首页 > AI前沿 > ROUNDS: Source-Framed Diagnostic Revision after Multi-Turn Clinical Evidence

ROUNDS: Source-Framed Diagnostic Revision after Multi-Turn Clinical Evidence

arXiv自然语言 2026-11-15 05:45 15 阅读 查看原文
用户现在需要把给定的新闻正文排版成语义化的HTML片段,遵循要求:段落用p,合适的标题?不对,原文是一段?不对看原文内容,首先原文是一段?不对,哦原文的内容,首先要准确,不能增删改,然后如果有合适的划分?等下原文是整段的学术新闻内容?不对,先看原文的结构,首先开头是提出问题,然后介绍ROUNDS框架,然后是实验结果和结论?不对,等下要严格保留原文内容,不能改。首先,原文的整个内容?不对,等下看用户给的正文,是不是就是一段?不对,仔细看:原文的内容是连贯的一段?不对,等下要按要求,每个自然段落用p包裹,有没有适合的h2/h3?哦不对,原文是整段的话?不对,等下再仔细读:首先,第一部分是提出研究的问题?然后介绍研究的内容(ROUNDS框架)?然后是实验结果,然后是结论?不对,但是不能增删内容,所以标题的话要基于原文?不对不对,规则是长文用h2/h3划分小节,那要看原文的内容逻辑。等下原文的内容: 首先第一部分:提出问题(LLM在医疗基准高分是否转化为对话中的可靠诊断),然后说诊断对话的特点,模型需要的能力,然后介绍ROUNDS框架,然后说实验的设置和结果,然后是结论。哦,那可以划分吗?等下不对,不能增删改内容,所以标题必须是原文里的?不对,不对,规则是不能增删改内容,那如果加标题的话会不会是增加?不对等下看规则3里的划分小节,是用原文的小节?不对,原文是纯文本的一段,没有现成的小节标题啊?哦不对,等下再仔细看用户给的正文,是不是就是一整段?对,用户给的正文就是连续的一段?不对,等下看: 用户给的正文: Large language models (LLMs) achieve high scores on medical benchmarks, but does this performance translate into reliable diagnosis as a conversation unfolds? In diagnostic conversations, users can provide clinical findings across multiple turns, ask for a diagnosis without supplying answer options, and introduce opinions from others. Models must therefore form diagnoses from the available evidence and decide how to respond to subsequent advice. We introduce ROUNDS, a framework that evaluates these abilities after a fixed sequence of clinical evidence delivered over multiple turns. ROUNDS compares diagnosis with and without answer options, presents the same incorrect diagnosis under clinician and lay framing, and varies support for competing diagnoses in written five-clinician panels. Across seven models and 1,035 cases from five datasets, average accuracy was 76.9% with answer options and 57.0% without them. Clinician-framed misleading advice produced lower final accuracy than lay-framed advice in every model, with differences of 4.6 to 38.2 percentage points. Initially correct diagnoses more often became incorrect answers than abstentions, and many adopted the suggested disease. Conversely, accuracy increased as more panel members supported the correct diagnosis, including when the majority opinion remained unchanged. These findings reveal a gap between medical benchmark performance and reliable diagnosis in conversation: models respond to external opinions, but remain vulnerable to abandoning correct diagnoses without new clinical evidence. 哦,这确实是一整段?不对,那要不要划分?不对,等下规则是长文用h2/h3划分小节,那如果没有现成的小节,是不是可以用p?不对,等下,有没有合适的划分?比如,第一部分是问题引入,然后是ROUNDS框架介绍,然后是实验结果,然后是结论?但是标题要怎么写?不对,不能增删内容啊,哦不对,规则4是保留原文所有信息,不得增删改内容,所以如果加标题的话,是不是属于增加内容?不对不对,等下再仔细看规则:规则2是“每个自然段落用

包裹,长文用

/

划分小节”,哦,那如果原文的纯文本里有自然段落的话,用户给的这个是不是就是一个自然段?那是不是就用一个p?不对,等下不对,再看内容,有没有可以划分的?不对,不对,用户给的正文就是一整段,那是不是就直接用p?不对,等下有没有重要的句子可以加strong?比如“We introduce ROUNDS, a framework that evaluates these abilities after a fixed sequence of clinical evidence delivered over multiple turns.” 这里的ROUNDS可以加strong?还有最后一句的结论也可以?不对,要合理,不能乱加。等下,再理清楚: 首先,原文是一整段的话,那是不是用一个

