The paper investigates the disparity between a vision‑language model’s ability to recognize text in compressed images and its ability to use that text for answering questions, a phenomenon termed the read‑use gap. Experiments show that while models can transcribe images accurately, they often fail to answer based on the transcription, and that recognition saturates at lower pixel resolutions than usage. The authors propose ALOUD, a method that allocates visual resolution to evidence detection and language tokens to reasoning, which doubles the accuracy of a frozen Qwen3‑VL‑8B on heavily compressed documents without retaining the original text.
Multimodal LLMs can see a document, but they often can't read it reliably. Small text, tables, visual cues, and topological elements still trip them up under direct visual inference, even when the pag...
The paper investigates the gap between retrieval and reading in document vision‑language models, showing that even when the correct page is retrieved, the model often fails to use the text on that page. By comparing answers derived from page images alone versus images plus extracted OCR text, the authors demonstrate that adding OCR text can improve strict accuracy by 13 to 16 points on their FoveDoc‑Bench benchmark. The study also reveals that OCR benefits textual evidence but not charts or figures, and that its advantage diminishes as retrieval quality worsens.
The paper investigates how general vision‑language models (VLMs) develop specialized optical character recognition (OCR) capabilities. By applying a causal intervention protocol, the authors identify sparse, stable OCR‑head sets in several VLMs and show that these heads largely overlap with textual retrieval/copy heads found in general VLMs. The study concludes that full‑sequence OCR functions as a dense multimodal copy‑and‑paste mechanism, and that when a VLM is fine‑tuned for OCR, it largely preserves the same head identities while redistributing their functional and causal strengths.
Vision‑Language Models (VLMs) are increasingly replacing traditional OCR for document understanding, but this study shows they often rewrite imperfect text into more plausible forms, a flaw that clean‑text OCR benchmarks miss. The authors created FaithC4, a multilingual perturbation benchmark of 1,455 single‑page documents with scramble, random substitution, and visually similar substitution attacks, and evaluated 15 systems across general‑purpose VLMs, OCR‑specialized VLMs, and traditional OCR pipelines. Results reveal that general‑purpose VLMs suffer up to 6.9 WER points under perturbation, OCR‑specialized VLMs 0.1–3.4 points, and traditional OCR less than 0.8 points on English; probing Qwen3‑VL‑4B shows rewriting occurs only when a perturbed word’s final‑layer representation remains close to the original, with short words (4–6 characters) rewritten up to 10% of the time. whyItMatters":"The findings highlight a critical limitation of VLMs in document transcription, underscoring the need for robust evaluation benchmarks that capture rewriting behavior beyond clean‑text accuracy."
arXiv:2609.38368v1 Announce Type: new Abstract: Vision-language models (VLMs) increasingly reason over visual evidence that is cropped, segmented, retrieved, or revealed over time. Yet most VQA bench...
LatentPress compresses conversational histories and long documents into continuous memory tokens that a frozen decoder can read directly, eliminating the need for text reconstruction at inference. The method achieves 4–16× compression with only a small adapter (0.1% of the decoder’s parameters) and outperforms text summaries and OCR-based compression on LongMemEval and LongBench-QA benchmarks. Writing and reading are significantly faster than traditional text summarization or OCR reconstruction, demonstrating a practical machine-facing context interface beyond text and vision.
arXiv:2610.29241v1 Announce Type: new Abstract: In this paper, we investigate why vision-language models (VLMs) produce errors by decomposing them into two sources: visual-representation failures, in...
arXiv:2608. 19739v1 Announce Type: cross Abstract: Multimodal LLMs can see a document, but they often can't read it reliably.
LensVLM is an inference framework and post‑training recipe that lets Vision‑Language Models (VLMs) process compressed images of text by selectively expanding only the relevant parts back to full resolution. Using Qwen3.5‑9B‑Base, LensVLM achieves accuracy comparable to full‑text models at 4.3× compression and outperforms other compression baselines up to 10.1× across seven text QA benchmarks, while also improving performance on multimodal document and code tasks as compression increases.
The paper introduces the concept of Squint Text—text embedded in an image’s coarse structure that multimodal LLMs often miss but humans can read by squinting. It presents SquintBench, a dataset of 1,076 moderation images, and SquintCtrl, a synthetic suite for diagnosing the issue. The authors show that spatial selection is a key bottleneck and propose Squint OPSD, a training method that uses a downscaled teacher view and privileged transcription to supervise a student model, achieving a significant performance boost on SquintBench.
The paper introduces Visual Retrieval Heads (VRHs), a small fraction of attention heads in vision‑language models that are causally responsible for grounding text descriptions to image regions. By recasting head‑scoring methods and evaluating across eleven VLMs and five benchmarks, the authors show that masking the top 20 VRHs can drop grounding accuracy by up to 80 percentage points, while random masking has little effect. VRHs generalize across various visual reference tasks, preserve output format while corrupting localization, and transfer causally across models sharing an LLM backbone.
The paper reports that scene text recognition models, while achieving 89–97% accuracy on standard benchmarks, perform significantly worse on rare word–trigram combinations, with a 10–18 point drop in accuracy at the rare‑word/rare‑trigram corner across multiple languages and models. Scaling the vision backbone improves overall accuracy but does not alleviate this corner‑specific deficit. The authors identify the autoregressive decoder’s lexical prior as the root cause and show that architectural changes—specifically moving from autoregressive to CTC decoding—yield the largest improvement for these rare compositions.