Prompt wording swings AI novelty judges: GPT-5.4 falls from 93.5% to 40.9%
allen_ai · x · 2026-10-28
Allen AI evaluated frontier models as novelty judges over 154 pairs pairing human research ideas (praised as novel by ICLR reviewers) with ideas generated by Claude Sonnet 4.5.
- Prompt wording matters hugely: asked which idea reviewers had judged novel, GPT-5.4 picked the human idea 93.5% of the time; asked simply which idea was novel, that fell to 40.9%.
- Models vary widely: Claude Opus 4.6 scored 100% under the reviewer-referencing prompt but only 72.7% on direct novelty judgment; Claude Sonnet 4.5 managed just 60.4% under the former.
- Prompt wording had a smaller effect when judging human-only pairs.
Takeaway: models are good at echoing cited human judgments but unstable at independently assessing novelty — LLM-assisted review needs careful prompt design.
Related event: Ai2 Benchmark Finds LLM Novelty Judgments Flip with Prompt Wording(10 posts)→
More from Research
- Using diffusion models and force matching to learn free-energy surfaces for molecular simulation — TimothyDuignan · 2026-10-28
- The AlphaFold Data Problem: Quantum Chemistry Models as Training Data for Larger Scales — TimothyDuignan · 2026-10-28
- Defining recursive self-improvement: clamped compute means RSI isn't happening — dfrsrchtwts · 2026-10-28
- Position Paper: Generative Molecular Dynamics Evaluations Let Models Claim Arbitrary Wins — rishabh16_ · 2026-10-28
- Pangram Calls for Third-Party AI Detector Benchmarks, Offering 200K-Credit Grants — gnukeith · 2026-10-28
- Economist re-runs viral Goldman AI labor chart, finds the inference fragile — danielrock · 2026-10-28