，长文用
/
划分小节。首先看正文有没有需要划分的小节，这里没有一个明确的引言、主体、结论的分割点，但可以考虑是否需要用小标题。不过仔细看内容，前面是问题和现有文献，然后是提出的框架，中间有实验部分吗？不对，用户给的正文里没有分节的小标题，所以可能不需要h2或h3？不对，再看原文，“We propose a prospective measurement protocol and investigate the framework using synthetic composition tasks across three small language models.”这部分是不是实验部分？然后后面是发现。 接下来，原文中的两个轴用了\emph{}，在HTML里应该用标签，因为用户说重要句子可用，但这里是强调两个轴的名称。然后三个regimes是amplification, composition-limited, and acquisition-limited，这部分可以用列表吗？比如“three regimes -- amplification, composition-limited, and acquisition-limited --”，可以改成
- amplification
- composition-limited
- acquisition-limited
，这样符合列表内容用
- 的规则。 然后检查是否有引用、代码，这里没有，所以不用考虑。现在开始组织： 首先第一个段落是开头的问题，用
包裹，里面的pass@$k$保留，因为是数学符号，HTML里可以直接写。然后“We argue that...post-trainability.”这部分是观点，接着“The first axis...\emph{effective full-solution support}...feasible exploration budget.”这里的\emph{effective full-solution support}要改成effective full-solution support，同样第二个轴的\emph{atomic primitive coverage}改成atomic primitive coverage。 然后“Together, these dimensions motivate three regimes -- amplification, composition-limited, and acquisition-limited -- each predicting...”这里把三个regimes用列表，所以改成“Together, these dimensions motivate three regimes:
- amplification
- composition-limited
- acquisition-limited
, each predicting a distinct empirical signature and calling for a distinct intervention.”。 接下来“We propose a prospective measurement protocol...synthetic composition tasks across three small language models.”这是一个段落，用
。然后“We find that...nonzero support enables substantial RL gains.”这也是一个段落。最后“The framework reframes mid-training as preparation for post-training: its role is to move models into states from which RL can operate effectively.”一个段落。 现在要检查是否保留所有信息，不能增删改。比如原文中的“pass@$k$”直接保留，“\emph{...}”换成，列表部分把破折号换成冒号加列表，这样符合规则。 现在整理成HTML片段：
Does reinforcement learning teach language models new capabilities, or merely amplify ones acquired during pre-training? The literature offers evidence for both: pass@$k$ studies suggest RL reweights latent behaviours while some controlled composition studies demonstrate gains beyond simple amplification.
We argue that this question has no answer independent of the capability state of the base model, and propose a two-dimensional state space for post-trainability.
The first axis, effective full-solution support, measures whether successful trajectories can be sampled under a feasible exploration budget.
The second axis, atomic primitive coverage, measures whether the constituent skills of a solution can be executed when the burden of composition is removed.
Together, these dimensions motivate three regimes:
- amplification
- composition-limited
- acquisition-limited
, each predicting a distinct empirical signature and calling for a distinct intervention.
We propose a prospective measurement protocol and investigate the framework using synthetic composition tasks across three small language models.
We find that primitive mastery alone does not guarantee that RL can learn a composition. Checkpoints that execute every primitive reliably can fail to improve when complete solutions are never sampled during training, whereas modest composition supervision that establishes nonzero support enables substantial RL gains.
The framework reframes mid-training as preparation for post-training: its role is to move models into states from which RL can operate effectively.
等等，这里有几个问题：1. 原文中的“We argue that...post-trainability.”是一个段落，中间的句号是对的，应该作为一个段落。2. 两个轴的描述是两个句子，可以放在一个段落吗？原文中是“The first axis...budget. The second axis...removed.”这是一个段落，所以应该合并成一个
。3. 三个regimes的部分，原文中的破折号是“--amplification, composition-limited, and acquisition-limited--”，在HTML里转换成冒号加列表是可以的，但要注意列表后的逗号，“each predicting...”，所以列表后面要加逗号吗？原文是“three regimes -- amplification, composition-limited, and acquisition-limited -- each predicting...”，所以替换成“three regimes:
- ...
, each predicting...”，这样是对的。 再调整： 第一个段落：
Does reinforcement learning teach language models new capabilities, or merely amplify ones acquired during pre-training? The literature offers evidence for both: pass@$k$ studies suggest RL reweights latent behaviours while some controlled composition studies demonstrate gains beyond simple amplification.
第二个段落：
We argue that this question has no answer independent of the capability state of the base model, and propose a two-dimensional state space for post-trainability. The first axis, effective full-solution support, measures whether successful trajectories can be sampled under a feasible exploration budget. The second axis, atomic primitive coverage, measures whether the constituent skills of a solution can be executed when the burden of composition is removed. Together, these dimensions motivate three regimes:
- amplification
- composition-limited
- acquisition-limited
, each predicting a distinct empirical signature and calling for a distinct intervention.
哦，不对，原文中“We argue that...post-trainability.”是一个完整的句子，然后“The first axis...budget.”是下一个句子，“The second axis...removed.”是下一个，“Together, these dimensions...intervention.”是下一个，这四个句子是不是一个段落？看原文的断句，是的，所以应该合并成一个段落。 然后接下来的“We propose a prospective measurement protocol...models.”是一个段落，“We find that...gains.”是一个段落，“The framework reframes...effectively.”是一个段落。 现在再调整，确保每个自然段落用
，所以：
Does reinforcement learning teach language models new capabilities, or merely amplify ones acquired during pre-training? The literature offers evidence for both: pass@$k$ studies suggest RL reweights latent behaviours while some controlled composition studies demonstrate gains beyond simple amplification.
