Research map / Large language models / Alignment and preferences
Pluralistic Leaderboards
Nika Haghtalab, Ariel D. Procaccia, Han Shao, Serena Lutong Wang, Kunhe Yang
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 |
|---|---|---|---|
| rank by Bradley-Terry scores instead of the local-stability ranking algorithms | LMArena leaderboard ranking · LMArena-based mixture of Mallows, top 20 models, phi=0.1 | stability ratio (gamma-hat) | Hurt |
| use committee selection via iterated rounding (Algorithm 1) instead of top-k Bradley-Terry prefix | committee selection · Synthetic mixture of Mallows, m=20 | stability ratio (gamma-hat) | Improved |
| use top-ranked candidate of each Mallows center instead of Algorithm 1 as committee | committee selection · Synthetic mixture of Mallows, m=20, phi=0.9 | stability ratio (gamma-hat) | Hurt |
| use ranking via geometric checkpoints or committee monotonicity instead of Bradley-Terry | leaderboard ranking · LMArena-based simulation, all phi | stability for all top-k prefixes | Improved |
| use the heuristic committee-monotone ranking (Algorithm 3) instead of the checkpoint ranking (Algorithm 2) (ablation) | pluralistic leaderboard ranking · LMArena-based simulation | stability ratio (gamma-hat) | Improved |
| increase user dispersion phi to 0.9 in simulation (ablation) | Bradley-Terry ranking · LMArena-based simulation | stability violations | Improved |
Closest papers by meaning
- A Unified Perturbation Framework for Analyzing Leaderboard Stability and Manipulation (NeurIPS 2026 · similarity 0.90)
- Inherent Trade-Offs between Diversity and Stability in Multi-Task Benchmarks (ICML 2024 · similarity 0.86)
- Prompt-to-Leaderboard: Prompt-Adaptive LLM Evaluations (ICML 2025 · similarity 0.85)
- Prediction-Powered Ranking of Large Language Models (NeurIPS 2024 · similarity 0.85)
- Aggregating Quantitative Relative Judgments: From Social Choice to Ranking Prediction (NeurIPS 2024 · similarity 0.85)
- Distributive Fairness in Large Language Models: Evaluating Alignment with Human Values (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.
