Can we trust LLMs to judge whether a research idea is novel?
A new Hugging Face study finds large language models often fail to reliably judge the novelty of research ideas, with minor prompt changes causing accuracy swings of over 50 percentage points.
Hugging Face has released Old Ideas, Novel Problems, a study examining how accurately large language models can assess the novelty of scientific research ideas. The paper includes a benchmark, datasets, and code for testing novelty judges, along with an automated pipeline for creating evaluation sets. The research is part of broader efforts at AI2 to develop AI systems that assist in scientific discovery, such as DataVoyager, AutoDiscovery, and Theorizer, which help researchers analyze data, generate hypotheses, and synthesize evidence.
The study highlights that novelty judges—LLMs tasked with determining whether a research idea is new—are not standardized and can produce inconsistent results based on choices like model selection, prompts, and evaluation formats. These judges are often validated using human-written papers rather than AI-generated ideas, which they are increasingly meant to evaluate. The findings reveal that even minor adjustments, such as changing a few words in a prompt, can alter a model’s accuracy in identifying novel ideas by more than 50 percentage points.
To test novelty judges, researchers built two evaluation pools: one with 154 human-authored ideas from top-rated ICLR 2026 papers where reviewers explicitly praised novelty, and another with 145 ideas from rejected papers where reviewers raised originality concerns. A separate check confirmed the accuracy of the extracted novelty signals in 47 of 50 cases. These pools form the Human-Only setting, designed to assess whether judges can distinguish between clearly novel and less novel human-authored ideas.
The study also created a Human+Generated setting, pairing the validated human ideas with 154 AI-generated ideas from a basic LLM generator. An expert review of 30 pairs found the human ideas preferred in all cases, with none of the generated ideas judged novel. The research tested six LLM judges—GPT-5.1, GPT-5.2, GPT-5.4, Claude Sonnet 4.5, Claude Opus 4.5, and Claude Opus 4.6—across various evaluation setups, revealing that prompt changes alone could drastically shift performance, particularly for AI-generated ideas.