Can we trust LLMs to judge whether a research idea is novel?
Web page · https://huggingface.co/blog/allenai/novelty-bench
First seen by Oriel on .
AI summary
Ai2’s study and benchmark find that LLM-based scientific novelty judgments are highly sensitive to prompts, models, and evaluation setups, especially for AI-generated research ideas.
- The benchmark includes 154 high-novelty and 145 lower-novelty human ideas mined from ICLR 2026 reviews, plus 154 ideas from a simple LLM generator.
- For GPT-5.4, a small prompt change reduced Human+Generated pairwise accuracy from 93.5% to 40.9% and reversed more than half of verdicts.
- GPT-5.6 Sol with iterative arXiv search achieved 0.78 macro-F1, versus Claude Opus 4.6 at 0.89, while costing roughly 10× more per decision.
- A Claude Opus 4.6 majority-vote baseline reached 0.854 macro-F1 for $2.59, outperforming Idea Novelty Checker and Sakana AI’s AI Scientist in accuracy and cost.
- The released benchmark, data, code, and automated pipeline support testing judges across prompts, models, research communities, and future peer-review datasets.
#llm evaluation #scientific discovery #benchmarking #research novelty
Generated by gpt-5.6-sol. Check the original for details.
Aliases
| Source | External ID | URL |
|---|---|---|
| Hugging Face Blog | https://huggingface.co/blog/allenai/novelty-bench | https://huggingface.co/blog/allenai/novelty-bench Read discussion |