MeReader paper describes offline AI reading assistant that avoids spoilers
A research paper introduces MeReader, an AI companion for eBooks that runs entirely on consumer hardware and restricts answers to the point a reader has reached, so the assistant cannot reveal plot events ahead of the reader.
In plain words
Most AI assistants process a whole book at once, which risks spoiling the story. MeReader is a research system that runs offline on a reader's own device and only uses text up to the reader's current position. The authors say this prevents future plot details from leaking while still answering questions usefully.
What was announced
The paper, posted on arXiv, describes MeReader as a privacy-preserving, offline and progress-aware assistant for narrative reading on consumer hardware. Instead of feeding an entire book to a large language model, the system limits both retrieval and generation to the reader's position in the text. It combines vector similarity, BM25 and summary embeddings in a hybrid retrieval architecture intended to keep answers faithful to the reader's current context.
In evaluation on classical English books, using quantitative comprehension metrics and qualitative judging by independent LLMs, the authors report a quality-latency frontier: mid-sized local models of roughly 2b-4b parameters reached comprehension parity with larger 5b+ models while keeping response quality acceptable. They also state the progress-aware boundary mechanism prevented future content leakage without hurting retrieval quality. The authors position the system as an augmentation of digital reading and an alternative to cloud-dependent services, noting that prior work had not combined progress awareness, privacy and narrative eBook AI support.
What it does not show
- The evidence is limited to the paper's abstract; full evaluation methods, datasets and results are not available in the supplied passages.
- Reported results, including model comparisons and the spoiler-prevention claim, come from the authors' own evaluation and have not been independently confirmed.
- No information is provided about public availability, supported devices or release of the system.
For security and IT teams
- Scope
- The supplied source is the paper's abstract only, covering the system design and headline evaluation findings rather than the full paper.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.