We argue that this question has no answer independent of the capability state of the base model, and propose a two-dimensional state space for post-trainability. The first axis, effective full-solution support, measures whether successful trajectories can be sampled under a feasible exploration budget. The second axis, atomic primitive coverage, measures whether the constituent skills of a solution can be executed when the burden of composition is removed. Together, these dimensions motivate three regimes:
- amplification
- composition-limited
- acquisition-limited
, each predicting a distinct empirical signature and calling for a distinct intervention.
We propose a prospective measurement protocol and investigate the framework using synthetic composition tasks across three small language models.
We find that primitive mastery alone does not guarantee that RL can learn a composition. Checkpoints that execute every primitive reliably can fail to improve when complete solutions are never sampled during training, whereas modest composition supervision that establishes nonzero support enables substantial RL gains.
包裹，里面的pass@$k$保留，因为是数学符号，HTML里可以直接写。然后“We argue that...post-trainability.”这部分是观点，接着“The first axis...\emph{effective full-solution support}...feasible exploration budget.”这里的\emph{effective full-solution support}要改成effective full-solution support，同样第二个轴的\emph{atomic primitive coverage}改成atomic primitive coverage。 然后“Together, these dimensions motivate three regimes -- amplification, composition-limited, and acquisition-limited -- each predicting...”这里把三个regimes用列表，所以改成“Together, these dimensions motivate three regimes:
- amplification
- composition-limited
- acquisition-limited
。然后“We find that...nonzero support enables substantial RL gains.”这也是一个段落。最后“The framework reframes mid-training as preparation for post-training: its role is to move models into states from which RL can operate effectively.”一个段落。 现在要检查是否保留所有信息，不能增删改。比如原文中的“pass@$k$”直接保留，“\emph{...}”换成，列表部分把破折号换成冒号加列表，这样符合规则。 现在整理成HTML片段：
Does reinforcement learning teach language models new capabilities, or merely amplify ones acquired during pre-training? The literature offers evidence for both: pass@$k$ studies suggest RL reweights latent behaviours while some controlled composition studies demonstrate gains beyond simple amplification.
We argue that this question has no answer independent of the capability state of the base model, and propose a two-dimensional state space for post-trainability.
The first axis, effective full-solution support, measures whether successful trajectories can be sampled under a feasible exploration budget.
The second axis, atomic primitive coverage, measures whether the constituent skills of a solution can be executed when the burden of composition is removed.
Together, these dimensions motivate three regimes:
- amplification
- composition-limited
- acquisition-limited
We propose a prospective measurement protocol and investigate the framework using synthetic composition tasks across three small language models.
We find that primitive mastery alone does not guarantee that RL can learn a composition. Checkpoints that execute every primitive reliably can fail to improve when complete solutions are never sampled during training, whereas modest composition supervision that establishes nonzero support enables substantial RL gains.
The framework reframes mid-training as preparation for post-training: its role is to move models into states from which RL can operate effectively.
等等，这里有几个问题：1. 原文中的“We argue that...post-trainability.”是一个段落，中间的句号是对的，应该作为一个段落。2. 两个轴的描述是两个句子，可以放在一个段落吗？原文中是“The first axis...budget. The second axis...removed.”这是一个段落，所以应该合并成一个。3. 三个regimes的部分，原文中的破折号是“--amplification, composition-limited, and acquisition-limited--”，在HTML里转换成冒号加列表是可以的，但要注意列表后的逗号，“each predicting...”，所以列表后面要加逗号吗？原文是“three regimes -- amplification, composition-limited, and acquisition-limited -- each predicting...”，所以替换成“three regimes:
- ...
Does reinforcement learning teach language models new capabilities, or merely amplify ones acquired during pre-training? The literature offers evidence for both: pass@$k$ studies suggest RL reweights latent behaviours while some controlled composition studies demonstrate gains beyond simple amplification.
第二个段落：We argue that this question has no answer independent of the capability state of the base model, and propose a two-dimensional state space for post-trainability. The first axis, effective full-solution support, measures whether successful trajectories can be sampled under a feasible exploration budget. The second axis, atomic primitive coverage, measures whether the constituent skills of a solution can be executed when the burden of composition is removed. Together, these dimensions motivate three regimes:
- amplification
- composition-limited
- acquisition-limited
，所以：
Does reinforcement learning teach language models new capabilities, or merely amplify ones acquired during pre-training? The literature offers evidence for both: pass@$k$ studies suggest RL reweights latent behaviours while some controlled composition studies demonstrate gains beyond simple amplification.
We argue that this question has no answer independent of the capability state of the base model, and propose a two-dimensional state space for post-trainability. The first axis, effective full-solution support, measures whether successful trajectories can be sampled under a feasible exploration budget. The second axis, atomic primitive coverage, measures whether the constituent skills of a solution can be executed when the burden of composition is removed. Together, these dimensions motivate three regimes:
- amplification
- composition-limited
- acquisition-limited
We propose a prospective measurement protocol and investigate the framework using synthetic composition tasks across three small language models.
We find that primitive mastery alone does not guarantee that RL can learn a composition. Checkpoints that execute every primitive reliably can fail to improve when complete solutions are never sampled during training, whereas modest composition supervision that establishes nonzero support enables substantial RL gains.