arXiv:2511.04934v3 Announce Type: replace Abstract: Unlearning in large language models (LLMs) is critical for regulatory compliance and for building ethical generative AI systems that avoid producing private, toxic, illegal, or copyrighted content. Despite rapid progress, in this work, we show that \textit{almost all} existing unlearning methods fail to achieve true forgetting in practice. Specifically, while evaluations of these `unlearned' models under deterministic (greedy) decoding often suggest successful knowledge removal using standard benchmarks, we show that sensitive information rel
The paper is published amidst increasing scrutiny on AI ethics, regulatory compliance, and the practical effectiveness of unlearning mechanisms in LLMs.
This finding reveals a fundamental limitation of current LLM unlearning methods, challenging the efficacy of compliance strategies and safety measures reliant on them.
The assumption that unlearning methods effectively remove sensitive information from LLMs is now heavily questioned, especially under probabilistic decoding scenarios.
- · AI safety researchers
- · Adversarial AI developers
- · Regulatory bodies (in clarifying requirements)
- · LLM developers (relying on current unlearning methods)
- · Companies using LLMs with sensitive data
- · Ethical AI frameworks (if not updated)
Current LLM unlearning methods are demonstrated to be insufficient for true forgetting, particularly with probabilistic decoding.
This will necessitate a re-evaluation and significant innovation in LLM unlearning techniques to meet regulatory and ethical demands.
Increased legal and reputational risks for companies deploying LLMs, potentially leading to new compliance standards for 'true' unlearning or restrictions on LLM deployment in sensitive contexts.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.LG