Knowing What Not to Answer: Selective Non-Compliance in Vision-Language Models

  • 类型：arxiv
  • 标识：2609.04720
  • 链接：https://arxiv.org/abs/2609.04720
  • 主分类：evaluation
  • 形态：benchmark
  • 被引：0
  • 被引来源：Semantic Scholar + OpenAlex
  • S2被引：0
  • OpenAlex被引：0
  • 影响力被引：0
  • TLDR：KoNA is introduced, a benchmark for evaluating selective non-compliance in VLMs across five categories: False Premise, Visual Inaccessibility, Universal Unknown, Task Feasibility, and Safety, and results suggest that the fine-tuned models can distinguish between answerable components and those requiring non-compliance and respond in a task-appropriate manner.
  • OpenAlex ID：W7211901234
  • OpenAlex DOI：10.48550/arxiv.2609.04720
  • DOI：10.48550/arxiv.2609.04720
  • DOI来源：OpenAlex
  • 开放获取：green
  • 开放获取链接：https://doi.org/10.48550/arxiv.2609.04720
  • OpenAlex更新：2026-09-09
  • 副分类：multimodal
  • 待LLM分类：否
  • 标题中文：知道不该回答什么：视觉语言模型中的选择性不遵从
  • TLDR中文：提出 KoNA，一个评估 VLM 选择性不遵从的基准，覆盖五类情形：False Premise、Visual Inaccessibility、Universal Unknown、Task Feasibility 与 Safety。实验结果显示，微调后的模型能够区分可回答部分与需要不遵从的部分，并以符合任务要求的方式作答。
  • 来源文件：
  • /inbox/tom/_candidates/2026-09-08-rag-retrieval-reranking-candidates.json
  • /inbox/tom/_candidates/2026-09-08-agent-rag-longcontext-candidates.json
  • [OpenAlex backfill]
  • [S2 enrich]