Event ·Official first-hand

GUI 智能体自报成功可能是假阳性,报告-轨迹审计可拦截

论文提出 report-trace auditing,用独立记录的执行轨迹核查 GUI 智能体的完成声明。aware/blind 对照实验显示自我报告会泄漏到判定:隐藏自报后,Qwen3.5-35B-A3B 的危险误报召回率提升 43.4 个百分点,DeepSeek-284B-A13B 提升 13.3 个百分点。结构化审计器将声明解释与无声明轨迹分析分离,在 465 条审计目标集上达到 F1 0.75-0.78。启用审计门控后,在 45-56% 的检查负载下可拦截 74-90% 的危险误报。

Current read

GUI 智能体嘴上说完成但轨迹不支持,这论文教你怎么用独立审计拦住这类假成功,还给了具体拦截率数据。

4 sources38° AI heatLast updated 2026/10/21 11:50:17

Evidence

4 sources
Primary sourcearXiv: DeepSeek
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.