已验证来源
每个故事都链接到最有力的可用证据：可用的原始来源，否则明确归因的报告。
简单的英语
发生了什么、为什么重要以及值得关注的内容——无需行话。
无填料
当信号很弱时，我们不会发布任何内容，而是填充提要。
更多故事
9 故事创新
预印本发现，通用的多语言 AI 指标可能导致跨语言比较出现偏差
一篇新的预印本报告指出，用于跨语言比较语言模型的几种归一化指标可能会引入与分词、编码和正字法差异相关的偏差。该研究认为，对语义等效序列进行句子级负对数似然计算，能产生更一致的比较结果。arxiv.org创新
MTDiag 测试医疗 LLM 是否能在多轮诊断对话中进行推理
一个新数据集将大型语言模型作为诊断代理，在交互式临床场景中而非通过孤立问题进行评估。MTDiag 结合了急诊医学、临床记录和已发表的病例报告中的案例，并提出了超越诊断准确性的基于知识的指标。arxiv.org创新
HealthBench-Psych 推出心理健康 AI 性能的开放基准
研究人员创建了 HealthBench-Psych，这是 OpenAI 的 HealthBench 中一个包含 610 段对话的心理健康子集，并发布了其评估流程、模型响应、评分和分析代码。arxiv.org创新
综述梳理了基础模型如何在训练后从自身输出中学习
一篇被 EMNLP 2026 Findings 接收的综述，列举了 80 种在无人工标签、偏好数据、更强教师模型或可执行验证器的情况下调整基础模型的方法。它警告说，内部学习信号既可能改进模型，也可能递归地放大其错误。arxiv.org创新
Study traces dialect bias through every stage of language-model development
A new preprint reports that language models encode dialectal English unequally during tokenization, pre-training, post-training and inference, even when paired texts have the same meaning.arxiv.org创新
Paper proposes adaptive timing for correcting bias in LLM reasoning
A new preprint proposes triggering bias corrections only when evidence of stereotyping accumulates during an LLM’s reasoning, reporting fewer interventions than fixed-interval methods but important accuracy tradeoffs across model types.arxiv.org创新
OmniPhys benchmark tests multimodal AI on physics reasoning and diagram generation
A paper accepted to Findings of EMNLP 2026 introduces OmniPhys, a benchmark of 15,246 physics questions and 19,850 images drawn from Chinese educational corpora. It evaluates multimodal AI systems on both physics reasoning and the generation of structured physics diagrams.arxiv.org创新
DCGC proposes global correction for flawed reasoning in masked diffusion language models
A new arXiv preprint introduces DCGC, a masked-diffusion framework that uses an upstream draft to revise reasoning globally instead of allowing early mistakes to propagate.arxiv.org创新
VietAIDetector offers an open-source detector for Vietnamese AI-generated text
A new arXiv paper introduces VietAIDetector, an open-source zero-shot tool designed to identify AI-generated Vietnamese text without domain-specific training data. The authors say it supports long and scanned documents and outperforms methods developed primarily for English on out-of-domain evaluations.arxiv.org