Reading, Not Thinking: Understanding and Bridging the Modality Gap When Text Becomes Pixels in Multimodal LLMs

• URL: http://arxiv.org/abs/2603.09095v1
• Date: Tue, 10 Mar 2026 02:14:23 GMT
• Title: Reading, Not Thinking: Understanding and Bridging the Modality Gap When Text Becomes Pixels in Multimodal LLMs
• Abstract summary: Multimodal large language models (MLLMs) can process text presented as images, yet they often perform worse than when the same content is provided as textual tokens.<n>We systematically diagnose this "modality gap" by evaluating seven MLLMs across seven benchmarks in five input modes.
• Score: 23.958966900531692
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Multimodal large language models (MLLMs) can process text presented as images, yet they often perform worse than when the same content is provided as textual tokens. We systematically diagnose this "modality gap" by evaluating seven MLLMs across seven benchmarks in five input modes, spanning both synthetically rendered text and realistic document images from arXiv PDFs to Wikipedia pages. We find that the modality gap is task- and data-dependent. For example, math tasks degrade by over 60 points on synthetic renderings, while natural document images often match or exceed text-mode performance. Rendering choices such as font and resolution are strong confounds, with font alone swinging accuracy by up to 47 percentage points. To understand this, we conduct a grounded-theory error analysis of over 4,000 examples, revealing that image mode selectively amplifies reading errors (calculation and formatting failures) while leaving knowledge and reasoning errors largely unchanged, and that some models exhibit a chain-of-thought reasoning collapse under visual input. Motivated by these findings, we propose a self-distillation method that trains the model on its own pure text reasoning traces paired with image inputs, raising image-mode accuracy on GSM8K from 30.71% to 92.72% and transferring to unseen benchmarks without catastrophic forgetting. Overall, our study provides a systematic understanding of the modality gap and suggests a practical path toward improving visual text understanding in multimodal language models.

