Novelty-Bench: Evaluating LLM Reliability in Scientific Research Assessment

Assessing the Reliability of AI Novelty Judgments

Automated scientific discovery relies on the ability of AI systems to distinguish between genuinely novel research ideas and those that merely repackage existing work. A new study, Old Ideas, Novel Problems, demonstrates that current Large Language Model (LLM) novelty judges are unreliable, with performance that fluctuates wildly based on minor changes to prompts, evaluation setups, and the nature of the ideas being assessed.

The Novelty-Bench Benchmark

To measure the effectiveness of novelty judges, researchers at the Allen Institute for AI (Ai2) developed a benchmark using ICLR 2026 peer reviews. Rather than relying on paper acceptance status, the team mined reviewer comments that explicitly discussed the originality of a contribution.

Two evaluation settings were created:

  • Human-Only: 154 high-novelty ideas (top 10% of accepted papers) paired with 145 lower-novelty ideas (bottom 10% of rejected papers).
  • Human+Generated: The 154 high-novelty human ideas paired with 154 ideas generated by a basic LLM, providing a baseline for comparing human research against AI-generated output.

Sensitivity to Prompting and Model Instability

Testing six LLM judges—including models from the GPT-5 and Claude 4.5/4.6 series—revealed that novelty evaluation is brittle.

  • Prompt Sensitivity: In one experiment, a simple change in prompt wording caused a model's accuracy in identifying the more novel idea to drop from 93.5% to 40.9%.
  • Inconsistency: The same prompt changes that caused minor shifts in the Human-Only dataset led to dramatic performance swings on the Human+Generated dataset, indicating that judges are particularly unstable when evaluating AI-generated content.
  • Lack of Transferability: Prompts that improved accuracy for one model often degraded performance in others, suggesting that there is no universal "best practice" for prompting novelty judges.

Performance of Specialized vs. Simple Systems

More complex, purpose-built systems do not necessarily yield better results than simple, prompted baselines. The study compared standard-retrieval judges against agentic systems with high reasoning budgets and live, iterative search.

  • Efficiency: A simple Claude Opus 4.6 prompt achieved a macro-F1 score of 0.854 at a cost of $2.59 per decision.
  • Cost-Performance Gap: Specialized systems like the "Idea Novelty Checker" and the "AI Scientist" evaluator scored lower (0.642 and 0.772 respectively) while costing significantly more ($30.66 and $82.05 per decision).

Implications for Automated Discovery

These findings suggest that current methods for verifying the novelty of AI-generated research are insufficient. Because novelty judges behave differently on AI-generated ideas than on human-authored ones, validating these systems on human papers alone is an inadequate proxy for real-world performance. The researchers have released the benchmark, data, and code to help the community develop more robust, dependable evaluation methods for AI-driven scientific discovery.

Sources