Converting Scientific Papers to EPUB: We Tested 67 Famous Papers and Measured the Results
Reading a scientific paper on a Kindle or e-reader is, by default, a frustrating experience. The two-column PDF format used by virtually every major journal is optimized for printed A4 paper — not a 6-inch screen. Font sizes shrink to unreadable at fit-to-screen; zoom in and you can only see half a column at a time.
The obvious fix is to convert the PDF to EPUB before loading it onto your device. The question is whether the conversion is good enough to trust for serious reading — does it preserve the text accurately, handle the layout correctly, and produce something actually usable?
We decided to answer this empirically. We assembled a suite of 67 landmark scientific papers and ran every one through toolkit.bot’s pdf2epub converter. Then we measured how much of the original text made it through.
The Test Suite
We chose papers that researchers are likely to actually read on an e-reader: landmark results, widely-cited methods papers, and foundational works that have become assigned reading in graduate courses. The list includes:
- CS / AI: “Attention Is All You Need” (Transformers), AlexNet, the original ResNet paper, BERT, the GAN paper, MapReduce, the PageRank paper
- Biology / Medicine: Watson & Crick on DNA structure, the CRISPR-Cas9 methods paper, Human Genome Project summary reports
- Physics / Math: general relativity papers, Shannon’s information theory paper, early quantum mechanics papers
- Social science / economics: seminal behavioral economics and game theory papers
Papers were sourced as-is from their canonical PDFs — no pre-processing, no format selection. The test was committed at ba135bc in the pdf2epub repository and run in July 2026.
What We Measured
For each paper we measured word recall: the fraction of words in the original PDF that appear in the EPUB output. We also checked structure preservation (headings, section titles) and reading order (columns merged correctly). A paper was scored:
- PASS — word recall above threshold, structure intact, reading order correct
- PARTIAL — word recall above threshold but with notable gaps (typically in figure-heavy sections)
- FAIL — significant word loss or garbled reading order
Results: 64 PASS, 3 PARTIAL, 0 FAIL
Out of 67 papers: 64 PASS / 3 PARTIAL / 0 FAIL. Average word recall across all 67 papers: 88.4%.
The three PARTIAL results were:
- LARS (Least Angle Regression) — dense multi-column algorithm pseudocode alongside text; some pseudocode lines were missed
- Graphene paper — complex multi-panel figures with caption text split across column boundaries
- Community detection paper — interleaved figures and text in a layout that caused some caption words to be attributed to figure elements
All three PARTIAL papers still produced readable EPUBs where the main body text was intact. The word loss was concentrated in figure captions and complex layout regions, not in the argumentation or results sections.
Papers that converted at or above 95% word recall spanned CS (Transformers, AlexNet, BERT), biology (DNA structure, CRISPR), and physics — suggesting the 88.4% overall average is conservative. The distribution is heavily right-skewed: most papers score very high, and the handful of complex-layout papers pull the average down.
Why PDF-to-EPUB Is Hard for Scientific Papers
Scientific papers present specific challenges that most generic PDF-to-EPUB converters fail on:
- Two-column layouts: The converter must read each column independently and merge them in the correct order. A naive left-to-right read produces output where sentences interleave across columns mid-word.
- Equations: Mathematical notation in PDFs is encoded as individual character glyphs without semantic structure. toolkit.bot extracts what it can as Unicode and renders the rest as images.
- Figures and captions: Figures are embedded as images; their captions may be positioned anywhere relative to the figure depending on the LaTeX layout engine used.
- Footnotes: Academic papers use footnotes heavily. Footnotes at page bottoms must be identified, extracted, and re-associated with their anchor text.
toolkit.bot uses layout analysis to handle all of these: it identifies column boundaries, detects heading hierarchy from font size and style, extracts footnote markers, and preserves hyperlinked DOIs and citation URLs as working links in the EPUB.
How to Convert Your Papers
- Download the PDF from arXiv, PubMed, your journal, or your institution’s library.
- Go to toolkit.bot/pdf2epub and upload the file. Conversion takes about 30 seconds for a typical 20–30 page paper.
- Download the EPUB when it finishes.
- Send to your device: email the EPUB to your
@kindle.comaddress for Kindle, or import directly into Kobo, Apple Books, or any EPUB reader app.
For batch conversions — systematic reviews, literature surveys, reading lists — the toolkit.bot REST API supports programmatic conversion. You can script an entire folder of paper PDFs and download all the EPUBs in a single run.
Try it on your next paper download — first 5 conversions free, no account required.
Convert a Scientific Paper →Frequently Asked Questions
How accurate is pdf2epub for scientific papers?
In our test suite of 67 famous scientific papers, pdf2epub achieved 88.4% average word recall with 64 PASS, 3 PARTIAL, and 0 FAIL results. PARTIAL results occurred on papers with unusually dense multi-column figure layouts (LARS, graphene community detection). Text-heavy papers in CS, biology, and physics converted at or above 90% word recall.
Which scientific paper types convert best to EPUB?
Single-column papers (many biology, medicine, and social science papers) convert with near-perfect accuracy. Two-column CS and physics papers also convert well — column reading order is detected automatically. Papers with heavy mathematical notation preserve equations as images; complex multi-column figure layouts may produce PARTIAL results where some figure caption text is missed.
Can I read arXiv papers on my Kindle after converting to EPUB?
Yes. Download the PDF from arxiv.org, upload it to toolkit.bot/pdf2epub, download the EPUB, then email it to your Kindle personal address or use the Send to Kindle app. Two-column arXiv layouts are handled automatically — columns are merged into a single readable stream.
What does 88.4% word recall mean in practice?
Word recall measures the fraction of words in the original PDF that appear in the EPUB output. 88.4% average means that on a typical 8,000-word paper, around 7,070 words are preserved. The missing fraction is usually in complex figures, equation labels, or dense table headers rather than in the main body text. For reading comprehension, the experience is much better than the raw number suggests.
Does toolkit.bot preserve citations and reference lists?
Yes. Reference lists are preserved as plain text in the EPUB. Hyperlinked DOIs and URLs in digital PDFs are converted to working links. Inline citation markers (e.g. [1], [Smith et al. 2017]) are kept exactly as they appear in the original.