapted multimodal model can discard correct answers, including capabilities acquired during adaptation. We formulate abstention repair as maximizing correct abstention on unanswerable questions under a budget on newly lost correct answers.
We propose ReFrain (Regression-Free Abstention), which combines a linear probe on first-token logits, threshold calibration on held-out image clusters, and a deterministic veto that replaces rejected outputs with the displayed "none of the above" option while preserving all other responses.
"Regression-free" denotes zero loss on the calibration set at the zero-budget target.
All abstention methods in our main comparison start from the same answering-SFT checkpoint; we separately track lost correct answers, repaired errors, and retained SFT gains.