大模型“幻觉”问题（输出事实性错误）在医疗、法律等高危领域尤为严重。业界对抗幻觉的两大主流方法——扩大训练数据与设置防御机制——均有局限：数据无法覆盖所有事实，防御机制则常导致AI过度谨慎。
在清华大学“AI医疗新范式”论坛上，百川智能CEO王小川发布新一代医疗大模型“Baichuan-M4”及AI家庭医生“百小医”。该模型登顶三大权威榜单，攻克了医疗AI的“事实性幻觉”问题，标志着AI在医疗垂直领域的精准度与应用形态取得突破性进展。
百川智能推出Baichuan-M3 Plus医疗大模型，专为高可靠性医疗场景设计。该模型将“六源循证范式”融入训练与推理全流程，确保诊断建议、用药提示等均基于权威医学证据。其事实性幻觉率降至2.6%，相比GPT-5.2降低超30%，显著提升医疗AI的临床可信度。
谷歌与Kaggle联合推出FACTS基准套件，旨在标准化评估AI模型的事实准确性。该基准针对法律、金融、医疗等对准确性要求高的行业，将“事实性”分为“上下文事实性”和“世界知识事实性”两个场景进行综合评估，填补了当前AI评估体系的空白。
用于评估大型语言模型事实性的最新基准
衡量语言模型回答事实性问题能力的基准测试
Aloe是一款专为医疗领域设计的高性能语言模型，提供先进的文本生成和对话能力。
Xai
$21
Input tokens/M
$105
Output tokens/M
128
Context Length
Baidu
$1
$4
64
Openai
$84
Clemylia
LAM-1是lamina系列中的首个完整版本的小型语言模型，由Clemylia开发。该模型专注于创造性内容生成，经过优化能够生成富有想象力、诗意和叙事性的回答，而非事实性信息。
Tesity-T5是由Clemylia开发的编码器-解码器类型语言模型，专注于带上下文的问答任务。该模型能够根据提供的文本上下文，精准提炼并合成答案，具有事实性和综合性的回答能力。
Sami92
基于XLM-R Large微调的文本分类模型，专门用于识别德语文本中的事实性声明和非事实性声明。模型采用弱监督学习方式，先在GPT-4o标注的Telegram数据集上训练，后在手动标注数据集上继续训练，在测试集上达到0.9的准确率。
Babelscape
基于T5架构的模型，专门用于从摘要文本中提取原子声明，是摘要事实性评估流程的关键组件。
Karko
Proctora是基于混合专家架构(MoE)的模型，结合了角色扮演和事实性回答的专家模块，支持32K上下文长度，在AI-RPG评估中表现优异。
stanford-oval
WikiChat是一个基于LLaMA-2 (7B)微调的大语言模型，通过在维基百科上进行少量样本锚定来阻止聊天机器人产生幻觉，显著提高了对话的准确性和事实性。
eibakke
一个在完整Natural Questions(NQ)数据集上训练的开放领域问答模型，能够回答各种事实性问题
%20--%3e%3cdefs%3e%3cstyle%3e%20.st0%20{%20fill:%20%23061b40;%20}%20.st1%20{%20fill:%20%23306af1;%20}%20.st2%20{%20fill:%20%235ce5cf;%20}%20%3c/style%3e%3c/defs%3e%3cg%3e%3cpath%20class='st0'%20d='M55,10.5h9v3h-9v9h12V7.5h-12v3ZM64,19.5h-6v-3h6v3Z'/%3e%3cpolygon%20class='st0'%20points='69%2016.5%2078%2016.5%2078%2019.5%2069%2019.5%2069%2022.5%2081%2022.5%2081%2013.5%2072%2013.5%2072%2010.5%2081%2010.5%2081%207.5%2069%207.5%2069%2016.5'/%3e%3cpolygon%20class='st0'%20points='95%2010.5%2095%207.5%2083%207.5%2083%2022.5%2095%2022.5%2095%2019.5%2086%2019.5%2086%2016.5%2095%2016.5%2095%2013.5%2086%2013.5%2086%2010.5%2095%2010.5'/%3e%3cpath%20class='st0'%20d='M40,1.5v21h11.6l1.4-1.4v-7.6h0c0,0-1.4-1.5-1.4-1.5l1.4-1.4V2.9l-1.4-1.4h-11.6ZM50,19.5h-7v-6h7v6ZM50,10.5h-7v-6h7v6Z'/%3e%3c/g%3e%3cpath%20class='st1'%20d='M23.1,24L14.7,4.8l-4.9,11.2h3.8l-1.8,4H3.3L12.1,0H2C.9,0,0,.9,0,2v20c0,1.1.9,2,2,2h21.1Z'/%3e%3cpath%20class='st2'%20d='M34,0h-16.8l10.6,24h6.2c1.1,0,2-.9,2-2V2C36,.9,35.1,0,34,0ZM32.5,20h-4V4h4v16Z'/%3e%3c/svg%3e)
'%3e%3cpath%20d='M55.2622%2038L48.2289%2022L44.1206%2031.3333H47.2539L45.7872%2034.6667H38.7622L46.0956%2018H37.6706C36.7539%2018%2036.0039%2018.75%2036.0039%2019.6667V36.3333C36.0039%2037.25%2036.7539%2038%2037.6706%2038H55.2622Z'%20fill='white'/%3e%3cpath%20d='M64.3368%2018H50.3535L59.1535%2038H64.3368C65.2535%2038%2066.0035%2037.25%2066.0035%2036.3333V19.6667C66.0035%2018.75%2065.2535%2018%2064.3368%2018ZM63.0868%2034.6667H59.7535V21.3333H63.0868V34.6667Z'%20fill='%23FFCC66'/%3e%3cpath%20d='M89.3379%2026.75V29.25H81.8379V36.75H91.8379V24.25H81.8379V26.75H89.3379ZM84.3379%2034.25V31.75H89.3379V34.25H84.3379Z'%20fill='white'/%3e%3cpath%20d='M103.504%2036.75V29.25H96.0039V26.75H103.504V24.25H93.5039V31.75H101.004V34.25H93.5039V36.75H103.504Z'%20fill='white'/%3e%3cpath%20d='M115.17%2026.75V24.25H105.17V36.75H115.17V34.25H107.67V31.75H115.17V29.25H107.67V26.75H115.17Z'%20fill='white'/%3e%3cpath%20d='M80.1712%2029.25V29.175L78.9962%2028L80.1712%2026.825V20.425L78.9962%2019.25H69.3379V36.75H78.9962L80.1712%2035.575V29.25ZM77.6712%2034.25H71.8379V29.25H77.6712V34.25ZM77.6712%2026.75H71.8379V21.75H77.6712V26.75Z'%20fill='white'/%3e%3c/g%3e%3cg%20clip-path='url(%23clip1_644_3948)'%3e%3cpath%20d='M169.59%2022.9827L169.59%2023.6491H177.409L168.175%2032.8822L169.118%2033.8248L178.351%2024.5917L178.352%2032.4106H179.685V22.3164L179.018%2022.3157H169.59V22.9827Z'%20fill='white'/%3e%3c/g%3e%3cdefs%3e%3cclipPath%20id='clip0_644_3948'%3e%3crect%20width='80'%20height='20'%20fill='white'%20transform='translate(36%2018)'/%3e%3c/clipPath%3e%3cclipPath%20id='clip1_644_3948'%3e%3crect%20width='20'%20height='20'%20fill='white'%20transform='translate(164%2018)'/%3e%3c/clipPath%3e%3c/defs%3e%3c/svg%3e)
