🤖 AI

Do VLMs Read or Rewrite? On Transcription Faithfulness in Vision-Language Models

This paper introduces the FaithC4 benchmark to demonstrate that Vision-Language Models often rewrite imperfect text into plausible forms rather than faithfully transcribing it, revealing distinct degradation patterns across model types and identifying specific layer representations and word lengths that trigger this rewriting behavior.

Original authors: Gwang Gook Lee, Kenan Emir Ak, Jay Mohta, Yan Xu, Dimitrios Dimitriadis

Published 2026-07-27

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Gwang Gook Lee, Kenan Emir Ak, Jay Mohta, Yan Xu, Dimitrios Dimitriadis

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a robot to read a handwritten note from your grandmother. You expect the robot to act like a super-accurate photocopier, capturing every squiggle, smudge, and misspelled word exactly as it appears. This is the job of Optical Character Recognition (OCR), a technology that has existed for decades to turn images of text into digital words. But recently, a new kind of robot has arrived: the Vision-Language Model (VLM). Think of these as robots that don't just "see" the letters but also "think" about what the sentence should mean, combining a camera with a very smart brain trained on the entire internet.

The big question is: when these smart robots see a messy word, do they copy it faithfully, or do they try to "fix" it? It's a bit like the "typoglycemia" effect humans experience, where we can still read a sentence even if the middle letters of the words are scrambled, because our brains guess the intended word. While this is helpful for humans, it can be a disaster for a robot if you need a perfect, literal copy of a legal contract or a historical letter. If the robot decides to "correct" a typo that was actually part of the original document, it has rewritten history rather than recording it. This paper dives into whether these new AI robots are faithful scribes or over-eager editors.

Do VLMs Read or Rewrite?

In this study, researchers from Amazon set out to test a specific habit of these Vision-Language Models (VLMs): when they see a word that looks a little weird, do they read it exactly as it is, or do they secretly rewrite it into something that makes more sense?

To find out, the team created a special playground called FaithC4. Imagine they took thousands of clean, normal pages of text in English, Chinese, and Korean. Then, they played a game of "telephone" with the words before turning them into images. They took 8% of the eligible words and messed them up in three creative ways:

Scramble: They shuffled the inside letters (like turning "standard" into "sdanartd").
Random: They swapped letters for completely random ones (like "xkpmewqi").
Visual: They swapped letters for ones that look almost identical (like turning an 'a' into a 'c' or an 'o' into a '0').

They then showed these "broken" images to 15 different systems. These systems fell into three groups:

The Old Guard: Traditional OCR tools (like Tesseract) that just look at pixels and say "that looks like an 'a'."
The Specialists: VLMs trained specifically for reading documents.
The Generalists: The big, powerful VLMs (like Qwen, GPT-4V, and Gemini) that are good at everything from math to poetry.

The Results: The "Smart" Robots Are Too Smart

The findings were surprising and a bit worrying for anyone who needs perfect copies. The researchers found that the "smarter" the robot, the more likely it was to rewrite the text.

The Old Guard (Traditional OCR): These were the most faithful. When the text was messed up, they barely changed their answer. Their error rate went up by less than 0.6 percentage points. They just saw the mess and copied the mess.
The Specialists: These did a bit better than the generalists but still made some changes. Their errors went up by 0.2 to 2 percentage points.
The Generalists: These were the worst at being faithful. When the text was scrambled or visually altered, their error rates jumped by up to 4.5 percentage points.

Why? Because these general-purpose models are so good at predicting what a sentence should look like, they often ignore what is actually written. If they see "prbolem," they confidently output "problem," even if the image clearly said "prbolem." It's like a student who knows the answer to a math problem so well that they ignore the teacher's scribbled mistake on the board and just write the correct answer.

The Hidden Mechanism: When Does the Rewrite Happen?

To understand why this happens, the researchers looked inside the brain of one of these models, Qwen3-VL-4B, layer by layer. They wanted to see if the robot was "blind" to the error or if it just chose to ignore it.

