Can we trust LLMs to judge whether a research idea is novel?
The paper “Old Ideas, Novel Problems” introduces a benchmark for evaluating LLM‑based novelty judges and reveals that even minor prompt tweaks can swing model accuracy by over 50 percentage points. Six judges (GPT‑5.1, GPT‑5.2, GPT‑5.4, Claude Sonnet 4.5, Claude Opus 4.5, Claude Opus 4.6) show highly inconsistent judgments, and added retrieval or reasoning does not reliably improve performance. Two purpose‑built novelty systems underperform a simple Claude Opus 4.6 baseline, highlighting the fragility of current approaches. Engineers building AI‑driven scientific discovery pipelines must therefore treat LLM novelty assessment as a provisional signal rather than a definitive filter.
⚡ Key Takeaways
- Accuracy of novelty judgments can change by >50 pp when only a few words in the prompt are altered.
- Six LLM judges evaluated: GPT‑5.1, GPT‑5.2, GPT‑5.4, Claude Sonnet 4.5, Claude Opus 4.5, Claude Opus 4.6.
- Retrieval augmentation and extra reasoning steps do not consistently boost novelty detection performance.
- A simple prompted Claude Opus 4.6 baseline outperforms two dedicated novelty‑evaluation systems.
- The benchmark construction relies on an automated pipeline that mines ICLR 2026 peer‑review texts for explicit novelty statements.
Without reliable novelty judging, AI systems risk propagating redundant or incremental ideas, undermining the value of automated hypothesis generation in real research workflows.
✅ Practical Steps
- Use the provided automated pipeline (built around ICLR 2026 peer‑review mining) to generate your own novelty evaluation sets.
- Prompt Claude Opus 4.6 with the baseline prompt from the benchmark as a quick, out‑of‑the‑box novelty judge.
- When designing novelty judges, experiment with minimal prompt variations and record accuracy shifts to gauge sensitivity before deployment.
Want the full story? Read the original article.
Read on Hugging Face Blog ↗