统一多模态模型视觉拒答能力研究
Get the story
2026年10月7日,arXiv Computation and Language 发布一手研究,提出视觉拒答(visual abstention)任务,要求统一多模态模型在请求的视觉变换违反任务规则时识别无解并拒绝生成。论文发布 Draw-or-Decline(DoD)基准,含 1,050 对可行/不可行请求,覆盖 7 类任务;评测 8 个统一多模态模型,结果显示最强编辑器编辑准确率为 68.4%,但仅拒绝 0.4% 的不可行请求,暴露其视觉拒答能力不足。
Generated from reports · updated 3 hr ago
Timeline
Follow the coverage from different angles.
- arXiv · Computation and Language统一多模态模型的视觉拒答能力研究
研究提出视觉拒答(visual abstention)任务,要求统一多模态模型在请求的视觉变换违反任务规则时识别无解并拒绝生成。论文发布 Draw-or-Decline(DoD)基准,含 1,050 对可行/不可行请求、覆盖 7 类任务;评测 8 个 UMM 显示,最强编辑器编辑准确率 68.4%,却仅拒绝 0.4% 的不可行请求。
Heat trend
Current heat 9·Comparable peak 10(Oct 7)·Comparable change over 24 hours –
The trend compares only the same participants observed continuously; its range may be smaller than the current heat count. Move or click on the chart to inspect hourly heat; use the left and right arrow keys to switch.