LLMs hallucinate obscure papers at 98% rate, study finds
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
Researchers have quantified a critical failure mode in LLM reasoning: models hallucinate author attributions at rates exceeding 98% for obscure papers while accuracy improves dramatically for well-cited work. Using 1.92 million academic papers across eight disciplines, the study establishes that factual prevalence in training data directly correlates with hallucination likelihood. This finding reshapes how practitioners should interpret LLM outputs on niche topics and informs model evaluation protocols. The inverse relationship holds consistently across three state-of-the-art systems, suggesting the problem is structural rather than model-specific, with implications for deployment in domains where rare but accurate information retrieval matters.
Modelwire’s AI-generated summary of coverage from arXiv cs.CL.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The study isolates a mechanical cause rather than just documenting failure: training data prevalence directly predicts hallucination likelihood. This means the problem isn't random or fixable by prompt engineering alone, but baked into what the model learned.
This finding extends the reliability audit pattern established in recent coverage. The peer review audit from November 12th showed LLM reviewers miss injected errors at scale; the version drift story from November 16th exposed instability in model outputs over time. This authorship work adds a third dimension: even within a single model version, accuracy degrades predictably based on training data scarcity. Together, these pieces suggest that LLM deployment in academic and professional gatekeeping roles requires not just better prompts but explicit prevalence-based confidence thresholds. The mechanistic interpretability work from November 13th hinted that internal model states could predict failures; this study confirms the upstream cause is data distribution itself.
If the same research team or follow-up work demonstrates that retrieval-augmented generation (RAG) with verified sources eliminates the hallucination gap for obscure papers, that would signal a practical mitigation path. If major venues adopt prevalence-weighted confidence scores for LLM-assisted review by Q2 2027, adoption of this finding has moved from research to practice.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.CL
Benchmark reveals gaps in LLM peer review accuracy across error types
As LLM-based peer review moves into production use across academic venues, a new benchmark exposes critical gaps in how well these systems catch errors. Researchers injected 357 deliberate flaws spanning mathematical, numerical, and citation issues into real papers across astrophysics and ML, then tested twelve reviewer systems. The work introduces Skepthical, a multi-agent approach that…
MentionsLarge Language Models · Academic authorship attribution
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Dreaming Up Authors: Factual Prevalence and LLM Hallucinations in Academic Authorship Attribution”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.