Comparing two texts beats single-text checks for author verification
A model that compares two texts at once shows higher accuracy when longer context is used, but results depend on training choices and data setup.
- Publication
- arXiv
- Stage
- arXiv version · journal reference supplied
- What we read
- Summary of the paper
- Authors
- Peter Kirby
- Universities and research institutions
- Georgia Institute of Technology
Institution metadata: OpenAlex record ↗
What they did and found
Researchers compared two setup options: having a model read two texts separately and compare them, versus feeding a single combined text to one classifier. The two-text, side-by-side approach performed better, especially when longer passages were available, in the tested tasks.
Why it matters
This shows that how data is arranged and how the model is trained can change results in real author checks for practical use today systems.
In plain terms, the researchers tested whether a model should read two texts separately and compare them, or read one combined version instead of together.
The two-text approach gave better results, especially when longer passages were available. They also used many examples and simple tricks to teach the model author style.
Original sources · 1
- Contrastive Learning for Authorship Verification ↗arXiv · 2026-09-23
Check the original paper for its authors, methods, version and access terms.
What remains uncertain
The results come from a similar author-check dataset; they may not apply to other tasks or very short texts.
What is your take?
Ask a question, add useful context or share a different perspective. Keep the conversation respectful and grounded.