LLM novelty judges fail on AI-generated research ideas

Hugging Face and Ai2 have released a new study titled Old Ideas, Novel Problems, which examines how reliably large language models can judge the novelty of scientific research ideas. Published in late October 2026, the work reveals that current AI systems often fail to distinguish genuinely new concepts from repackaged existing work, especially when evaluating ideas generated by other AI models.

What happened

The research team developed a benchmark to test six leading LLMs, including GPT-5.1 through GPT-5.4 and Claude Sonnet 4.5 through Claude Opus 4.6. They found that seemingly small adjustments to evaluation prompts could cause accuracy swings of more than 50 percentage points. In one specific test involving GPT-5.4, changing a few words in the instruction caused the model’s ability to identify the more novel idea in a pair to drop from 93.5% to 40.9%, which is below random chance.

The study highlights a critical gap in automated scientific discovery tools. While systems like AutoDiscovery and Theorizer can generate hypotheses, they rely on novelty judges to filter out incremental or duplicate work. The results show that these judges are highly unstable. A prompt that works well for one model may degrade performance for another, and adding more reasoning steps or retrieval capabilities does not consistently improve accuracy. In some cases, expensive, purpose-built evaluation systems performed worse than simple prompted baselines.

To create the benchmark, the researchers mined peer reviews from ICLR 2026. They identified 154 accepted papers where reviewers explicitly praised novelty and 145 rejected papers where reviewers questioned originality. This created a Human-Only dataset with strong evidence of novelty levels. They then created a Human+Generated dataset by pairing these human ideas with ideas produced by a basic LLM generator. This setup allowed them to test how judges perform on the AI-generated content they are increasingly expected to evaluate.

How it works

The evaluation pipeline tests novelty judges in two modes: pointwise, where the model rates the absolute novelty of a single idea, and pairwise, where it compares two ideas to determine which is more novel. The researchers varied several factors independently, including the prompt text, the reasoning budget allocated to the model, whether the judge could access related literature, and the formatting of the input ideas.

A key finding is that novelty judges are brittle when facing AI-generated text. When the prompt explicitly told the model that one idea in a pair was novel, GPT-5.4 performed well. However, when the prompt simply asked which idea was novel without that hint, performance collapsed. This suggests that models may rely heavily on contextual cues rather than deep semantic understanding of scientific originality. The study also found that rewriting ideas into a common structure did not eliminate these inconsistencies, indicating the issue is not just about writing style differences.

The researchers also compared simple prompting against agentic workflows. They tested GPT-5.6 Sol with an extra-high reasoning budget and live iterative search against standard retrieval setups. Despite costing roughly ten times more per decision, the agentic judge achieved a macro-F1 score of 0.78, lagging behind a simple Claude Opus 4.6 baseline that scored 0.89. This demonstrates that increased compute and search complexity do not automatically translate to better novelty detection.

Key details

• The study evaluates six models: GPT-5.1, GPT-5.2, GPT-5.4, Claude Sonnet 4.5, Claude Opus 4.5, and Claude Opus 4.6.
• Prompt changes caused accuracy drops of over 50 percentage points for GPT-5.4 on AI-generated ideas.
• The Human-Only dataset contains 154 high-novelty and 145 lower-novelty human-authored ideas from ICLR 2026 reviews.
• Purpose-built evaluators like Idea Novelty Checker cost $30.66 per decision but scored only 0.642 macro-F1.
• A simple Claude Opus 4.6 prompt cost $2.59 and achieved a higher macro-F1 of 0.854.
• An expert annotator agreed with the review-mining polarity in 47 out of 50 checked cases.

Why it matters

For engineering teams building RAG systems or automated research assistants, this study serves as a warning against trusting LLM judgments of originality without rigorous validation. If your system uses an LLM to filter or rank generated content, the output may be highly sensitive to prompt phrasing rather than actual content quality. This instability can lead to the rejection of valid ideas or the promotion of derivative work, undermining the value of the automation.

The findings also challenge the assumption that more advanced agentic workflows solve evaluation problems. Spending significantly more on API calls for iterative search and reasoning may yield worse results than a carefully tuned, simpler prompt. This has direct implications for cost optimization and system architecture. Developers should prioritize robust benchmarking on AI-generated data rather than relying on performance metrics derived from human-written texts.

What you can do

• Test your novelty judges on AI-generated ideas, not just human-written samples, to expose instability.
• Measure sensitivity by varying prompt wording slightly and observing changes in verdicts.
• Avoid assuming that higher reasoning budgets or retrieval steps will improve novelty detection accuracy.
• Use pairwise comparisons with caution, as verdicts can flip based on minor prompt adjustments.
• Consider using simple, low-cost prompted baselines before investing in complex agentic evaluation pipelines.
• Adopt the provided benchmark and pipeline to evaluate new models as they become available.