Can we trust LLMs to judge whether a research idea is novel?
- ID
- 40028
- Status
- summarized
- Published
- 28 Oct 2026, 12:12 AM
- Fetched
- 28 Oct 2026, 12:54 AM
- Provider
- Hugging Face Blog
- Category
- developer-ai
- Original URL
- https://huggingface.co/blog/allenai/novelty-bench
- Source URL
- https://huggingface.co/blog/feed.xml
Summary
- Score
- 7.5
- Created
- 28 Oct 2026, 12:54 AM
- Tags
- Audience
- ai_ml_learnersdevelopersvibe_coderssaas_founders
What happened
Ai2 released "Old Ideas, Novel Problems," a study plus benchmark, data, code, and an automated pipeline for testing whether LLMs can judge the novelty of research ideas. Across six judges (GPT-5.1, GPT-5.2, GPT-5.4, Claude Sonnet 4.5, Claude Opus 4.5, Claude Opus 4.6), changing just a few words in a prompt shifted accuracy on a pairwise "which idea is more novel" task by more than 50 percentage points. Retrieval and extra reasoning did not consistently help, and two purpose-built novelty evaluation systems scored worse than a simple prompted Claude Opus 4.6 baseline; the authors also note judges are usually validated on human-written papers, not the AI-generated ideas they are meant to score.
Why it matters
If you use an LLM as a judge anywhere in your pipeline, this is direct evidence that your prompt wording is a bigger variable than your model choice: the same ideas flipped verdicts by over 50 percentage points from small prompt edits, and adding retrieval or reasoning budget did not reliably fix it. Concretely, version-pin your judge prompts, re-run a fixed calibration set whenever you touch the wording, and don't assume a fancier evaluation harness beats a plain strong-model baseline — the two purpose-built systems here lost to one.
Discussion angle
If a few words of prompt can swing a judge's accuracy by 50+ points, what is the minimum calibration harness a small team should keep — fixed labeled set, pinned prompt versions, and pairwise rather than absolute scoring — and where have you seen an LLM judge quietly drift after a prompt 'cleanup'?