Research map / Large language models / Language, safety and interpretability
How Reliable is Language Model Micro-Benchmarking?
Gregory Yauney, Shahzaib Saqib Warraich, Swabha Swayamdipta
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase micro-benchmark size from 10 to 1000 examples (ablation) | tinyBenchmarks · BBH | MDAD: 16 -> 4 -> 1 | Improved |
| Use Anchor Points at 10 examples vs other methods | micro-benchmarking · BBH | MDAD: 6 vs ~15 for others | Improved |
| Use Anchor Points at 1000 examples | Anchor Points · MMLU-Pro | MDAD | Hurt |
| Use tinyBenchmarks instead of random sampling at 10 examples | micro-benchmarking · MMLU | MDAD: 20 -> 12.5 | Improved |
| Use model-dependent selection (tinyBenchmarks) instead of random at 500 examples | micro-benchmarking · MMLU | MDAD: 2 vs 2 | No clear effect |
| Use model-dependent selection instead of random at >=250 examples | micro-benchmarking · MMLU, MMLU-Pro, BBH | MDAD: all MDADs 2 or less | No clear effect |
| Use rank correlation vs MDAD to distinguish methods at <=100 examples | meta-evaluation · MMLU | tau vs MDAD | Improved |
| Evaluate micro-benchmark on held-out draw of full benchmark | micro-benchmarks · full benchmarks | MDAD | No clear effect |
| Select micro-benchmarks per subtask and evaluate on fresh draw | micro-benchmarks · subtasks | MDAD | Hurt |
Closest papers by meaning
- BenchmarkCards: Standardized Documentation for Large Language Model Benchmarks (NeurIPS 2025 · similarity 0.91)
- Signal and Noise: A Framework for Reducing Uncertainty in Language Model Evaluation (NeurIPS 2025 · similarity 0.91)
- Aligning Language Model Benchmarks with Pairwise Preferences (NeurIPS 2026 · similarity 0.91)
- tinyBenchmarks: evaluating LLMs with fewer examples (ICML 2024 · similarity 0.91)
- How Benchmark Prediction from Fewer Data Misses the Mark (NeurIPS 2025 · similarity 0.90)
- metabench - A Sparse Benchmark of Reasoning and Knowledge in Large Language Models (ICLR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
