An AI that researches papers as page images instead of OCR text ── DR Pixel, a preprint, retrieves and reasons over rendered pages without converting them to text

1. What problem does the paper address?

The problem is that AI systems that research papers read them only after turning them into plain text. Deep research agents, which search many documents, read them and combine what they find into an answer, usually convert papers to text with OCR (optical character recognition) and then search and reason over that text. But much of a paper's information lives in tables, figures and equations. Once a page becomes a string of text, the alignment of rows and columns in a table, or the axes and legend of a figure, can be scrambled or lost^[1].

The authors reframe the question. If vision-language models (VLMs) are capable enough, why not skip the conversion, retrieve pages as images and reason over the images directly? They call this page-native deep research^[1].

Retrieving pages as images is not itself a new idea; ColPali and its benchmark ViDoRe had already demonstrated it^[3]. The paper's question goes a step further: can the whole job of an agent, which moves back and forth between searching, reading and reasoning many times, run on page images alone? This article covers a preprint posted to arXiv that has not been peer reviewed, so what follows should be read as the authors' claims.

2. What does it propose?

The core idea is to make the rendered page image the single unit of both retrieval and reasoning. DR Pixel is an open-source agent that uses page images for both. The authors describe it as the first open-source, end-to-end multimodal deep research agent built this way. It operates over a corpus of 1.06M pages^[1].

To evaluate this setting, the authors built DRPixelBench, a benchmark of questions about scientific papers that require reasoning across text, figures, tables and equations^[1]. The repository includes evaluation code that works with any model behind an OpenAI-compatible API^[2].

The second pillar is training. When an agent is fine-tuned on teacher trajectories, those trajectories mix the reasoning and actions that actually led to the answer with redundant or failed attempts. The authors propose counterfactual credit reweighting: for each reasoning step and action in a trajectory, estimate how much it contributed to the final answer, by asking what would change without it, and weight training by that contribution^[1].

3. What did it show? Only what the abstract states

The abstract's results fall into two groups: comparisons without training, and gains after training^[1].

• Without training: The page-native approach, with no training at all, outperformed OCR-based baselines across models such as GPT-5.4, Gemini-3.5-Flash and Qwen3.6-27B, while making fewer tool calls. The authors themselves introduce this result with the word "surprisingly."
• After training: On DRPixelBench, retrieval recall improved by about 15 points and final accuracy by over 10 points. The trained agent also beat the untrained baseline on four existing benchmarks.
• Out of distribution: On ViDoRe-V3, which was not used in training, DR Pixel came within 1.8 points of an oracle given the gold pages directly.

From this, the authors conclude that the conventional OCR-then-retrieve pipeline is not only unnecessary but suboptimal. The abstract does not say which OCR methods were used as baselines, what it cost in compute or storage to handle page images, or which four existing benchmarks were used and by what margins.

4. A worked example

Consider someone in a drug company's medical affairs team reviewing several papers on a drug's safety and pulling the frequency of a particular adverse event from a table. This is a hypothetical scenario for illustration, not an experiment from the paper.

What the reviewer is worried about is the loss of information through OCR that the paper describes. DR Pixel tries to keep tables and figures intact by making the page image the unit of retrieval and reasoning. But the paper's evaluation reports accuracy on a benchmark built from scientific papers; it does not guarantee that each number extracted is correct.

5. What is not new, and where are the limits?

Retrieving page images is not new. ColPali proposed embedding page images directly instead of extracting text to build an index, and created ViDoRe to evaluate it^[3]. Its successor, ViDoRe V3, evaluates RAG on visually rich documents from retrieval through to answer, and reports that non-text elements such as figures and tables remain hard for models^[4]. What is new here is running the whole job of an agent that searches and reasons repeatedly on page images, and the reweighting method for learning from noisy teacher trajectories.

There are at least four limits. First, the main evaluation uses DRPixelBench, which the authors built themselves, so whether its design favours their method needs checking by outside evaluation. Second, the material is scientific papers; the abstract does not show whether internal documents with different layouts, or handwritten records, give the same result. Third, fewer tool calls is not the same as lower compute cost. The cost of running image models, and of storing and indexing large numbers of page images, is not given in the abstract. Fourth, the conclusion that OCR is "not only unnecessary but suboptimal" is a claim about the settings the authors tested; reading it as "OCR is never needed" goes beyond the evidence.

6. Why is this topic clustering now?

It is not clustering. The collection pipeline found no other paper on the same question, so the cluster size is 1. The paper has 6 upvotes on Hugging Face and 40 GitHub stars, and it was picked because the implementer signal rose. This site uses "hot topic" for several independent signals rising together as a cluster, and this paper is still at the stage of a single paper drawing attention. Upvotes and stars measure attention, not whether the conclusions hold.

Even so, handling documents without converting them to text has been a steady line of retrieval research. ColPali moved the retrieval unit to page images^[3], and ViDoRe V3 extended evaluation to the whole RAG pipeline^[4]. DR Pixel can be placed as an extension of that line to agents that research a question through repeated searches. The repository says the benchmark, corpus, page images, indexes and a trained model are released on Hugging Face^[2], which makes outside verification easier to start.

7. What connects to pharmaceutical and regulatory practice?

The closest links are literature-based safety information gathering, and checking data cited in promotional materials. Both require picking numbers correctly from tables and figures in papers, and being able to show which paper, page and table each number came from. When OCR breaks a table, the link to the source breaks even before the number itself goes wrong. Treating the page image as the unit keeps answers tied to pages, so it may suit the work of checking sources.

Suiting the work, however, is not the same as removing the need to check. In material review, a person compares each cited number against the table in the original source. Even if AI can point to sources page by page, that is no reason to skip the comparison. If anything, a condition for adoption is that the system always attaches the source page, so the reviewer can open it and check.

The paper does not show that literature review systems should switch from OCR to page images. What it supports is narrower: in documents rich in tables and figures, information can be lost at the text conversion step, and one way to avoid that has shown results on a benchmark of scientific papers. Whether the same holds for a company's own documents is something to test with the released code and evaluation tools.