A Calibrated Reflection Approach for Enhancing Confidence Estimation in LLMs

Umesh Bodhwani; Yuan Ling; Shujing Dong; Yarong Feng; Hongfei Li; Ayush Goyal

cs.CLSep 3, 2026

A Calibrated Reflection Approach for Enhancing Confidence Estimation in LLMs

Authors: Umesh Bodhwani, Yuan Ling, Shujing Dong, Yarong Feng, Hongfei Li, Ayush Goyal

Organizations: Amazon.com Seattle, Washington, USA

Abstract

A critical challenge in deploying Large Language Models (LLMs) is developing reliable mechanisms to estimate their confidence, enabling systems to determine when to trust model outputs versus seek human intervention. We present a Calibrated Reflection approach for enhancing confidence estimation in LLMs, a framework that combines structured reasoning with distance-aware calibration technique. Our approach introduces three key innovations: (1) a Maximum Confidence Selection (MCS) method that comprehensively evaluates confidence across all possible labels, (2) a reflection-based prompting mechanism that enhances reasoning reliability, and (3) a distance-aware calibration technique that accounts for ordinal relationships between labels. We evaluate our framework on diverse datasets, including HelpSteer2, Llama T-REx, and a proprietary conversational dataset, demonstrating its effectiveness across both conversational and fact-based classification tasks. This work contributes to the broader goal of developing reliable and well-calibrated confidence estimation methods for LLMs, enabling informed decisions about model trust and human judgement.

Figures & tables

Figure 1: Calibrated Reflection Workflow: Reflection Prompting generates an initial confidence score based on rubrics, reflects on its reasoning, and updates the scores. Distance-Aware Calibration adjusts the scores based on ordinal distances between labels. The output includes the predicted score and the calibrated confidence score

Panel of Figure 2 — Figure 2: MCS-R prompt template for conversation quality assessment. The model follows a structured evaluation process (initial feedback → reflection → final feedback → probability) for each potential label

Figure 3: Performance comparison on the HelpSteer2 dataset across different evaluation metrics. Results are reported for VC, MCS, MCS-C, MCS-R, and MCS-RC. Each bar group represents performance on five conversational dimensions and their average, highlighting the impact of advanced prompting techniques and calibration on confidence estimation.

Method	Brier	ECE	AUROC	AUPRC
Log Prob	0.117	0.101	0.623	0.895
CoT	0.098	0.093	0.667	0.908
Self-Consistency	0.095	0.086	0.672	0.913
Debate	0.104	0.078	0.692	0.927
Reflection	0.110	0.080	0.687	0.918

Table 1: Average Scores for Confidence Score Estimation Experiments on Verbalized Method

Method	Brier	ECE	AUROC	AUPRC
MCS	.138	.169	.591	.899
+Reflection	.256	.301	.697	.925
+Debate	.256	.322	.700	.924

Table 2: Comparison of Results for MCS with Reflection and Debate Prompts

Method	Brier Score	ECE	AUROC	AUPRC
MCS	0.13799	0.16877	0.59146	0.89905
MCS+Debate (Ours)	0.25574	0.32192	0.69999	0.9243
MCS+Reflection (Ours)	0.25617	0.30092	0.6969	0.92534
MCS+Debate+Calibration (Proposed)	0.1282	0.19243	0.73752	0.93334
MCS+Reflection+Calibration (Proposed Best)	0.12502	0.17472	0.73994	0.93516

Table 3: Final Evaluation Scores Combining MCS + Enhanced Prompts + Calibration Technique

Figure 4: Confidence score distribution across different methods: VC, MCS, MCS, MCS-R, and MCS-RC. The histogram illustrates the frequency of confidence scores, while the blue line represents the kernel density estimation. The progression from VC to MCS-RC demonstrates improved confidence score distribution, with reduced overconfidence and better alignment with model uncertainty

Method	AUPRC
Verbalized Confidence	0.700
Log Prob	0.709
Trained Probe (SOTA)	0.910
Verbalized Debate (Proposed)	0.887
Verbalized Reflection (Proposed)	0.890

Table 4: Performance Comparison on T-REx Dataset

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Confidence Estimation in Language Models

Confidence estimation in language models measures whether a model’s uncertainty signals predict answer correctness and support reliable decisions.

See all 247 papers

Instinct vs. Reflection: Unifying Token and Verbalized Confidence in Multimodal Large Models
Apr 19, 2026Yunkai Dang, Yifan Jiang, Yizhu Jiang +3Selective Prediction Confidence Estimation in Language Models
School of Artificial Intelligence, Nanjing University.
ProbPlug: A Plugin Uncertainty Network for Reliable Confidence in LLM Binary Classification
Sep 9, 2026Jianzong Wang, Chuhang Liu, Botao Zhao +6Confidence Estimation in Language Models Language Model Calibration
Ping An Technology (Shenzhen) Co., Ltd., Shenzhen, China · Tsinghua Shenzhen International Graduate School, Shenzhen, China · Hubei University of Technology, Hubei, China
Confidence Before Answering: A Paradigm Shift for Efficient LLM Uncertainty Estimation
Mar 6, 2026Changcheng Li, Jiancan Wu, Hengheng Zhang +5Confidence Estimation in Language Models LLM Uncertainty Estimation
University of Science and Technology of China · Huawei Inc.

Apr 19, 2026J/K move · Enter open · S save

Instinct vs. Reflection: Unifying Token and Verbalized Confidence in Multimodal Large Models

Yunkai Dang, Yifan Jiang, Yizhu Jiang, Anqi Chen+2

School of Artificial Intelligence, Nanjing University.

Selective Prediction Confidence Estimation in Language Models Multimodal Large Language Models Language Model Calibration LLM Reliability LLM Uncertainty Estimation+1

Open paper AskPDF arXiv

Multimodal Large Language Models (MLLMs) have demonstrated exceptional capabilities in various perception and reasoning tasks. Despite this success, ensuring their reliability in practical deployment necessitates robust confidence estimation. Prior works have predominantly focused on text-only LLMs, often relying on computationally expensive self-consistency sampling. In this paper, we extend this to multimodal settings and conduct a comprehensive evaluation of MLLMs' response confidence estimation. Our analysis reveals a significant instinct-reflection misalignment: the model's implicit token-level support frequently diverges from its verbal self-assessment confidence. To address this misalignment, we propose a monotone confidence fusion framework to merge dual-channel signals and cross-channel consistency to estimate correctness. Subsequently, an order-preserving mean alignment step is applied to correct global bias, which improves calibration while preserving the risk-coverage trade-off for selective prediction. Experiments on diverse open-source and closed-source MLLMs show that our method consistently yields more reliable confidence estimates and improves both calibration and failure prediction. Code will be available at https://github.com/Yunkaidang/Instinct-vs.-Reflection.

Abstract

Confidence Estimation in Language Models

Explore similar work

Instinct vs. Reflection: Unifying Token and Verbalized Confidence in Multimodal Large Models

ProbPlug: A Plugin Uncertainty Network for Reliable Confidence in LLM Binary Classification

Confidence Before Answering: A Paradigm Shift for Efficient LLM Uncertainty Estimation