Research map / Large language models / AI and society
Forest vs Tree: The (N, K) Trade-off in Reproducible ML Evaluation
Deepak Pandita, Flip Korn, Chris Welty, Christopher M Homan
Read the paperCheck your idea against 137,217 papers
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 responses per item K (ablation) | accuracy metric · all five datasets | p-value: p-value increases as K increases | Hurt |
| increase responses per item K (ablation) | TV metric · all five datasets | p-value: K=120 at NK=1000 on Toxicity | Improved |
| increase K (ablation) | Wins metric · DICES, D3code, JobsQ3 | p-value: optimal K in [25, 100] | Mixed |
| increase K (ablation) | KL-Div metric · all datasets | p-value: best K in [2,5] then worse, settles by K=300 | Mixed |
| increase K (ablation) | TV metric · all datasets | CI width: best CI for K in [1, 10] | Hurt |
| increase K (ablation) | KL-Div metric · all datasets | CI width: settles around K=100 | Improved |
| increase perturbation epsilon (ablation) | all metrics · five datasets | p-value: epsilon 0.1 to 0.4 lowers p-values | Improved |
| increase number of categories M (ablation) | TV metric · balanced simulations | p-value: p-values monotonically decrease with K as M grows | Improved |
| use unbalanced instead of balanced category prior (ablation) | TV metric · simulations | p-value: similar trends | No clear effect |
| use TV rather than accuracy | evaluation metric · five datasets | lowest N x K: TV needs smallest N x K, 250 to 1000 | Improved |
Closest papers by meaning
- Reproduce, Replicate, Reevaluate. The Long but Safe Way to Extend Machine Learning Methods (AAAI 2024 · similarity 0.89)
- Limits to scalable evaluation at the frontier: LLM as judge won’t beat twice the data (ICLR 2025 · similarity 0.87)
- quanda: An Interpretability Toolkit for Training Data Attribution Evaluation (NeurIPS 2026 · similarity 0.85)
- Position: Machine Learning Conferences Should Establish a "Refutations and Critiques" Track (NeurIPS 2025 · similarity 0.85)
- Benchmark Data Repositories for Better Benchmarking (NeurIPS 2024 · similarity 0.85)
- The State of Data Curation at NeurIPS: An Assessment of Dataset Development Practices in the Datasets and Benchmarks Track (NeurIPS 2024 · similarity 0.85)
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.
