包裹,长文用h2/h3,列表的话原文里有几个(i)(ii)(iii)(iv)的点,对吧?首先,先处理:首先开头的部分是介绍self-consistency认证的,然后说假设不成立,然后讲建模依赖,然后是四个结论,然后是实验验证,最后是总结。 首先,首先,先把前面的介绍部分放在一个
里?不对,先看,首先第一部分:Self-consistency认证的内容,到changes那个部分?不对,不对,原文的结构: 第一段?不对,原文是一个连贯的文本,但是里面有四个结论,那四个结论应该用
- 对吧?然后,首先,开头的部分: 第一个
里放:Self-consistency certification treats repeated sampling from a language model as sequential evidence accumulation and stops when an anytime-valid e-process crosses a threshold, guaranteeing finite-sample Type I error control by Ville's inequality. Prior constructions assume sampled rollouts are conditionally independent. We show this assumption fails in practice: reasoning streams are positively correlated through shared reasoning prefixes or across-prompt heterogeneity, changing not merely certificate efficiency but the hypothesis certified. Modeling dependence via a per-unit latent state (the random-effects / intraclass-correlation structure) induces exchangeable, equi-correlated increments with correlation $ρ$. Correlation splits the single null into a realized null about the drawn belief and a population null about average model behavior, which coincide only when $ρ=0$. 这部分对吧? 然后接下来是We show: 然后后面是四个点,所以这里可以先写
We show:
?不对,或者把We show:放在p里,然后后面接ul?不对,或者,哦,对,We show:后面的内容是列表,所以:We show:
- the standard betting e-process remains anytime-valid for the realized null, not the population null;
- read as a population-null test, a suitable mixture has asymptotic crossing probability at least $π=P(μ(Z)>0)$, potentially exceeding nominal $ε$;
- the within-unit mean's effective sample size saturates at $1/ρ$, so extra within-unit rollouts cannot restore population validity;
- a prompt-level e-process restores population validity, with optimal allocation at one rollout per prompt under a pure rollout budget.
里对吧? 等下,有没有需要加h2或者h3的?原文是学术内容,没有明显的小节,要不要加?不对,用户说长文用h2/h3划分小节,这个内容不算特别长的多小节?不对,再看,有没有重要的句子可以用strong?比如,比如那个四个结论的前面的?不对,或者,哦,对,比如提到核心结论的部分?不对,不能增删内容,只是排版。等下,再检查规则:保留原文所有信息,不得增删改内容。所以不能随便加内容,h2/h3是划分小节,如果原文没有明显的小节划分,是不是不用加?不对,再看原文的结构,哦,不对,原文是一段?不对,用户给的正文是一段?不对,看用户给的正文:就是从Self-consistency开头到最后的一段?不对,那怎么分?等下,再看,哦,原文里的内容,前面的是背景和问题,然后是研究结论(四个点),然后是实验验证和总结,那要不要用h2?比如
研究背景与问题
?不对不行,因为不能增删内容啊!哦对!规则4是保留原文所有信息,不得增删改内容。哦对!所以h2/h3的标题必须是原文里有的内容?不对,原文里没有小标题啊,那是不是就不用加h2/h3?哦对,因为原文没有可以作为小节标题的内容,所以不用加,不然就增内容了。 哦对,这点很重要,不能自己加标题内容,所以如果原文没有明确的小节标题,就不用h2/h3,就用p和ul等就行。 那现在整理: 首先第一个,内容是:Self-consistency certification treats repeated sampling from a language model as sequential evidence accumulation and stops when an anytime-valid e-process crosses a threshold, guaranteeing finite-sample Type I error control by Ville's inequality. Prior constructions assume sampled rollouts are conditionally independent. We show this assumption fails in practice: reasoning streams are positively correlated through shared reasoning prefixes or across-prompt heterogeneity, changing not merely certificate efficiency but the hypothesis certified. Modeling dependence via a per-unit latent state (the random-effects / intraclass-correlation structure) induces exchangeable, equi-correlated increments with correlation $ρ$. Correlation splits the single null into a realized null about the drawn belief and a population null about average model behavior, which coincide only when $ρ=0$. 然后接下来,We show: 后面是四个结论,所以可以写成
We show:
?不对,或者把We show:放在p里,然后接ul?不对,哦,也可以是We show:
然后-
...
,内容是:Simulations and Claude Haiku 4.5 rollouts confirm this regime dependence: on GSM8K ($\hatρ=0.435$, $\hatθ=+0.96$) the gap is dormant, whereas on AIME ($\hatρ=0.480$, $\hatθ=-0.15$) the single-unit rule crosses on 24% of prompts at nominal $ε=0.05$, while the prompt-level procedure declines (0.00). This correctness result identifies which reliability claim an anytime-valid certificate actually supports under dependence. 等下,有没有需要用strong的地方?比如,有没有重要句子?比如那个核心的发现?不对,不能随便加,除非原文有重要的?哦,或者,比如“We show this assumption fails in practice”这句?不对,用户说重要句子可用strong,是可以用,不是必须,对吧?不过也可以不用?不对,再想想,有没有需要?或者,比如,那四个结论里有没有?不用也行,不过只要符合规则就行。不对,规则是“可用”,不是必须,所以如果没有必要的话可以不用,但是要确保不增删内容。 等下,再检查有没有内容遗漏:原文的所有内容都有对吧? 再核对: 1. Self-consistency... $ρ=0$. 这部分都在第一个p里,对。 2. We show: 然后四个li,对,四个点都在。 3. Simulations...under dependence. 都在最后一个p里,对。 有没有需要引用或者代码的?没有,原文没有引用,没有代码,所以不用blockquote和pre code。 那现在写出来:
Self-consistency certification treats repeated sampling from a language model as sequential evidence accumulation and stops when an anytime-valid e-process crosses a threshold, guaranteeing finite-sample Type I error control by Ville's inequality. Prior constructions assume sampled rollouts are conditionally independent. We show this assumption fails in practice: reasoning streams are positively correlated through shared reasoning prefixes or across-prompt heterogeneity, changing not merely certificate efficiency but the hypothesis certified. Modeling dependence via a per-unit latent state (the random-effects / intraclass-correlation structure) induces exchangeable, equi-correlated increments with correlation $ρ$. Correlation splits the single null into a realized null about the drawn belief and a population null about average model behavior, which coincide only when $ρ=0$.
We show:
- the standard betting e-process remains anytime-valid for the realized null, not the population null;
- read as a population-