The field of large language models (LLMs) is moving towards developing more reliable and trustworthy models. Recent research has focused on improving the uncertainty awareness of LLMs, enabling them to better estimate their confidence in their predictions and generate more accurate results. This includes work on uncertainty-aware frameworks, reliability estimation, and human-alignment of inference-time uncertainty. Notable papers in this area have proposed innovative methods for detecting model confabulations, jointly generating and attributing answers, and introducing conformal p-value frameworks for multiple-choice question answering tasks. Overall, these developments have the potential to significantly improve the performance and trustworthiness of LLMs in various applications. Noteworthy papers include: Can LLMs Detect Their Confabulations, which proposes a reliability estimation method that leverages token-level uncertainty to guide the aggregation of internal model representations. Jointly Generating and Attributing Answers using Logits of Document-Identifier Tokens, which introduces a method that jointly generates and faithfully attributes answers in RAG by leveraging specific token logits during generation.
Advances in Uncertainty-Aware Large Language Models
Sources
Interpreting Fedspeak with Confidence: A LLM-Based Uncertainty-Aware Framework Guided by Monetary Policy Transmission Paths
Can LLMs Detect Their Confabulations? Estimating Reliability in Uncertainty-Aware Language Models
Human-Alignment and Calibration of Inference-Time Uncertainty in Large Language Models
Jointly Generating and Attributing Answers using Logits of Document-Identifier Tokens
Prospect Theory Fails for LLMs: Revealing Instability of Decision-Making under Epistemic Uncertainty
Conformal P-Value in Multiple-Choice Question Answering Tasks with Provable Risk Control
Reflect then Learn: Active Prompting for Information Extraction Guided by Introspective Confusion