?不对,等下有没有逻辑上的划分?哦,不对,用户给的正文是一段,那会不会我误解了?等下再看,哦,不对,是不是可以按照内容层次划分,但是标题必须是原文里的表述?不对,不能加额外的内容啊,哦!哦对,规则是不能增删改内容,所以不能自己加标题文字,因为标题文字是原文没有的,对吧?哦,原来如此!我刚才差点犯错,不能自己加h2/h3的文字,因为那属于新增内容,违反规则4。那如果原文的纯文本里没有小节标题,是一整段,那是不是就直接用

包裹?不对,等下再看规则2,“长文用

/

划分小节”,但前提是原文有小节的内容,而且不能用原文没有的文字。哦,那这里用户给的正文是一整段,没有现成的小节,所以是不是就直接一个

?不对,等下再仔细看内容,有没有可以拆分的?不对,拆分段落的话会不会改变原文的段落?原文是一个自然段,所以不能拆成多个p,对吧?因为规则1是每个自然段落用p包裹,原文是一个自然段,所以一个p。然后重要的部分可以加strong,比如ROUNDS这个框架,还有关键的数字?不对,比如“ROUNDS”可以加strong?还有最后结论的部分?等下,合理的加:比如“We introduce ROUNDS, a framework that evaluates these abilities after a fixed sequence of clinical evidence delivered over multiple turns.” 然后还有“These findings reveal a gap between medical benchmark performance and reliable diagnosis in conversation: models respond to external opinions, but remain vulnerable to abandoning correct diagnoses without new clinical evidence.” 这句是结论,可以加strong?不对,要不要?或者还有其他的?比如关键的数字?比如76.9%、57.0%?不对,不要乱加,只加重要的。等下,再确认规则,不得增删改内容,加strong是强调,不是改内容,对吧?对,规则3说重要句子可用,所以没问题。 等下,不对,再想想,有没有可能原文的结构可以划分?比如: 第一部分是问题的提出:从开头到“decide how to respond to subsequent advice.” 然后第二部分是ROUNDS框架的介绍:“We introduce ROUNDS... five-clinician panels.” 然后第三部分是实验结果:“Across seven models... unchanged.” 第四部分是结论:“These findings reveal... clinical evidence.” 但是这样的话,如果不用标题,是不是可以拆成多个p?不对,原文是一个自然段,拆分段落的话是不是属于改变原文?哦,规则是每个自然段落用p包裹,原文的纯文本是一个自然段落,所以不能拆成多个p,对吧?因为用户给的正文是连续的一段,没有换行,所以是一个自然段,所以用一个

包裹。然后里面的重点内容加strong,比如ROUNDS,还有最后的结论句?不对,再看,有没有其他的?比如“Clinician-framed misleading advice produced lower final accuracy than lay-framed advice in every model, with differences of 4.6 to 38.2 percentage points.” 这句是不是重要的?可以加吗?或者不要加太多,合理就行。 等下,再仔细核对:不能增删改内容,所以所有原文的文字都要保留,顺序不能变。那现在写出来的话:

Large language models (LLMs) achieve high scores on medical benchmarks, but does this performance translate into reliable diagnosis as a conversation unfolds? In diagnostic conversations, users can provide clinical findings across multiple turns, ask for a diagnosis without supplying answer options, and introduce opinions from others. Models must therefore form diagnoses from the available evidence and decide how to respond to subsequent advice. We introduce ROUNDS, a framework that evaluates these abilities after a fixed sequence of clinical evidence delivered over multiple turns. ROUNDS compares diagnosis with and without answer options, presents the same incorrect diagnosis under clinician and lay framing, and varies support for competing diagnoses in written five-clinician panels. Across seven models and 1,035 cases from five datasets, average accuracy was 76.9% with answer options and 57.0% without them. Clinician-framed misleading advice produced lower final accuracy than lay-framed advice in every model, with differences of 4.6 to 38.2 percentage points. Initially correct diagnoses more often became incorrect answers than abstentions, and many adopted the suggested disease. Conversely, accuracy increased as more panel members supported the correct diagnosis, including when the majority opinion remained unchanged