Harmonizing Multi-Objective LLM Unlearning via Unified Domain Representation and Bidirectional Logit Distillation

Researchers propose a multi-objective unlearning framework for Large Language Models that simultaneously removes hazardous information, preserves general utility, avoids over-refusal, and resists adversarial attacks. The method uses unified domain representation and bidirectional logit distillation to harmonize competing optimization goals, achieving state-of-the-art performance across diverse unlearning requirements.

LLM unlearning has emerged as a critical challenge in AI safety, addressing the need to remove sensitive, proprietary, or harmful information from trained models without degrading their overall performance. This research tackles a previously underexplored problem: existing unlearning methods optimize for narrow objectives like efficacy and utility preservation while neglecting robustness against adversarial attacks and boundary behaviors—scenarios where models incorrectly refuse benign requests adjacent to unlearned concepts.

The framework's innovation lies in its co-design approach that treats seemingly conflicting objectives as cooperative optimization tasks. By standardizing diverse training corpora into a unified representation, the method reduces domain gaps that traditionally cause task interference. The bidirectional distillation mechanism simultaneously extracts desired behaviors from a teacher model while suppressing undesirable outputs in the student, creating a more balanced unlearning process.

For the AI industry, this addresses a growing regulatory and practical concern: enterprises deploying LLMs must ensure they can remove proprietary data, personal information, or harmful knowledge after deployment. Companies developing language models face pressure from privacy regulations and competitive concerns requiring post-training model modifications. Current methods' failure to maintain robustness against adversarial probing represents a genuine security vulnerability.

Developers implementing unlearning will benefit from a more comprehensive framework preventing the common tradeoff where removing problematic knowledge paradoxically makes models less usable. As LLM deployment becomes more scrutinized by regulators, particularly around copyright and privacy, robust unlearning methods become infrastructure-level requirements rather than optional features. The methodology's focus on adversarial robustness particularly matters as bad actors increasingly test model boundaries through jailbreak techniques.

• →Multi-objective LLM unlearning framework balances knowledge removal, utility preservation, boundary behavior control, and adversarial robustness simultaneously.
• →Unified domain representation reduces task interference by treating conflicting optimization objectives as cooperative rather than competing.
• →Bidirectional logit distillation transfers desired behaviors from teacher models while suppressing unlearned information in student models.
• →Framework addresses critical gap in existing methods that overlook adversarial robustness and over-refusal behaviors in practical deployments.
• →State-of-the-art results suggest this approach enables reliable post-deployment model modifications crucial for regulatory compliance and data protection.