Related papers

• When Prompts Become Pixels: Prompt-Region Grounding for Multimodal Reasoning [73.56343820592399]
  We introduce Visualized Task Semantics (VTS), a controlled intervention that moves the question into the image while keeping the source problem and answer fixed.<n>Across six MLLMs and four benchmarks, accuracy drops in all 24 model-task pairs, by 17.8 points on average.<n>We present prompt-region grounding, whose core design aligns the question region with typed semantics and recovers its clean representation from a masked view.
  arXiv  Detail & Related papers  (2026-08-05T11:46:15Z)
• Do Vision-Language Models See or Guess? Measuring and Reducing Textual-Prior Reliance with a Phrasing-Controlled Benchmark [2.379911867541422]
  Vision-language models (VLMs) are increasingly deployed where answers must follow from what is in the image.<n>They often answer from textual priors, the question's phrasing together with memorized world knowledge, rather than from the image itself.<n>Existing benchmarks rarely isolate this behavior, since each image is usually paired with a single fixed question.
  arXiv  Detail & Related papers  (2026-06-09T04:18:38Z)
• LogicOCR: Do Your Large Multimodal Models Excel at Logical Reasoning on Text-Rich Images? [80.4577892387028]
  We introduce LogicOCR, a benchmark comprising 1,100 multiple-choice questions designed to evaluate LMMs' logical reasoning abilities on text-rich images.<n>We develop a scalable, automated pipeline to convert a text corpus into multimodal samples.<n>We evaluate a range of representative open-source and proprietary LMMs under both Chain-of-Thought (CoT) and direct-answer settings.
  arXiv  Detail & Related papers  (2025-05-18T08:39:37Z)
• Towards Visual Text Grounding of Multimodal Large Language Model [74.22413337117617]
  We introduce TRIG, a novel task with a newly designed instruction dataset for benchmarking text-rich image grounding.<n>Specifically, we propose an OCR-LLM-human interaction pipeline to create 800 manually annotated question-answer pairs as a benchmark.<n>A comprehensive evaluation of various MLLMs on our proposed benchmark exposes substantial limitations in their grounding capability on text-rich images.
  arXiv  Detail & Related papers  (2025-04-07T12:01:59Z)
• Towards Text-Image Interleaved Retrieval [49.96332254241075]
  We introduce the text-image interleaved retrieval (TIIR) task, where the query and document are interleaved text-image sequences.<n>We construct a TIIR benchmark based on naturally interleaved wikiHow tutorials, where a specific pipeline is designed to generate interleaved queries.<n>We propose a novel Matryoshka Multimodal Embedder (MME), which compresses the number of visual tokens at different granularity.
  arXiv  Detail & Related papers  (2025-02-18T12:00:47Z)
• ViLBias: Detecting and Reasoning about Bias in Multimodal Content [6.710013984078675]
  ViLBias is a framework for detecting and reasoning about bias in multimodal news.<n> dataset comprises 40,945 text--image pairs from diverse outlets.<n>Results show that images alongside text improves detection accuracy by 3--5%.
  arXiv  Detail & Related papers  (2024-12-22T15:05:30Z)
• ASCIIEval: Benchmarking Models' Visual Perception in Text Strings via ASCII Art [83.95594027644124]
  We frame the problem as a recognition task, and construct a novel benchmark, ASCIIEval.<n>It covers over 3K samples with an elaborate categorization tree, along with a training set for further enhancement.<n>Given textual input, language models shows their visual perception ability on ASCII art concepts.<n>For image inputs, we reveal that open-source MLLMs suffer from a trade-off between fine-grained text recognition and collective visual perception.
  arXiv  Detail & Related papers  (2024-10-02T16:46:01Z)
• AdaptVision: Dynamic Input Scaling in MLLMs for Versatile Scene Understanding [96.01726275876548]
  We present AdaptVision, a multimodal large language model specifically designed to dynamically process input images at varying resolutions. We devise a dynamic image partitioning module that adjusts the number of visual tokens according to the size and aspect ratio of images. Our model is capable of processing images with resolutions up to $1008times 1008$.
  arXiv  Detail & Related papers  (2024-08-30T03:16:49Z)
• A Picture is Worth a Thousand Words: Principled Recaptioning Improves Image Generation [9.552642210681489]
  We show that by relabeling the corpus with a specialized automatic captioning model and training a text-to-image model on the recaptioned dataset, the model benefits substantially across the board. We analyze various ways to relabel the corpus and provide evidence that this technique, which we call RECAP, both reduces the train-inference discrepancy and provides the model with more information per example.
  arXiv  Detail & Related papers  (2023-10-25T14:10:08Z)
• Dynamic Visual Semantic Sub-Embeddings and Fast Re-Ranking [0.5242869847419834]
  We propose a Dynamic Visual Semantic Sub-Embeddings framework (DVSE) to reduce the information entropy. To encourage the generated candidate embeddings to capture various semantic variations, we construct a mixed distribution. We compare the performance with existing set-based method using four image feature encoders and two text feature encoders on three benchmark datasets.
  arXiv  Detail & Related papers  (2023-09-15T04:39:11Z)
• Is it an i or an l: Test-time Adaptation of Text Line Recognition Models [9.149602257966917]
  We introduce the problem of adapting text line recognition models during test time. We propose an iterative self-training approach that uses feedback from the language model to update the optical model. Experimental results show that the proposed adaptation method offers an absolute improvement of up to 8% in character error rate.
  arXiv  Detail & Related papers  (2023-08-29T05:44:00Z)
• RoCOCO: Robustness Benchmark of MS-COCO to Stress-test Image-Text Matching Models [39.37985751472905]
  We create new variants of texts and images in the MS-COCO test set and re-evaluate the state-of-the-art (SOTA) models with the new data.<n>Specifically, we alter the meaning of text by replacing a word, and generate visually altered images that maintain some visual context.<n>Our evaluations on the proposed benchmark reveal substantial performance degradation in many SOTA models.
  arXiv  Detail & Related papers  (2023-04-21T03:45:59Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.