SciPredict: Can LLMs Predict the Outcomes of Research Experiments?

Paper: SciPredict: Can LLMs Predict the Outcomes of Research Experiments in Natural Sciences?

Overview

SciPredict is a benchmark evaluating whether AI systems can predict experimental outcomes in physics, biology, and chemistry. The dataset comprises 405 questions derived from recently published empirical studies (post-March 2025), spanning 33 subdomains.

Dataset Structure

• Total Questions: 405 (5,716 rows including model responses)
• Domains: Physics (9 subdomains), Chemistry (10 subdomains), Biology (14 subdomains)
• Question Formats: Multiple-choice (MCQ), Free-format, Numerical

Key Fields

• DOMAIN: Scientific domain (Physics, Biology, Chemistry)
• FIELD: Specific field within the domain
• PQ_FORMAT: Question format (MCQ, Free-Format, Numerical)
• TITLE: Paper title
• URL: Paper URL
• PUBLISHING_DATE: Publication date
• EXPERIMENTAL_SETUP: Description of the experimental configuration
• MEASUREMENT_TAKEN: What was measured in the experiment
• OUTCOME_PREDICTION_QUESTION: The prediction task
• GTA: Ground truth answer
• BACKGROUND_KNOWLEDGE: Expert-curated background knowledge
• RELATED_PAPERS_DATA: Related papers information

Key Findings

• Model accuracy: 14-26% (vs. ~20% human expert accuracy)
• Poor calibration: Models cannot distinguish reliable from unreliable predictions
• Background knowledge helps: Providing expert-curated context improves performance
• Format matters: Performance degrades from MCQ → Free-form → Numerical