Unlearners Can Lie: Evaluating and Improving Honesty in LLM Unlearning

2026-05-09

AI Summary

Existing unlearning methods for large language models often induce hallucinations or behavioral inconsistencies when removing harmful data, lacking formal guarantees of "honesty." This work introduces the first formal definition of honesty in the context of model unlearning and establishes a multidimensional evaluation framework encompassing utility, honesty on retained knowledge, unlearning efficacy, and rejection stability. Building upon this foundation, the authors propose ReVa, a representation alignment method that combines feature randomization with fine-tuning to encourage models to honestly acknowledge uncertainty after unlearning. Experimental results demonstrate that ReVa achieves nearly twice the refusal rate of the next-best baseline on unlearned-set question-answering tasks after two rounds of interaction, while simultaneously yielding significant improvements in honesty over retained knowledge.

Abstract

Unlearning in large language models (LLMs) aims to remove harmful training data while preserving overall utility. However, we find that existing methods often hallucinate, generate abnormal token sequences, or behave inconsistently, raising safety and trust concerns. According to prior literature on LLM honesty, such behaviors are often associated with dishonesty. This motivates us to investigate the notion of honesty in the context of model unlearning. We propose a formal definition of unlearning honesty, which includes: (1) preserving both utility and honesty on retained knowledge, and (2) ensuring effective forgetting while encouraging the model to acknowledge its limitations and respond consistently to questions related to forgotten knowledge. To systematically evaluate the honesty of unlearning, we introduce a suite of metrics that cover utility, honesty on the retained set, effectiveness of forgetting, rejection rate and refusal stability in Q&A and MCQ settings. Evaluating 9 methods across 3 mainstream families shows that all current methods fail to meet these standards. After experimental and theoretical analyses, we present ReVa, a representation-alignment procedure that fine-tunes feature-randomized unlearned models to better acknowledge forgotten knowledge. On Q&A tasks from the forget set, ReVa achieves the highest rejection rate after two rounds of interaction, nearly doubling the performance of the second-best method. Remarkably, It also improves honesty on the retained set. We release our data and code at https://github.com/renjiegu.

Problem

• unlearning
• honesty
• large language models
• forgetting
• trustworthiness

Innovation

• LLM unlearning
• honesty
• representation alignment
• forgetting evaluation
• ReVa