Research map / Large language models / Instruction and fine-tuning
SteerConf: Steering LLMs for Confidence Elicitation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SteerConf, a training-free black-box method that elicits verbalized confidences under five steering levels (very cautious to very confident), combines mean confidence with answer and confidence consistency, and selects answers via linear quantization to improve calibration and failure prediction.
Problem
LLMs accessed only via black-box APIs are overconfident in verbalized confidence scores, which hurts calibration and failure prediction in high-stakes use.
Approach
Explicitly prompt the LLM to be more cautious or more confident at several steering levels, then aggregate the steered answers and confidences using a steered confidence consistency measure (coefficient-of-variation based) and answer-frequency consistency, and select the final answer by linearly quantizing the calibrated confidence onto the steering index.
Key results
- With CoT, GPT-4 + SteerConf reaches average AUROC 81.1% vs 65.5% for vanilla GPT-4 CoT across seven datasets.
- On GPT-3.5 with CoT, SteerConf achieves average ECE 11.7% vs 17.3% for the best baseline (Misleading).
- Without CoT, GPT-3.5 + SteerConf average ECE drops from 41.8% to 20.7%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SteerConf to vanilla verbalized confidence elicitation | LLaMA3, without CoT · Seven benchmarks, average | ECE: 38.0 -> 24.6 | Helped |
| Add SteerConf to vanilla verbalized confidence elicitation | GPT-3.5, without CoT · Seven benchmarks, average | ECE: 41.8 -> 20.7 | Helped |
| Add SteerConf to vanilla verbalized confidence elicitation | GPT-4, without CoT · Seven benchmarks, average | AUROC: 58.2 -> 69.3 | Helped |
| Add SteerConf to vanilla verbalized confidence elicitation | LLaMA3, with CoT · Seven benchmarks, average | AUROC: 58.1 -> 77.0 | Helped |
| Add SteerConf to vanilla verbalized confidence elicitation | GPT-3.5, with CoT · Seven benchmarks, average | ECE: 29.8 -> 12.7 | Helped |
| Add SteerConf to vanilla verbalized confidence elicitation | GPT-4, with CoT · Seven benchmarks, average | AUROC: 65.5 -> 81.1 | Helped |
| Use SteerConf instead of Misleading | GPT-3.5 with CoT · Five benchmarks, average | ECE: 17.3 -> 11.7 | Helped |
| Replace confidence-based answer selection with majority voting (ablation) | SteerConf with LLaMA3 and CoT · Seven benchmarks, average | accuracy: 82.3 -> 83.6 | Helped |
| Replace confidence-based answer selection with majority voting (ablation) | SteerConf with LLaMA3 and CoT · Seven benchmarks, average | ECE: 7.3 -> 7.8 | Hurt |
| Use a very cautious prompt instead of the vanilla prompt | GPT-3.5 without CoT · GSM8K | mean confidence difference: -25.98 | Hurt |
| Use a very confident prompt instead of the vanilla prompt | GPT-3.5 without CoT · GSM8K | mean confidence difference: +5.36 | Helped |
| Use a confident prompt instead of the vanilla prompt | GPT-3.5 without CoT · Ethics | mean confidence difference: -1.79 | Hurt |
Surprising findings
- Steering toward 'confident' can deviate from the intended direction at moderate intensity, while very cautious and very confident steering behave as expected.
- Majority-vote answer selection gives slightly higher accuracy but weaker calibration than the proposed linear quantization selection.
Limitations
- Relies on manually designed steering prompts.
- Multiple prompts per question increase computational cost.
- Calibration gains come with a slight accuracy reduction relative to vanilla elicitation.
Experimental setup
Datasets: GSM8K StrategyQA Sports Understanding Date Understanding Object Counting MMLU Law MMLU Business Ethics
Benchmarks: GSM8K StrategyQA Sports Understanding Date Understanding Object Counting MMLU Law MMLU Business Ethics
Metrics: ECE AUROC PR-P PR-N
Baselines: Top-K Misleading Self-Random Vanilla verbalized confidence
Builds on: chain-of-thought Top-K verbalized confidence prompting self-consistency
Closest papers by meaning
- Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs (ICLR 2024 · similarity 0.92)
- Learning to Route LLMs with Confidence Tokens (ICML 2025 · similarity 0.91)
- ConfTuner: Training Large Language Models to Express Their Confidence Verbally (NeurIPS 2025 · similarity 0.90)
- Multicalibration for Confidence Scoring in LLMs (ICML 2024 · similarity 0.90)
- Your Pre-trained LLM is Secretly an Unsupervised Confidence Calibrator (NeurIPS 2025 · similarity 0.90)
- A Semantic-Sampling Framework for Evaluating Calibration in Open-Ended Question Answering (NeurIPS 2026 · similarity 0.89)
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.