A series of studies conducted in 2025 and 2026 indicates that frontier AI models struggle to generate original research methods when lacking prior context. Researchers developed the InnoEval framework, which debuted in an arXiv paper in February 2026 and was presented at ICML 2026, to measure innovation through knowledge-grounded evaluations.
InnoEval outperformed traditional LLM-as-judge methods by up to 16.18% in F1 scores. However, testing revealed limitations in AI originality. In the Reconstruction benchmark, which analyzed 643 papers as of August 2026, models successfully recovered core ideas only 3–15% of the time using only pre-publication reference lists.
An October 2026 study in Science compared AI-generated hypotheses to human ones, assigning them average novelty scores of 3.406 and 3.968, respectively. Another large-scale analysis from October 2026 examined over 121,000 preprints, finding that LLMs frequently produce narrowly focused ideas with limited diversity.
Additional benchmarks, including InnoGym and InnovatorBench, were introduced between 2025 and 2026 to evaluate AI performance across complete research workflows. These tests also demonstrated low success rates in creating genuinely new methods, suggesting that current models are more effective as research assistants than as primary sources for new hypotheses.
