Research map / Large language models / Retrieval and knowledge
Query-Level Uncertainty in Large Language Models
Lihu Chen, Gerard de Melo, Fabian M. Suchanek, Gaël Varoquaux
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 |
|---|---|---|---|
| aggregate P(Yes) over layers and tokens with attenuated weights instead of naive averaging (ablation) | Internal Confidence · TriviaQA/SciQ/GSM8K average | AUROC: 57.4 -> 59.8 (Phi), 61.7 -> 64.2 (Llama), 63.5 -> 67.1 (Qwen) | Improved |
| aggregate P(Yes) around the decision center instead of using only the last layer and last token (ablation) | Internal Confidence · TriviaQA/SciQ/GSM8K average | AUROC: 59.8 -> 59.8 (Phi), 55.5 -> 64.2 (Llama), 60.9 -> 67.1 (Qwen) | Mixed |
| use Internal Confidence instead of Predictive Entropy on GSM8K | query-level uncertainty · GSM8K | AUROC: 63.6 vs 53.9 (Phi), 66.1 vs 65.7 (Llama), 66.4 vs 66.8 (Qwen) | Mixed |
| use Internal Confidence instead of answer-level MSP | uncertainty estimation · GSM8K (Qwen-14B) | AUROC: 68.5 -> 66.8 | Hurt |
| use query-level Internal Confidence instead of answer-level SAR | uncertainty estimation · GSM8K (Qwen-14B) | time per sample (s): 180.6 -> 0.3 | Improved |
| use query-level Internal Confidence instead of answer-level methods | uncertainty estimation · TriviaQA (Qwen-14B) | AUROC: 78.1 (EV Laplacian) -> 71.9 | Hurt |
| use Internal Confidence instead of Perplexity | uncertainty estimation · SciQ (Qwen-14B) | AUROC: 71.5 -> 62.6 | Hurt |
| use Internal Confidence instead of P(Yes) top right | query-level uncertainty · SimpleQA/MuSiQue/TruthfulQA average | AUROC: 55.4 -> 61.0 (Phi), 53.7 -> 61.0 (Llama), 52.0 -> 57.2 (Qwen) | Improved |
| use Internal Confidence instead of P(Yes) naive avg | query-level uncertainty · TriviaQA ECE (Llama-8B) | ECE: 12.6 -> 25.4 | Hurt |
| use Internal Confidence instead of P(Yes) naive avg | query-level uncertainty · SimpleQA/MuSiQue/TruthfulQA average ECE (Llama-8B) | ECE: 16.8 -> 52.4 | Hurt |
| vary locality alpha (default 1.0, locality about 0.7) (ablation) | Internal Confidence · TriviaQA/SciQ/GSM8K | AUROC | Mixed |
| add in-context examples (1 to 5 shots) (ablation) | Internal Confidence · TriviaQA/SciQ/GSM8K | AUROC | No clear effect |
Closest papers by meaning
- FLUE: Streamlined Uncertainty Estimation for Large Language Models (AAAI 2025 · similarity 0.89)
- Inference Scaling for Long-Context Retrieval Augmented Generation (ICLR 2025 · similarity 0.89)
- Distinguishing the Knowable from the Unknowable with Language Models (ICML 2024 · similarity 0.89)
- Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.88)
- AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions (NeurIPS 2025 · similarity 0.88)
- Making Retrieval-Augmented Language Models Robust to Irrelevant Context (ICLR 2024 · similarity 0.88)
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.
