发现视觉语言模型会错答计数，但内部其实知道正确答案。
The Count Is There, but Misaligned: Understanding and Correcting Counting Failures in VLMs
- 用探测器分析模型激活值，发现计数错误时内部已有正确知识
- 改进探测方向后，计数准确率提升15.6个百分点
- 仅在检测到错误时重提示，无需训练就能纠错
尽管在多项多模态任务中表现优异，视觉语言模型（VLMs）仍难以完成基础物体计数。我们探究这是否源于内部知识缺失，还是内部表征与输出之间存在偏差。在四个VLM上，对五个计数数据集的激活值进行简单探测，结果表明非线性探测器能可靠识别计数错误，说明模型通常已编码正确数量，却在输出时出错。SVCCA分析显示，基于真实计数和模型输出训练的探测器共享部分激活子空间，但读取方向不一致。通过因果干预验证，强化计数探测方向可显著提升计数性能。受此启发，我们提出一种探测器引导的自校正方法：仅当内部探测器预测失败时，才对模型进行选择性重提示。该推理阶段干预方法无需参数更新，计数准确率最高提升15.6个百分点。结果表明，基于激活值的错误探测既是提升VLM计数能力的实用工具，也是揭示内部知识与输出间鸿沟的机制视角。
原文摘要 · Abstract (English)
Despite strong performance on many multimodal tasks, vision-language models (VLMs) still struggle with basic object counting. We investigate whether this reflects missing internal knowledge or a gap between internal representations and verbalized outputs. Training simple probes on activations from four VLMs across five counting datasets reveals that nonlinear probes can reliably detect counting errors, suggesting that VLMs often encode the correct count even when they output the wrong answer. SVCCA analysis shows that probes trained on ground-truth counts and probes trained on model outputs occupy a partially shared activation subspace but read out along misaligned directions. We further validate our findings using a causal steering intervention, proving that strengthening the direction of count-identified probes does improve model counting performance. Motivated by this result, we propose a detector-guided self-correction method that selectively re-prompts the model only when an internal error detector predicts failure. This simple inference-time intervention improves counting accuracy by up to 15.6 absolute percentage points, without any parameter updates. Our results establish activation-based error probing as both a practical tool for improving VLM counting and a mechanistic lens on the gap between internal knowledge and model outputs.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。