Event ·Multi-source

GPT-6 Astra 仅靠计算工具即可审计临床预测网络,无需直接读取权重

一项仿真实验测试了 GPT-6 Astra 等 5 个 GPT 模型能否从文件层面审计临床预测模型。在无工具条件下,Astra 只能处理 353 权重网络,超过 3,457 权重即不可靠,且默认指令下拒绝 12 项任务中的 11 项。接入仅暴露网络预测的黑盒计算工具后,它在全部 29 次、最高 38,401 权重的实验中找出了植入关系,并在 6 次实验中发现标签少报死亡的偏置亚组。黑盒与白盒访问效果相当,说明审计关键在于计算能力与提示设计,而非读取原始权重。

Current read

实验设计挺巧的:给 GPT-6 一个只能算数不能读权重的工具,审计临床模型的效果和白盒一样好,对做模型审计的人很有参考价值。

11 sources37° AI heatLast updated 2026/11/5 11:28:04

Evidence

11 sources
Related source 1shao__meng
View source
Related source 2Simon Willison
View source
Related source 3Artificial Analysis
View source
Related source 4@koltregaskes
View source
Related source 5lmarena.ai
View source
Related source 6Gorden Sun
View source
Related source 8智源研究院 BAAI
View source
Related source 9AI Engineer
View source
Related source 10rohanpaul_ai
View source

Conflict 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.