Primary sourcearXiv: DeepSeek
View sourceEvent ·Official first-hand
NumTrace 基准:审计金融分析师智能体的数字来源
arXiv 论文 NumTrace 针对金融分析师智能体生成的数字论断,审计其来源与财务语义。该基准包含 540 道题目,覆盖 2,163 份 SEC EDGAR 文件,首次覆盖 Form 13F 持股和 Item 402(i) 高管薪酬数据。基准对每个论断给出 value、binding、support 三项独立判定,用来识别“数字对但归属错”(RNWA)错误。审计 Claude Opus 5.5、GPT-6 Astra 和 DeepSeek-V4.1-Flash 三个智能体后发现,RNWA 错误占各自失败论断的 67.5%-81.8%,在 13F 联合申报组问题上约 91% 的论断通过数值检查,但通过 binding 检查的不足一半。
Current read
论文测了三个主流模型在 SEC 文件上算数字的表现,发现数字本身往往算对,但挂错了公司或科目,这个问题之前没人单独量化过。
11 sources31° AI heatLast updated 2026/10/26 05:58:10
Evidence
11 sourcesRelated source 1rohanpaul_ai
View sourceRelated source 2vLLM
View sourceRelated source 3TestingCatalog
View sourceRelated source 4The Rundown AI
View sourceRelated source 5IT之家
View sourceRelated source 6掘金本周最热
View sourceRelated source 7Decoder
View sourceRelated source 8OpenAI Blog
View sourceRelated source 9AIGCLINK
View sourceRelated source 10Simon Willison’s Weblog
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.