Research map / Large language models / Language, safety and interpretability
Complementing Self-Consistency with Cross-Model Disagreement for Uncertainty Quantification
Kimia Hamidieh, Veronika Thost, Walter Gerych, Mikhail Yurochkin, Marzyeh Ghassemi
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 |
|---|---|---|---|
| add epistemic (cross-model disagreement) term to aleatoric self-consistency | AU (self-consistency) · 8 datasets, Mistral-7B ref | AUROC: 0.617 -> 0.718 | Improved |
| add EU to AU on HotpotQA | AU · HotpotQA (5 models avg) | AUROC: +0.15 | Improved |
| add EU to AU on pooled data | AU · All datasets pooled | AUROC: 0.707 -> 0.746 | Improved |
| add EU to AU for API models | AU · SimpleQA (GPT-4o) | AUROC: 0.59 -> 0.70 | Improved |
| add EU to AU | AU · SimpleQA (Claude 3.7 Sonnet) | AUROC: 0.53 -> 0.58 | Improved |
| add EU to AU on strongest models for XSum | AU · XSum (Llama, Qwen) | AUROC: 0.708 -> 0.693 (Llama) | Hurt |
| add EU to AU on TruthfulQA | AU · TruthfulQA | AUROC: 0.528 -> 0.579 (Mistral); 0.47 -> 0.456 (Llama) | Mixed |
| use TU for selective abstention | AU · HotpotQA, XSum | AURC: 0.304 -> 0.206 (HotpotQA) | Improved |
| increase number of auxiliary models from 2 to 5 (ablation) | TU · Mistral-7B ref | AUROC | Improved |
| increase samples per model from 5 to 50 (ablation) | TU · multiple datasets | AUROC | No clear effect |
| increase reference model size (Gemma3, Qwen2.5) (ablation) | TU and AU · TriviaQA | AUROC | Hurt |
| replace diverse auxiliary models with noise-perturbed copies of reference (ablation) | TU · multiple datasets | TU-AU AUROC gain | Hurt |
Closest papers by meaning
- Distinguishing the Knowable from the Unknowable with Language Models (ICML 2024 · similarity 0.91)
- To Believe or Not to Believe Your LLM: Iterative Prompting for Estimating Epistemic Uncertainty (NeurIPS 2024 · similarity 0.90)
- Efficient semantic uncertainty quantification in language models via diversity-steered sampling (NeurIPS 2025 · similarity 0.90)
- Decomposing Uncertainty for Large Language Models through Input Clarification Ensembling (ICML 2024 · similarity 0.90)
- Kernel Language Entropy: Fine-grained Uncertainty Quantification for LLMs from Semantic Similarities (NeurIPS 2024 · similarity 0.90)
- Conformal Prediction with Paraphrase-Aware Scoring for LLM Uncertainty Quantification (NeurIPS 2026 · similarity 0.90)
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.
