AI2 researchers tested whether large language models can tell genuinely novel research ideas from less novel ones, and found that small changes to the judging prompt can dramatically alter the verdict. In one test, GPT-5.4’s accuracy fell from 93.5% to 40.9% on the same human-and-AI idea pairs after a prompt change.
Allen Institute for AI Watch analysis
What happened
The study introduces a benchmark, data and code for testing AI systems that judge research novelty. The researchers built two evaluation sets from ICLR 2026 peer reviews: one compares human-authored ideas with stronger and weaker evidence of novelty; the other pairs human ideas with ideas generated by a basic language model.
Across six LLM judges, the researchers varied prompts, reasoning budgets, access to related papers and other evaluation choices. Prompt changes produced some of the largest swings. On the human-only set, a prompt change shifted Claude Opus 4.6’s accuracy by 21 percentage points. The researchers also found that more search and reasoning did not reliably improve results.
In one cost comparison on the human-and-generated set, a simple Claude Opus 4.6 setup scored 0.854 macro-F1 for $2.59. Two purpose-built evaluators scored lower and cost more: Idea Novelty Checker scored 0.642 for $30.66, while Sakana AI’s AI Scientist scored 0.772 for $82.05.
The study and benchmark announcement comes from Ai2. The researchers note that ideas generated for one part of the benchmark were given weak labels: they were assumed to be less novel than the paired human ideas, rather than individually verified as such.
Why it matters
AI systems are increasingly asked to propose research directions, but a plausible-sounding idea is not necessarily a new one. If the model judging novelty can flip its answer when a few words change, researchers may mistake a prompt’s influence for a reliable assessment of originality.
The results also complicate the tempting fix of spending more on search, reasoning or specialist evaluation tools. In this benchmark, extra machinery did not guarantee better scores. The evidence is a specific test built from selected peer-reviewed ideas, not a verdict on every scientific field or every novelty-evaluation system. But it gives researchers a practical warning: test the judge’s stability, not just its headline accuracy.
Our read
This is a useful reality check for automated science. A novelty score looks satisfyingly tidy; the hard part is showing it means the same thing after a modest prompt edit. Ai2’s benchmark gives researchers a way to probe that weakness rather than taking the score on faith. If you are building an AI ideation pipeline, vary the wording and evaluation setup before treating its originality verdict as a gatekeeper.
What to watch
- Whether other research groups reproduce the prompt sensitivity across fields and newer models.
- Whether future benchmarks test ideas against broader and more independently checked evidence of novelty.
- Whether AI research tools report stability across prompts alongside their accuracy scores.
Discussion spark: Should an AI-generated research idea be treated as novel only when its evaluation stays stable across prompts, even if that makes the screening process slower?
Sources and evidence
- Can we trust LLMs to judge whether a research idea is novel? (27 October 2026, 16:12 UTC)
Independent WittyWires tracker for public updates about Allen Institute for AI. Not affiliated with or endorsed by Allen Institute for AI; this is not an official account.