They discovered a fascinating rule: The robot rewrites the word only when it still "thinks" the word looks familiar.

If the robot's internal representation of the scrambled word stays very close to the original word (like a blurry photo that still looks like a cat), the robot says, "Oh, that's definitely 'cat'!" and rewrites it.
But if the scramble is so bad that the internal representation drifts far away from the original (the photo is now just a blob), the robot stops guessing and finally just reads what it sees.

This explains why word length matters so much. Short words (4 to 6 letters) get rewritten up to 10% of the time because it's easy for the robot to guess the original. But once a word gets to 8 characters or longer, the rewriting drops to 0%. The scramble is just too big for the robot's "guessing brain" to recover the original, so it gives up and transcribes faithfully.

The Ripple Effect

Here is the most surprising part: messing up just a few words doesn't just hurt those specific words. The researchers found that corrupting just 5% of the words in a document caused errors on the perfect, untouched words to jump by 5 to 10 times for the general-purpose models.

It's as if the robot gets so confused by the one typo that it starts misreading the whole paragraph around it. Traditional OCR systems didn't have this problem; they stayed calm and accurate even when surrounded by chaos.

What This Means

The paper concludes that while these fancy new AI models are amazing at understanding documents and answering questions, they are not reliable for tasks that require a literal, perfect copy of the text. If you are digitizing a historical manuscript, checking a legal contract, or analyzing medical records where every typo matters, you should stick with traditional OCR or specialized tools. The general-purpose VLMs are too eager to "help" by fixing things that shouldn't be fixed.

The researchers suggest that we need to be careful about how we use these models. Just because a model can read a document doesn't mean it will read it faithfully. For now, if you need the truth exactly as it was written, the old-school robots are still the most honest scribes.

Technical Summary: Do VLMs Read or Rewrite? On Transcription Faithfulness in Vision-Language Models

Problem Statement

Vision-Language Models (VLMs) are increasingly replacing traditional Optical Character Recognition (OCR) pipelines for document understanding due to their ability to handle text recognition, layout analysis, and downstream tasks within a single model. However, standard OCR benchmarks, which rely on clean text, fail to capture a critical behavioral flaw: VLMs often act as "silent rewriters" rather than faithful transcribers. When presented with imperfect inputs (typos, visual artifacts, or unclear characters), VLMs tend to output what they expect to see based on language priors, rather than what is actually written. This behavior, reminiscent of the typoglycemia effect in human reading, poses significant risks for domains requiring literal transcription, such as legal documents, medical records, and historical manuscripts. Existing research has not systematically investigated this rewriting behavior, as current benchmarks focus on clean-text recognition.

Methodology

To address this gap, the authors introduce FaithC4, a multilingual perturbation benchmark designed to measure transcription faithfulness under imperfect inputs.

Dataset Construction: The authors sourced raw text from the Colossal Clean Crawled Corpus (C4) and rendered it into single-page document images. They applied controlled perturbations to 8% of eligible words (those with $\ge$ 4 characters) across three languages: English (EN), Chinese (ZH), and Korean (KO). The perturbation families include:
- Scramble: Internal characters are shuffled while preserving the first and last letters (e.g., "standard" $\to$ "sdanartd").
- Random Substitution: Characters are replaced with random letters from the same script.
- Visual Substitution: Characters are replaced with visually similar alternatives (e.g., "standard" $\to$ "stsndord").
- The perturbed text serves as the ground truth; the model's output is compared against this to determine if it faithfully transcribed the error or "rewrote" it back to the original.
Systems Evaluated: The study evaluates 15 systems across three categories:
1. General-purpose VLMs: Open-weight (e.g., Qwen3-VL, InternVL, Gemma) and closed-weight (e.g., GPT-5.4-mini, Gemini) models.
2. OCR-specialized VLMs: Models fine-tuned for document extraction (e.g., MinerU, DeepSeek-OCR).
3. Traditional OCR: Character-level pipelines (Tesseract, docTR).
Evaluation Metrics:
- Word Error Rate (WER): Measures order-sensitive word-level accuracy (or Character Error Rate, CER, for Chinese).
- Edit Distance Similarity (EDS): Measures character-level similarity.
- Error Classification: Errors on perturbed words are categorized as Rewritten (matching the original unperturbed word), Close (minor edits), Moderate, or Wrong.
Mechanistic Probing: The authors perform layer-by-layer probing on Qwen3-VL-4B to identify the origin of rewriting. They analyze the cosine similarity of Feed-Forward Network (FFN) outputs between original and perturbed inputs and measure attention weights on highlighted word patches.

