首页 > AI前沿 > ChronoDossier: Long-Form Personas as Temporal Memory Representations

ChronoDossier: Long-Form Personas as Temporal Memory Representations

arXiv自然语言 2026-11-19 10:48 16 阅读 查看原文
用户现在需要把给定的纯文本新闻正文排版成语义化的HTML片段,要遵循给定的规则。首先,先看正文的内容,整体是介绍ChronoDossier这个基准的内容。首先,开头的几段?不对,原文的段落,先看:首先第一句到“tied to life phases.”是一段?然后接下来到“evidence delivery.”是一段?然后接下来“In a contextual-stress control.”到哪里?不对,得准确划分自然段落,不能随便加内容,要保留所有信息,不增删。 首先,第一个自然段落:Long-horizon agents rely on stored records to recover a person's state as facts change. Reliable answers depend on how facts are recorded, retrieved, and presented to the reader. We treat a long-form persona dossier as a memory representation of a known latent state, an invariant identity plus a timeline of facts, routines, and decision rules tied to life phases. 这一段用

包裹。 然后第二个自然段落:We present ChronoDossier, a benchmark of 311 synthetic dossiers of 1k-10k words written by seven generators from dated timelines. Intrinsic metrics evaluate the records against their intended states; extrinsic metrics test recovery by five readers on a seeded 32-persona GPT-5.1 panel balanced across lengths (520 rated turns per condition). The linked states, records, evidence packets, and answers support tests of construction and evidence delivery. 这一段也用

。 接下来,这里可以加个

?不对,看内容,接下来是两个控制实验相关的内容?不对,原文接下来的部分:In a contextual-stress control, tag-scoped writer inputs remove detected out-of-scope anchor copying. Frozen questions authored on staged dossiers transfer less reliably to one-pass dossiers of the same states. In matched evidence controls, whole sections repair 29 of 32 GPT-5.1 and 25 of 30 Sonnet 4.6 failures labeled as packaging, compared with 2 of 8 and 1 of 8 answer-formation failures. Attribution uses separate key-based and key-free support measures, with model disagreements retained. Length remains confounded with persona and question composition; the observed between-bucket widening depends on excerpt construction. 这部分可以加个小节标题?比如

Benchmark Controls and Findings

?不对,等下,规则说长文用h2/h3划分小节,但是不能增删内容,标题是合理的划分对吧?不对,哦不对,不能增删内容?哦等下规则4是保留原文所有信息,不得增删改内容。那标题不能乱加?不对,再看规则,哦规则2是每个自然段落用p包裹,长文用h2/h3划分小节,那划分小节的标题是不是属于结构,不是内容?不对,要仔细想,原文没有标题,那是不是可以合理的划分,但是不能加原文没有的内容?不对,不对,再看原文的结构: 第一部分是背景和核心定义:从开头到“tied to life phases.” 第二部分是ChronoDossier的基本信息和评估方式:从“We present ChronoDossier”到“evidence delivery.” 第三部分是控制实验的结果:从“In a contextual-stress control”到“excerpt construction.” 第四部分是最后两句:The benchmark supports offline evaluation against known facts. Memory acquisition from interaction and validation against real users remain outside its scope. 那划分小节的话,是不是可以用

来划分不同的部分,但是标题要准确,不能加原文没有的信息?不对,等下规则4是不得增删改内容,那如果我加小节标题,是不是属于新增内容?哦,不对,再看规则2的描述:“每个自然段落用

包裹,长文用

/

划分小节”,这是结构要求,小节标题是结构标识,应该是合理的,只要不改变原文内容,只是结构划分。不对,或者是不是可以不用加标题?不对,用户要求用h2/h3划分,那得合理划分,同时标题不能凭空造,要贴合内容,而且不能添加原文没有的信息?不对,等下,再仔细看原文的内容,有没有可以作为小节标题的?不对,或者是不是可以不用硬加,但是按照内容结构,比如: 首先,第一段(背景与核心概念):

Long-horizon agents rely on stored records to recover a person's state as facts change. Reliable answers depend on how facts are recorded, retrieved, and presented to the reader. We treat a long-form persona dossier as a memory representation of a known latent state, an invariant identity plus a timeline of facts, routines, and decision rules tied to life phases.

然后

ChronoDossier Benchmark Overview

,然后下面的段落:

We present ChronoDossier, a benchmark of 311 synthetic dossiers of 1k-10k words written by seven generators from dated timelines. Intrinsic metrics evaluate the records against their intended states; extrinsic metrics test recovery by five readers on a seeded 32-persona GPT-5.1 panel balanced across lengths (520 rated turns per condition). The linked states, records, evidence packets, and answers support tests of construction and evidence delivery.

然后

Control Experiments and Results

,然后下面的段落:

In a contextual-stress control, tag-scoped writer inputs remove detected out-of-scope anchor copying. Frozen questions authored on staged dossiers transfer less reliably to one-pass dossiers of the same states. In matched evidence controls, whole sections repair 29 of 32 GPT-5.1 and 25 of 30 Sonnet 4.6 failures labeled as packaging, compared with 2 of 8 and 1 of 8 answer-formation failures. Attribution uses separate key-based and key-free support measures, with model disagreements retained. Length remains confounded with persona and question composition; the observed between-bucket widening depends on excerpt construction.

然后最后一段:

The benchmark supports offline evaluation against known facts. Memory acquisition from interaction and validation against real users remain outside its scope.

哦不对,但是等下,小节标题是不是属于新增内容?会不会违反规则4?规则4是“保留原文所有信息,不得增删改内容”,那标题是对内容的概括,是不是允许的?不对,再仔细看规则,规则2说“长文用

/

划分小节”,这是排版要求,所以划分小节的标题是排版结构的一部分,不是新增原文的信息,只是结构标识,只要标题准确概括该部分内容,不添加原文没有的事实,就没问题?不对,或者有没有可能不用加标题,直接用段落?但是用户要求长文用h2/h3划分,那这篇不算太短,应该划分。 然后,有没有需要用strong的地方?比如ChronoDossier?因为是重要的基准名称,对,这个可以。还有那些数据,比如29 of 32这些,要不要加strong?不用,除非是重要句子,但是主要是名称可以加。 等下,再检查有没有漏内容,所有原文的字都要保留,不能改。现在整理一下: 首先第一个p,里面的内容,要不要加strong?比如不需要,就是普通段落。 然后

ChronoDossier Benchmark Overview

,然后p里把ChronoDossier加strong:

We present ChronoDossier, a benchmark of 311 synthetic dossiers of 1k-10k words written by seven generators from dated timelines. Intrinsic metrics evaluate the records against their intended states; extrinsic metrics test recovery by five readers on a seeded 32-persona GPT-5.1 panel balanced across lengths (520 rated turns per condition). The linked states, records, evidence packets, and answers support tests of construction and evidence delivery.

然后

Control Experiments and Results

,然后p:

In a contextual-stress control, tag-scoped writer inputs remove detected out-of-scope anchor copying. Frozen questions authored on staged dossiers transfer less reliably to one-pass dossiers of the same states. In matched evidence controls, whole sections repair 29 of 32 GPT-5.1 and 25 of 30 Sonnet 4.6 failures labeled as packaging, compared with 2 of 8 and 1 of 8 answer-formation failures. Attribution uses separate key-based and key-free support measures, with model disagreements retained. Length remains confounded with persona and question