LiveBench
LiveBench resists contamination by refreshing questions from recent sources and grading automatically against ground truth across math, coding, reasoning, and more. Compare scores within a question-set release and use per-category results to match workloads.
LiveBench is a general-capability benchmark built to resist data contamination, the problem where models memorize test items that leaked into training data. It does this by drawing questions from recently published material and refreshing the question set over time, so a model cannot have seen the current items during training.
What It Measures
LiveBench covers a broad set of categories including math, coding, reasoning, data analysis, language tasks, and instruction following. Each category contains tasks with objective, verifiable answers. The headline is an overall score averaged across categories, with per-category scores that profile a model's strengths. Because every answer has a ground truth, scoring needs no human or LLM judge.
Methodology
Questions are sourced from recent inputs such as newly released math competitions, recent papers, and fresh news articles, and the set is periodically rotated. Every task is designed so the correct answer can be checked automatically against ground truth, avoiding the bias and cost of model-as-judge grading. Models are run under standardized prompting, and results are published on a public leaderboard that updates as new questions and models are added. The combination of fresh data and objective grading is the benchmark's defining feature.
How to Interpret Results
Because LiveBench rotates questions, scores are most meaningful when compared within the same release of the question set; older numbers may not be comparable to newer ones. Use the per-category breakdown to match a model to your workload, since a strong overall score can hide weakness in, say, data analysis. The contamination resistance makes LiveBench a useful cross-check when a model posts suspiciously high scores on older, well-known benchmarks.
Limitations
Frequent refreshes improve fairness but reduce historical comparability, complicating long-term tracking. Sourcing fresh, high-quality, objectively gradable questions at scale is hard, so coverage in some categories is thinner than in mature benchmarks. Objective grading favors tasks with crisp answers and under-tests open-ended generation. No benchmark fully eliminates contamination risk, and LiveBench's effectiveness depends on disciplined, ongoing maintenance.