跳转至

文章背景与核心概要

本文探讨了视觉语言模型（VLM）是否能够可靠地预测自身行为并遵循其声明的内省推理，这对于构建值得信赖的AI部署至关重要。作者引入了分级颜色归因（Graded Color Attribution, GCA）基准测试，这是一个通过控制像素级颜色覆盖率来考察世界知识重着色、反事实重着色及无颜色先验形状的受控数据集。

通过对比人类参与者和VLM在阈值规则生成以及后续颜色归因决策上的表现，研究揭示了一个显著的差异：人类能够忠实于自己阐述的规则（微小偏差可用对颜色覆盖率的高估来解释），而VLM则系统性地违背了自身的内省规则。例如，在具有强颜色先验的物体上，GPT-5-mini在近60%的情况下会与其自身阐述的规则相矛盾。最终，研究结果表明VLM的内省自我知识存在校准偏差，这对“VLM推理失败纯粹由任务难度驱动”的传统假设提出了挑战。

何时将苹果称为红色：人类遵循内省规则，而视觉语言模型（VLM）则不然 (When to Call an Apple Red: Humans Follow Introspective Rules, VLMs Don't)

摘要 (Summary)

This paper investigates whether Vision-Language Models (VLMs) can reliably predict their own behavior and adhere to their stated introspective reasoning—critical components for trustworthy AI deployment. The authors introduce the Graded Color Attribution (GCA) benchmark, a controlled dataset consisting of line drawings that vary pixel-level color coverage across world-knowledge recolorings, counterfactual recolorings, and shapes without color priors.

By comparing human participants and VLMs on threshold-rule generation and subsequent color attribution decisions, the study reveals a striking discrepancy: while humans remain faithful to their stated rules (with minor deviations explained by overestimating color coverage), VLMs systematically violate their own introspective rules. For instance, GPT-5-mini contradicts its own stated rules in nearly 60% of cases for objects with strong color priors. Ultimately, the findings suggest that VLM introspective self-knowledge is miscalibrated, challenging the assumption that VLM reasoning failures are purely difficulty-driven.

本文探讨了视觉语言模型（VLM）是否能够可靠地预测自身行为并遵循其声明的内省推理——这是实现值得信赖的AI部署的关键组成部分。作者引入了分级颜色归因（Graded Color Attribution, GCA）基准测试，这是一个受控数据集，包含通过像素级颜色覆盖率变化而生成的线条图，涵盖了世界知识重着色、反事实重着色以及无颜色先验的形状。

通过对比人类参与者和VLM在阈值规则生成及后续颜色归因决策上的表现，该研究揭示了一个显著的差异：人类能够忠实于自己阐述的规则（微小偏差可以通过对颜色覆盖率的高估来解释），而VLM则系统性地违背了自身的内省规则。例如，对于具有强颜色先验的物体，GPT-5-mini在近60%的情况下会与其自身阐述的规则相矛盾。最终，研究结果表明，VLM的内省自我知识存在校准偏差，这挑战了“VLM推理失败纯粹由难度驱动”的假设。

元数据 (Metadata)

  • arXiv Identifier: arXiv:2604.06422 [cs.CL]
  • Authors: Jonathan Nemitz, Carsten Eickhoff, Junyi Jessy Li, Kyle Mahowald, Michal Golovanevsky, William Rudman
  • Submitted: April 7, 2026 (v1); Last revised: August 19, 2026 (v2)
  • Accepted at: COLM 2026
  • Primary Subject: Computation and Language (cs.CL)
  • Secondary Subjects: Artificial Intelligence (cs.AI), Computer Vision and Pattern Recognition (cs.CV)
  • DOI: 10.48550/arXiv.2604.06422
  • arXiv 标识符： arXiv:2604.06422 [cs.CL]
  • 作者： Jonathan Nemitz, Carsten Eickhoff, Junyi Jessy Li, Kyle Mahowald, Michal Golovanevsky, William Rudman
  • 提交时间： 2026年4月7日（v1）；最后修订：2026年8月19日（v2）
  • 录用会议： COLM 2026
  • 主要学科： 计算与语言 (cs.CL)
  • 次要学科： 人工智能 (cs.AI)、计算机视觉与模式识别 (cs.CV)
  • DOI： 10.48550/arXiv.2604.06422

摘要正文 (Abstract)

Understanding when Vision-Language Models (VLMs) will behave unexpectedly, whether models can reliably predict their own behavior, and if models adhere to their introspective reasoning are central challenges for trustworthy deployment. To study this, we introduce the Graded Color Attribution (GCA) dataset, a controlled benchmark designed to elicit decision rules and evaluate participant faithfulness to these rules. GCA consists of line drawings that vary pixel-level color coverage across three conditions: world-knowledge recolorings, counterfactual recolorings, and shapes with no color priors. Using GCA, we ask both VLMs and human participants to state a threshold rule: the share of an object's pixels that must be a given color for the object to receive that color label. We then compare these rules with their subsequent color attribution decisions.

Our findings reveal that models systematically violate their own introspective rules. For example, GPT-5-mini violates its stated introspection rules in nearly 60% of cases on objects with strong color priors. Human participants remain faithful to their stated rules, with any apparent violations being explained by a well-documented tendency to overestimate color coverage. In contrast, we find that VLMs can accurately estimate color coverage, yet directly contradict their own reasoning in their final responses. Across all models and strategies for eliciting introspective rules, world-knowledge priors systematically degrade faithfulness in ways that do not mirror human cognition. Our findings challenge the view that VLM reasoning failures are difficulty-driven and suggest that VLM introspective self-knowledge is miscalibrated, with direct implications for high-stakes deployment.

理解视觉语言模型（VLM）何时会表现出意料之外的行为、模型是否能可靠地预测自身行为、以及模型是否遵循其内省推理，是实现可信部署的核心挑战。为了研究这些问题，我们引入了分级颜色归因（GCA）数据集，这是一个受控基准，旨在引发决策规则并评估参与者对这些规则的忠实度。GCA由线条图组成，这些线条图在三种条件下改变像素级颜色覆盖率：世界知识重着色、反事实重着色以及无颜色先验的形状。利用GCA，我们要求VLM和人类参与者阐述一项阈值规则：物体必须有多少比例的像素呈现特定颜色，该物体才会被赋予该颜色标签。随后，我们比较了这些规则与其后续的颜色归因决策。

我们的研究结果表明，模型会系统性地违背自身的内省规则。例如，在具有强颜色先验的物体上，GPT-5-mini在近60%的情况下违背了其声明的内省规则。人类参与者则对他们阐述的规则保持忠实，任何表面的违背都可以通过人们众所周知的高估颜色覆盖率的倾向来解释。相比之下，我们发现VLM能够准确估计颜色覆盖率，却在最终的回答中直接与其自身的推理相矛盾。在所有模型和激发内省规则的策略中，世界知识先验以一种不反映人类认知的方式系统性地降低了忠实度。我们的研究结果挑战了“VLM推理失败是由难度驱动的”这一观点，并表明VLM的内省自我知识存在校准偏差，这对高风险部署具有直接影响。

访问与资源 (Access and Resources)