LLM novelty judges prove fragile on AI‑generated research ideas, new benchmark shows

If AI is going to suggest fresh research directions, how can we be sure those suggestions aren’t just repackaged old work? Researchers are already using large language models to draft hypotheses, but the next step—deciding whether a proposal is genuinely new—still relies on another LLM acting as a “novelty judge.”

Allen Institute for AI (Ai2) has now released a benchmark called Old Ideas, Novel Problems that puts those judges under a microscope. The benchmark comes with data, code, and an automated pipeline so the community can build new evaluation sets and test how stable novelty judgments really are.

Source: Hugging Face: a]:hidden"> Can we trust LLMs to judge whether a research idea is novel? ↗

To create a reliable test set, Ai2 mined ICLR 2026 peer reviews for explicit statements about originality. They assembled a high‑novelty pool of 154 human‑authored ideas from top‑rated accepted papers and a low‑novelty pool of 145 ideas from rejected papers where reviewers flagged originality concerns. An independent annotator agreed with the extracted novelty signals in 47 of 50 cases, confirming the pools’ validity.

Six LLM judges—including three GPT‑5 variants and three Claude models—were evaluated on both human‑only and human‑plus‑generated idea pairs. Prompt wording proved decisive: a minor rephrase dropped GPT‑5.4’s accuracy from 93.5 % to 40.9 %, below chance. Adding retrieval or extra reasoning did not consistently improve scores, and a simple Claude Opus 4.6 prompt outperformed two purpose‑built novelty systems while costing a fraction of the price.

Source: Hugging Face: a]:hidden"> Can we trust LLMs to judge whether a research idea is novel? ↗

The results expose a brittle core: small prompt changes can swing judgments by 20‑plus points, and the effect is even larger on AI‑generated ideas than on human‑authored ones. No single prompt works best across models, and more compute or literature search does not guarantee better consistency. This instability means current novelty judges cannot be trusted as definitive arbiters.

The benchmark also has limits. AI‑generated ideas were produced by a very basic generator and labeled only as “less novel” by assumption, not by independent verification. The data come solely from ICLR 2026 reviews, so findings may not transfer to other fields or to more sophisticated idea‑generation pipelines.

If you’re experimenting with AI‑driven ideation, treat novelty judgments as provisional signals. Run several prompts, compare multiple LLMs, and always back the final decision with human expert review before committing resources to a project.

All the benchmark data, code, and the automated review‑mining pipeline are publicly available on Hugging Face. Researchers can contribute new peer‑review datasets, test their own judges, or adapt the pipeline to other scientific venues. Keeping an eye on follow‑up work from Ai2 will be essential as the community works toward more reliable novelty evaluation.