Research map / Large language models / Language, safety and interpretability
SCOPE: Selective Conformal Optimized Pairwise LLM Judging
Sher Badshah, Ali Emami, Hassan Sajjad
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 |
|---|---|---|---|
| use BPE (average forward/reverse probabilities, binary entropy) instead of Swap-and-Aggregate discrete voting (ablation) | bidirectional LLM judge · MT-Bench (Qwen2.5-14B) | AUROC: 0.679 -> 0.777 | Improved |
| use BPE instead of Swap-and-Aggregate (ablation) | bidirectional LLM judge · RewardBench (Qwen2.5-7B) | AUROC: 0.673 -> 0.744 | Improved |
| use BPE instead of Predictive Probability | LLM judge · MT-Bench (Qwen2.5-7B) | ECE: 0.239 -> 0.143 | Improved |
| use BPE instead of Verbalized Confidence | LLM judge · MT-Bench (Qwen2.5-7B) | AUROC: 0.497 -> 0.685 | Improved |
| use BPE instead of Verbalized Confidence | LLM judge · MT-Bench (Qwen2.5-14B) | ECE: 0.114 -> 0.174 | Hurt |
| use BPE instead of Simulated Annotators | LLM judge · MT-Bench (Qwen2.5-7B) | AUROC: 0.59 -> 0.69 | Improved |
| use BPE instead of Simulated Annotators | LLM judge · MT-Bench (Qwen2.5-14B) | AUROC: 0.60 -> 0.78 | Improved |
| use BPE instead of Simulated Annotators | LLM judge · RewardBench (Qwen2.5-7B) | ECE: 0.103 -> 0.104 | No clear effect |
| use SCOPE conformal thresholding instead of naive calibration | selective LLM judging · MT-Bench (Qwen-7B, alpha=0.10) | empirical risk: 0.116 -> 0.097 | Improved |
| use SCOPE instead of naive calibration | selective LLM judging · MT-Bench (Qwen-14B, alpha=0.10) | coverage: 0.566 -> 0.504 | Hurt |
| use SCOPE instead of naive calibration | selective LLM judging · MT-Bench (Qwen-7B, alpha=0.10) | coverage: 0.102 -> 0.246 | Improved |
| use SCOPE instead of heuristic thresholding | selective LLM judging · MT-Bench (Qwen-7B, alpha=0.10) | empirical risk: 0.251 -> 0.097 | Improved |
Closest papers by meaning
- From Uncertain Judgments to Calibrated Rankings: Conformal Elo Estimation for LLM Evaluation (NeurIPS 2026 · similarity 0.89)
- Bridging Human and LLM Judgments: Understanding and Narrowing the Gap (NeurIPS 2025 · similarity 0.88)
- Style Outweighs Substance: Failure Modes of LLM Judges in Alignment Benchmarking (ICLR 2025 · similarity 0.88)
- RocketEval: Efficient automated LLM evaluation via grading checklist (ICLR 2025 · similarity 0.87)
- BayesJudge: Uncertainty-Aware Bayesian Meta-Evaluation of Human and LLM Judgments (NeurIPS 2026 · similarity 0.87)
- MLLM-as-a-Judge: Assessing Multimodal LLM-as-a-Judge with Vision-Language Benchmark (ICML 2024 · similarity 0.87)
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.