Key Results

1. Performance Degradation and Robustness Hierarchy

The study reveals a consistent three-tier ordering of robustness across all three languages, correlating with the strength of a model's language prior:

Traditional OCR: Most robust, with degradation $< 0.6$ percentage points (pp) in WER/CER. They transcribe character-by-character without overriding visual evidence.
OCR-specialized VLMs: Show moderate degradation (0.2–2.0 pp), retaining some language priors but less than general models.
General-purpose VLMs: Suffer the largest degradation (up to 4.5 pp). For example, InternVL3.5-4B degraded by 4.45 pp under visual perturbation in English, while Gemini-3-Flash was an exception with lower degradation (0.23–0.60 pp).

2. Rewriting Behavior

Rewriting Rates: General-purpose VLMs rewrite a substantial fraction of perturbed words. For instance, under scramble perturbation, Gemini-2.5-Flash rewrote 64.6% of errors, while Qwen3-VL rewrote 29.2%. Traditional OCR rewrote 0%.
Non-Local Error Propagation: Perturbing only ~4.7% of words caused error rates on the unperturbed remainder of the text to increase by 5–10 $\times$ for general-purpose VLMs, whereas traditional OCR showed only 1.7–3.1 $\times$ amplification. This suggests that rewriting is a non-local failure mode.
Cross-Lingual Consistency: The degradation hierarchy holds for English, Chinese, and Korean, though the most damaging perturbation type varies by script (e.g., random substitution is most damaging for Chinese general-purpose VLMs).

3. Mechanistic Insights (Qwen3-VL-4B)

Probing the internal layers of Qwen3-VL-4B revealed that rewriting is not caused by a failure to notice the perturbation:

Attention vs. Representation: The model allocates more attention to the perturbed word's image patches at early and late layers, indicating it "sees" the corruption.
The Gating Mechanism: Rewriting occurs only when the perturbed word's internal FFN representation at the final layer remains close to the original encoding.
- If the representation diverges sufficiently (low cosine similarity), the model transcribes faithfully.
- If the representation stays close (high similarity), the language prior overrides the visual evidence to "correct" the word.
Word Length Effect: Rewriting is highly dependent on word length. Short words (4–6 characters) are rewritten up to 10% of the time. There is a sharp cutoff at 8 characters, where the rewriting rate drops to 0%. Shorter words cause less disruption to the internal representation, allowing the language prior to recognize and "correct" them.

Significance and Claims

The paper claims that FaithC4 exposes a systematic limitation in VLMs that clean-text benchmarks cannot detect: the tendency to silently rewrite imperfect text based on language priors. The authors argue that for applications requiring literal transcription (legal, medical, historical), traditional OCR or carefully selected OCR-specialized VLMs remain the safest choice, as general-purpose VLMs introduce systematic rewriting errors.

The study establishes that the "rewriting" behavior is not a hallucination in the traditional sense of generating non-existent objects, but a specific failure mode where the model's language prior overrides visual evidence when the perturbation is subtle enough to leave the internal representation recognizable. This behavior is gated by the magnitude of the shift in the model's internal representation, not by the amount of attention allocated to the input.