Primary sourcearXiv cs.AI
View sourceEvent ·Single source
视频语言模型在简单事件计数上失准，研究揭示时间推理瓶颈
新论文提出trace-grounded参数化测试方法，在2,190个视频上审计事件计数能力。Gemini 3.6 Flash在80%可靠性阈值下最多可靠计数12个持续状态转换事件（0.5和1.0 Hz），对瞬时闪烁事件则无可靠计数区间。高计数高频率场景下仅0.2%的最终计数正确，模型只能恢复18.1%的真实事件。提高采样率将Bounce Ball准确率从19.6%提升至29.3%，但报告序列与真值一致率仅3.7%。
Current read
用可执行真值逐帧审计视频模型计数，发现Gemini 3.6 Flash处理瞬时事件几乎失效，加帧只虚涨分数，值得看。
2 sources53° AI heatLast updated 2026/8/6 17:57:06
Evidence
2 sourcesRelated source 1向阳乔木
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.