Two independent benchmarking studies undercut the assumption that newer model generations are becoming epistemically safer: Linardon et al. 2025 found no clear evidence that GPT-4o reduces citation hallucination relative to earlier versions, and Peters & Chin-Yee 2025 found newer models generalize less accurately than their predecessors.
E35 reports that despite expectations of enhanced accuracy, GPT-4o showed no clear evidence of reduced hallucination rates relative to earlier versions, and E36 reports newer models tended to perform worse in generalization accuracy than earlier ones.
Written by Kimi K3 via Ollama Cloud · checked by GLM-5.3 via Ollama Cloud · 24 Oct, 04:56
Source chain
AEvery quote below was checked, without a model, to appear verbatim in its source.
- 01
“Despite expectations of enhanced accuracy, we found no clear evidence of reduced hallucination rates”
Full-text passage · no page number · evidence E35
This approach enabled a more fine-grained analysis of how topic characteristics influence LLM performance, revealing meaningful variation even within a single area of research. Second, while most earlier studies evaluated older versions of GPT [8,15,16], our use of GPT-4o allowed us to assess whether citation reliability has improved in newer iterations. Despite expectations of enhanced accuracy, we found no clear evidence of reduced hallucination rates, although cross-study comparisons are inherently limited due to variation in topic selection, task framing, and evaluation methods. A key novelty of this work in comparison to previous literature lies in its design. Rather than contrasting different disciplines or older model versions, we systematically manipulated topic familiarity and review specificity within the same research field. This approach revealed important topic-level differences in citation reliability that would have been obscured in cross-disciplinary comparisons. Thus, this study not only extends existing evidence on LLM hallucinations but also introduces a new framework for evaluating how characteristics of the research question itself influence model performance.
Passage read from www.ebi.ac.uk, which may be a preprint rather than the published version.
- 02
“Notably, newer models tended to perform worse in generalization accuracy than earlier ones”
Full-text passage · no page number · evidence E36
Notably, newer models tended to perform worse in generalization accuracy than earlier ones. Our results indicate a strong bias in many widely used LLMs towards overgeneralizing scientific conclusions, posing a significant risk of large-scale misinterpretations of research findings. We highlight potential mitigation strategies, including lowering LLM temperature settings and benchmarking LLMs for generalization accuracy. Keywords: large language models, algorithmic bias, science communication, overgeneralization Introduction Accurately communicating findings of scientific studies is vital for educating the public, informing policy, guiding behaviour, and advancing research [1,2]. To learn about, review, and communicate scientific findings, both experts (e.g. researchers) and laypeople (e.g. reporters and students) now increasingly use artificial intelligence (AI) chatbots (e.g. ChatGPT) powered by large language models (LLMs) [3–5]. AI chatbots can process vast amounts of scientific information and summarize content in easily understandable language, thus helping to spread scientific knowledge, promote evidence uptake, and facilitate research [3,6,7].
Passage read from www.ebi.ac.uk, which may be a preprint rather than the published version.