SEPS: A Separability Measure for Robust Unlearning in LLMs

TL;DR

This paper introduces SEPS, a new evaluation framework for unlearning in LLMs that assesses the model's ability to forget and retain information simultaneously within mixed prompts, addressing limitations of existing metrics.

Contribution

The paper proposes SEPS and Mixed Prompt unlearning, a novel training strategy that enhances robustness of unlearning in LLMs in complex, mixed-query scenarios.

Findings

SEPS effectively measures mixed-query unlearning performance.

Mixed Prompt unlearning improves robustness in multi-query settings.

Existing methods often overfit or erase too much information.

Abstract

Machine unlearning aims to selectively remove targeted knowledge from Large Language Models (LLMs), ensuring they forget specified content while retaining essential information. Existing unlearning metrics assess whether a model correctly answers retain queries and rejects forget queries, but they fail to capture real-world scenarios where forget queries rarely appear in isolation. In fact, forget and retain queries often coexist within the same prompt, making mixed-query evaluation crucial. We introduce SEPS, an evaluation framework that explicitly measures a model's ability to both forget and retain information within a single prompt. Through extensive experiments across three benchmarks, we identify two key failure modes in existing unlearning methods: (1) untargeted unlearning indiscriminately erases both forget and retain content once a forget query appears, and (2) targeted unlearning overfits to single-query scenarios, leading to catastrophic failures when handling multiple queries. To address these issues, we propose Mixed Prompt (MP) unlearning, a strategy that integrates both forget and retain queries into a unified training objective. Our approach significantly improves unlearning effectiveness, demonstrating robustness even in complex settings with up to eight mixed forget and retain queries in a single prompt.