Primary sourcearXiv: OpenAI
View sourceEvent ·Official first-hand
论文提出 DS-QA 基准与蒙特卡洛协议评估 LLM 数据科学任务可靠性
论文介绍 DS-QA 基准,包含 100 道来自英国大学课程的数据科学与统计题目,覆盖 16 个数据集和 3 个难度等级。评估框架分三层:通用响应属性(运行时间、冗长度、文本相似度)、粗粒度指标(准确率、代码可执行性、完整性)和任务特定指标。团队用蒙特卡洛协议对 OpenAI 模型重复运行 100 次,发现数据可视化与监督学习任务在中高难度下响应变异性显著。论文还测得词冗长度比每增加 0.1,token 冗长度比约增加 0.013。
Current read
这篇论文做了件常被忽略的事:同一题让模型跑 100 遍看答案稳不稳,结果数据可视化和监督学习这类任务波动最大,做评测的人该看看。
11 sources31° AI heatLast updated 2026/11/6 17:06:21
Evidence
11 sourcesRelated source 1Stanford AI Lab
View sourceRelated source 2IT之家
View sourceRelated source 3TestingCatalog
View sourceRelated source 4宝玉
View sourceRelated source 5Greg Brockman
View sourceRelated source 6swyx (AI Engineer)
View sourceRelated source 7rohanpaul_ai
View sourceRelated source 8kimmonismus
View sourceRelated source 9DAIR.AI
View sourceRelated source 10Decoder
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.