AI BriefingAi2 releases benchmark showing LLM novelty judges are unreliable on AI-generated research ideas
·2026.10.28 01:12
Key point
A new benchmark from Ai2 reveals that small prompt changes can shift LLM novelty judgment accuracy by over 50 percentage points, and simple baselines outperform expensive purpose-built systems.
Details
- Ai2 releases the paper "Old Ideas, Novel Problems" and a benchmark to evaluate how reliably LLMs judge the novelty of scientific research ideas.
- Novelty evaluation is highly unstable for AI-generated ideas; changing a few words in a prompt shifted GPT-5.4's accuracy from 93.5% to 40.9% on the Human+Generated dataset.
- A simple prompted Claude Opus 4.6 baseline achieved 0.854 macro-F1 at a cost of $2.59, outperforming purpose-built systems like Idea Novelty Checker (0.642 macro-F1, $30.66) and Sakana AI's AI Scientist (0.772 macro-F1, $82.05).
- Adding live search and high reasoning budgets did not consistently improve judgments; an agentic GPT-5.6 Sol setup scored 0.78 macro-F1, trailing the simple baseline while costing roughly 10x more per decision.
- The benchmark includes a high-novelty pool of 154 human-authored ideas and a lower-novelty pool of 145, derived from ICLR 2026 peer reviews where reviewers explicitly discussed originality.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
NeurIPS 2026 paper proposes sampling coherent but cognitively unavailable research directionsgpu-unlock Linux module doubles FP/BF16 instruction rate on Nvidia 3000 series GPUs