Confidence Elicitation

Confidence elicitation in large language models (LLMs) focuses on developing methods to accurately reflect a model's certainty in its predictions, aligning its reported confidence with the actual probability of correctness. Current research emphasizes improving calibration—the agreement between predicted confidence and accuracy—through various techniques, including prompt engineering to elicit verbalized confidence, analyzing consistency across multiple model outputs, and exploring the impact of model size and fine-tuning methods. This work is crucial for building trustworthy AI systems, particularly in high-stakes applications where understanding and managing uncertainty is paramount, such as misinformation detection and code summarization.

Papers

October 17, 2024

Atomic Calibration of LLMs in Long-Form Generations
Caiqi Zhang, Ruihan Yang, Zhisong Zhang, Xinting Huang, Sen Yang, Dong Yu, Nigel Collier
Large Language Model Text Generation Calibration Performance LLM Fine Tuning Confidence Calibration Long Form Generation Confidence Elicitation

June 21, 2024

Benchmarking Uncertainty Quantification Methods for Large Language Models with LM-Polygraph
Roman Vashurin, Ekaterina Fadeeva, Artem Vazhentsev, Lyudmila Rvanova, Akim Tsvigun, Daniil Vasilev, Rui Xing, Abdelrahman Boda Sadallah, Kirill Grishchenkov, Sergey Petrakov, Alexander Panchenko, Timothy Baldwin, Preslav Nakov, Maxim Panov, Artem Shelmanov
Large Language Model Uncertainty Quantification Text Generation Task Lie Detection Uncertainty Quantification Method Confidence Elicitation

April 30, 2024

Enhancing Trust in LLM-Generated Code Summaries with Calibrated Confidence Scores
Yuvraj Virk, Premkumar Devanbu, Toufique Ahmed
Code Summarization Building Trust Human Written Summary LLM Generated Code Generated Summary Program Comprehension Confidence Elicitation

February 9, 2024

Calibrating Long-form Generations from Large Language Models
Yukun Huang, Yixin Liu, Raghuveer Thirukovalluru, Arman Cohan, Bhuwan Dhingra
LLM Fine Tuning Confidence Score Long Form Generation Confidence Elicitation

January 13, 2024

Combining Confidence Elicitation and Sample-based Methods for Uncertainty Quantification in Misinformation Mitigation
Mauricio Rivera, Jean-François Godbout, Reihaneh Rabbany, Kellin Pelrine
Uncertainty Quantification Misinformation Mitigation Confidence Elicitation Sample Based

November 16, 2023

Examining LLMs' Uncertainty Expression Towards Questions Outside Parametric Knowledge
Genglin Liu, Xingyao Wang, Lifan Yuan, Yangyi Chen, Hao Peng
Large Language Model Yes No Question Parametric Knowledge Uncertainty Expression Confidence Elicitation

June 22, 2023

Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs
Miao Xiong, Zhiyuan Hu, Xinyang Lu, Yifei Li, Jie Fu, Junxian He, Bryan Hooi
Large Language Model Medical LLM High Uncertainty Anticipation Empirical Evaluation Confidence Calibration Black Box LLM LLM Uncertainty Confidence Elicitation

May 24, 2023

Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback
Katherine Tian, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, Christopher D. Manning
Large Language Model Language Model Human Feedback Calibration Performance General Strategy Fine Tuned Language Model Confidence Score Calibrated Confidence Low Confidence Confidence Elicitation

Confidence Elicitation

Papers

Atomic Calibration of LLMs in Long-Form Generations

Benchmarking Uncertainty Quantification Methods for Large Language Models with LM-Polygraph

Enhancing Trust in LLM-Generated Code Summaries with Calibrated Confidence Scores

Calibrating Long-form Generations from Large Language Models

Combining Confidence Elicitation and Sample-based Methods for Uncertainty Quantification in Misinformation Mitigation

Examining LLMs' Uncertainty Expression Towards Questions Outside Parametric Knowledge

Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs

Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback