Research map / Label-efficient and robust learning / Recognition and long-tail
STABLEVAL: Disagreement-Aware and Stable Evaluation of AI Systems
Akash Bonagiri, Gerard Janno Anderias, Saee Patil, Angelina Lai, Devang Borkar, Gezheng Kang, Ishant Gandhi, Setareh Rafatirad, Houman Homayoun
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 |
|---|---|---|---|
| replace majority vote with posterior expected credit (PEC) aggregation | agent ranking on MT-Bench · MT-Bench human judgments | mean rank std under annotator subsampling: 0.259 -> 0.197 | Improved |
| replace majority vote with PEC | aggregation under adversarial annotators · synthetic, 40% adversarial annotators | MSE (x10^-3): 8.42 -> 5.54 | Improved |
| replace majority vote with PEC | aggregation under adversarial annotators · synthetic, 40% adversarial annotators | Kendall's tau: 0.977 -> 0.996 | Improved |
| use PEC instead of Dawid-Skene for label recovery | aggregation · synthetic stress tests | MSE vs latent ground truth: 3.47 -> 5.54 at 40% adversarial | Hurt |
| use PEC instead of majority vote with only 3 labels per item | aggregation · synthetic, 3 labels per item | MSE (x10^-3): 4.78 -> 5.71 | Hurt |
| use PEC instead of majority vote when agent quality gaps are tight | aggregation · synthetic, tight agent gap | Kendall's tau: 0.939 -> 0.953 | Improved |
| use PEC instead of majority vote in high-consensus datasets | aggregation · ConvAbuse and QAGS | ranking stability: 0.000 -> 0.000 mean rank std | No clear effect |
| use PEC instead of majority vote with only two annotators | aggregation · MSLR | mean rank std: 0.503 -> 0.587 | Hurt |
| increase labels per item from 3 to 9 (ablation) | all aggregation methods · synthetic annotation density | MSE (x10^-3): 4.78 -> 1.09 (MV); 5.71 -> 0.96 (PEC) | Improved |
| vary MT-Bench pairwise-to-correctness mapping (tie as win, loss, or partial values) (ablation) | PEC agent ranking · MT-Bench | agent rank | No clear effect |
| fraction of strict or lenient annotators raised to 40% (ablation) | majority vote · synthetic | MSE (x10^-3): 1.93 -> 3.65 (strict); 1.77 -> 4.44 (lenient) | Hurt |
| raise hard item rate from 0% to 20% (ablation) | majority vote · synthetic | MSE (x10^-3): 1.10 -> 2.62 | Hurt |
Closest papers by meaning
- AI Evaluation Should Require Standardized Item-Level Data Releases (NeurIPS 2026 · similarity 0.86)
- Consistency as a Testable Property: Statistical Methods to Evaluate AI Agent Reliability (NeurIPS 2026 · similarity 0.86)
- Validating LLM-as-a-Judge Systems under Rating Indeterminacy (NeurIPS 2025 · similarity 0.86)
- Measuring Error Alignment for Decision-Making Systems (AAAI 2025 · similarity 0.85)
- Let's Think in Two Steps: Mitigating Agreement Bias in MLLMs with Self-Grounded Verification (ICLR 2026 · similarity 0.85)
- Fantastic Bugs and Where to Find Them in AI Benchmarks (NeurIPS 2025 · 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.
