Research map / Large language models / AI and society

Position: AI Competitions Provide the Gold Standard for Empirical Rigor in GenAI Evaluation

ICML 2025 · Large language models › AI and society · approach: Social-science view of AI evaluation

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Position paper arguing that AI competitions (e.g., Kaggle) offer a renewable, leak-resistant structure that should be treated as the gold standard for empirical rigor in GenAI evaluation.

Problem

Empirical evaluation of generative AI models is insufficient because of unbounded input/output spaces, lack of ground truth, feedback loops, and especially data leakage and contamination in static benchmarks.

Approach

Reframe GenAI generalization as novelty-based and treat leakage as the central pitfall; propose that parallel, time-bound AI competitions with hidden test data, prospective ground truth, and post-deadline data collection provide a renewable source of novel, leak-resistant evaluations, and that meta-analyses of competitions should be valued.

Key results

• Reports that every time Kaggle submissions were ensembled, little-to-no improvement to top-ranked solutions was obtained, suggesting competitions extract near-maximal signal from data.
• OpenVaccine Challenge competitors improved state-of-the-art mRNA degradation rate prediction by 25% within 4 weeks, and solutions generalized to longer RNA sequences.
• CAFA 5 drew nearly two thousand participants across 1,625 teams using prospective ground truth; the median Kaggle featured competition had 1,400 forum messages last year.

Surprising findings

• Prior work found ImageNet model rankings stayed consistent on new data despite heavy reuse, so overfitting from benchmark reuse was not the main issue.
• Kaggle leaderboard performance on public data was a good indicator of rank ordering on private holdout data despite many thousands of teams.

Limitations

• Competitions have artificial deadlines and may prohibit valuable late submissions; community and dynamic benchmarks have their own maintenance and bias issues.
• Competitions cover only a limited range of problems and may not be feasible for tasks requiring very specialized knowledge or long verification.
• Position is argued via case studies and anecdotes rather than systematic empirical evaluation.

Experimental setup

Datasets: Kaggle competition datasets (SETI Breakthrough Listen, TalkingData, LANL Earthquake, CAFA 5, AIMO, OpenVaccine) Meta Kaggle Nectar

Baselines: SEAL Leaderboards ARC-AGI FrontierMath Humanity's Last Exam LiveBench LiveCodeBench SWE-Bench LM Arena

Builds on: IID test-train splitting MapReduce

Closest papers by meaning

• Position: Evaluating Generative AI Systems Is a Social Science Measurement Challenge (ICML 2025 · similarity 0.87)
• The Leaderboard Illusion (NeurIPS 2025 · similarity 0.86)
• Pricing and Competition for Generative AI (NeurIPS 2024 · similarity 0.86)
• Adaptive auditing of AI systems with anytime-valid guarantees (NeurIPS 2026 · similarity 0.85)
• MLR-Bench: Evaluating AI Agents on Open-Ended Machine Learning Research (NeurIPS 2025 · similarity 0.85)
• Braess’s Paradox of Generative AI (AAAI 2025 · similarity 0.85)