Research map / Large language models / Language, safety and interpretability
Can LLMs Detect Their Confabulations? Estimating Reliability in Uncertainty-Aware Language Models
Tianyi Zhou, Johanne Medina, Sanjay Chawla
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 LogProb with uncertainty-guided Probe(AVG) | Fanar1-9b · TruthfulQA | AUROC: 0.597 -> 0.734 | Improved |
| replace LogProb with uncertainty-guided Probe(AVG) | Fanar1-9b · Math | AUROC: 0.757 -> 0.833 | Improved |
| replace LogProb with Probe(AVG) | Qwen2.5-7B · TruthfulQA | AUROC: 0.591 -> 0.761 | Improved |
| replace LogProb with Probe(AVG) | Qwen2.5-7B · Math | AUROC: 0.635 -> 0.699 | Improved |
| replace LogProb with Probe(AVG) | Gemma3-12B · TruthfulQA | AUROC: 0.545 -> 0.733 | Improved |
| replace LogProb with Probe(AVG) | Gemma3-12B · TriviaQA | AUROC: 0.806 -> 0.818 | Improved |
| replace LogProb with Probe(AVG) | Fanar1-9b · TriviaQA | AUROC: 0.774 -> 0.765 | Hurt |
| replace LogTokU with Probe(AVG) | Gemma3-12B · Math | AUROC: 0.791 -> 0.786 | Hurt |
| average hidden states of selected tokens (Probe(AVG)) instead of EOS token only (ablation) | Fanar1-9b · TruthfulQA | AUROC: 0.706 -> 0.734 | Improved |
| average hidden states (Probe(AVG)) instead of EOS token only (ablation) | Qwen2.5-7B · TriviaQA | AUROC: 0.812 -> 0.786 | Hurt |
| average hidden states (Probe(AVG)) instead of single highest/lowest EU token (Probe(EU)) (ablation) | Gemma3-12B · Math | AUROC: 0.669 -> 0.786 | Improved |
| average hidden states (Probe(AVG)) instead of single EU token (Probe(EU)) (ablation) | Fanar1-9b · TriviaQA | AUROC: 0.751 -> 0.765 | Improved |
Closest papers by meaning
- An Empirical Analysis of Uncertainty in Large Language Model Evaluations (ICLR 2025 · similarity 0.90)
- Distinguishing the Knowable from the Unknowable with Language Models (ICML 2024 · similarity 0.90)
- Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs (ICLR 2024 · similarity 0.90)
- To Believe or Not to Believe Your LLM: Iterative Prompting for Estimating Epistemic Uncertainty (NeurIPS 2024 · similarity 0.90)
- Position: Uncertainty Quantification Needs Reassessment for Large Language Model Agents (ICML 2025 · similarity 0.90)
- Large Language Models Must Be Taught to Know What They Don’t Know (NeurIPS 2024 · 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.
