Research map / Large language models / Language, safety and interpretability
Annotation-Efficient Honesty Alignment via Confidence Elicitation and Calibration
Shiyu Ni, Keping Bi, Jiafeng Guo, Minghao Tang, Jingtong wu, Zengxin Han, Xueqi Cheng
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 confidence elicitation stage before calibration (1k annotations) (ablation) | Cal-Only · HonestyBench in-domain avg (Qwen2.5-7B) | AUROC: 73.41 -> 84.36 | Improved |
| add confidence elicitation stage before calibration (1k annotations) (ablation) | Cal-Only · HonestyBench OOD avg (Qwen2.5-7B) | AUROC: 77.32 -> 84.47 | Improved |
| add confidence elicitation before calibration with all 560k annotations (ablation) | Cal-Only (560k) · HonestyBench in-domain avg (Qwen2.5-7B) | AUROC: 86.20 -> 86.49 | No clear effect |
| add calibration stage with 1k annotations after elicitation (ablation) | Eli-Only · HonestyBench in-domain avg (Qwen2.5-7B) | AUROC: 71.19 -> 84.36 | Improved |
| replace Consis-Sem inference-time sampling with trained Eli-Only one-shot confidence | Consis-Sem · HonestyBench in-domain avg (Qwen2.5-7B) | AUROC: 73.62 -> 71.19 | No clear effect |
| add elicitation then 1k calibration at 32B scale | Cal-Only (1k) · HonestyBench in-domain avg (Qwen2.5-32B) | AUROC: 79.31 -> 84.39 | Improved |
| add elicitation then 1k calibration at 32B scale | Cal-Only (1k) · HonestyBench OOD avg (Qwen2.5-32B) | AUROC: 79.62 -> 84.01 | Improved |
| add EliCal 1k calibration versus best training-free Consis-Sem | Consis-Sem · HonestyBench in-domain avg (Qwen2.5-7B) | AUROC: 73.62 -> 84.36 | Improved |
| add temperature scaling (Thermometer) instead of learned elicitation | EliCal · HonestyBench in-domain avg (Qwen2.5-7B) | AUROC: 84.36 -> 59.48 | Hurt |
| increase self-consistency samples k from 2 to 20 for Consis-Sem (ablation) | Consis-Sem · HonestyBench in-domain avg (Qwen2.5-7B) | AUROC: 68.04 -> 73.62 | Improved |
| increase self-consistency samples k from 2 to 20 for EliCal (ablation) | EliCal (1k) · HonestyBench in-domain avg (Qwen2.5-7B) | AUROC: 84.41 -> 84.36 | No clear effect |
| increase elicitation training data size (ablation) | Eli-Only · HonestyBench in-domain avg (Qwen2.5-7B) | AUROC: approaches Consis-Sem 73.620 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Thermometer — Thermometer: Towards Universal Calibration for Large Language Models (ICML 2024) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Alignment for Honesty (NeurIPS 2024 · similarity 0.91)
- Your Pre-trained LLM is Secretly an Unsupervised Confidence Calibrator (NeurIPS 2025 · similarity 0.89)
- HonestLLM: Toward an Honest and Helpful Large Language Model (NeurIPS 2024 · similarity 0.88)
- Aligning Large Language Models for Faithful Integrity Against Opposing Argument (AAAI 2025 · similarity 0.88)
- Calibrating Large Language Models with Sample Consistency (AAAI 2025 · similarity 0.88)
- MoHoBench: Assessing Honesty of Multimodal Large Language Models via Unanswerable Visual Questions (AAAI 2026 · 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.
