Do VLMs Read or Rewrite? On Transcription Faithfulness in Vision-Language Models VLMs 是阅读还是重写？关于视觉语言模型的转录保真度

TL;DR

• Vision Language Models (VLMs) often rewrite imperfect text into more plausible forms rather than transcribing it faithfully, a behavior not detected by clean-text OCR benchmarks.
• The paper introduces FaithC4, a multilingual perturbation benchmark with 1,455 single-page documents (English, Chinese, Korean) featuring three perturbation families: scramble, random substitution, and visually similar substitution.
• Evaluation of 15 systems reveals significant differences in Word Error Rate (WER) degradation under perturbation: general-purpose VLMs degrade by up to 4.5 points, OCR-specialized VLMs by 0.2-2 points, and traditional OCR by less than 0.6 points on English.
• Layer-by-layer probing of Qwen3-VL-4B identifies that rewriting occurs when a perturbed word's final layer FFN representation stays close to the original encoding; sufficient divergence leads to faithful transcription.
• Short words (4-6 characters) are rewritten up to 10% of the time, with a sharp cutoff at 8 characters above which rewriting drops to 0%.

Why It Matters

This research is crucial for AI practitioners and researchers working on document understanding, as it highlights a critical limitation of VLMs: their tendency to "hallucinate" or rewrite text when faced with imperfect input, which can lead to significant errors in real-world applications where text is often noisy or degraded. The FaithC4 benchmark provides a much-needed tool for evaluating transcription faithfulness, enabling better model selection and development for tasks requiring accurate text extraction.

Technical Details

• FaithC4 benchmark comprises 1,455 single-page documents in English, Chinese, and Korean, perturbed using three methods: scramble (reordering characters), random substitution (replacing characters with random ones), and visually similar substitution (replacing characters with visually similar ones).
• The study evaluates 15 systems: general-purpose VLMs (e.g., GPT-4V, Claude 3), OCR-specialized VLMs (e.g., PaddleOCR, DocVLM), and traditional OCR pipelines (e.g., Tesseract).
• Analysis of Qwen3-VL-4B reveals that rewriting behavior is tied to the similarity between the perturbed word's final layer FFN representation and the original encoding; when representations diverge beyond a threshold, the model transcribes faithfully.
• Word length is a key factor: short words (4-6 characters) are rewritten up to 10% of the time, while words longer than 8 characters are never rewritten.

Industry Insight

• Organizations relying on VLMs for document processing should be aware of the risk of unintended text rewriting, especially for short words, and consider using OCR-specialized models or traditional OCR pipelines for critical applications requiring high transcription fidelity.
• The FaithC4 benchmark should be adopted as a standard evaluation metric for VLMs in document understanding tasks, complementing existing clean-text benchmarks to better assess real-world performance.
• Future VLM development should focus on reducing rewriting behavior for imperfect text, potentially through architectural modifications or training strategies that prioritize transcription faithfulness over plausibility.

TL;DR

• 视觉语言模型（VLMs）在文档理解中常被用作OCR的替代品，但它们在文本不完美时倾向于重写而非忠实转录。
• 引入了FaithC4基准，包含1,455个单页文档（英语、中文、韩语），具有三种扰动家族：打乱、随机替换和视觉相似替换。
• 评估了15个系统，包括通用VLMs、OCR专用VLMs和传统OCR管道，发现它们在扰动下的WER退化程度不同。
• 对Qwen3-VL-4B的分层探测发现，重写行为仅在扰动词的最终层FFN表示接近原始编码时触发；当表示充分发散时，模型会忠实转录。
• 单词长度影响重写率：短词（4-6字符）的重写率最高可达10%，超过8字符时重写率降至0%。

为什么值得看

• 该研究揭示了VLMs在文档理解中的潜在局限性，特别是在处理不完美文本时的行为，这对于依赖VLMs进行文档处理的应用至关重要。
• FaithC4基准为评估VLMs的转录忠实性提供了一个新的工具，有助于推动更可靠和透明的文档理解系统的发展。

技术解析

• FaithC4基准：包含1,455个单页文档，覆盖英语、中文和韩语，具有三种扰动家族：打乱、随机替换和视觉相似替换，用于评估VLMs的转录忠实性。
• 系统评估：评估了15个系统，包括通用VLMs、OCR专用VLMs和传统OCR管道，发现它们在扰动下的WER退化程度不同，通用VLMs退化最严重，传统OCR最稳定。
• Qwen3-VL-4B分析：通过分层探测发现，重写行为仅在扰动词的最终层FFN表示接近原始编码时触发，表示充分发散时模型会忠实转录。
• 单词长度影响：短词（4-6字符）的重写率最高可达10%，超过8字符时重写率降至0%，表明模型对短词更倾向于重写。

行业启示

• VLMs的局限性：VLMs在处理不完美文本时可能不如传统OCR可靠，行业在选择文档理解工具时需考虑这一局限性。
• 基准的重要性：FaithC4基准为评估VLMs的转录忠实性提供了新工具，有助于推动更可靠和透明的文档理解系统的发展。
• 模型优化方向：研究结果提示，优化VLMs在处理短词时的转录忠实性可能是未来改进的方向，以提升其在文档理解中的可靠性。