Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

LACUNA: A Testbed for Evaluating Localization Precision for LLM Unlearning

Source

arXiv

arxiv.org/abs/2607.02513v1

Read the source

Source

Plain-Language Brief

The researchers introduce a new testbed called LACUNA for evaluating the precision of post hoc removal methods (unlearning) for LLMs. The testbed focuses on injecting personal identifiable information into specific parameters of OLMo-based models and provides ground-truth parameter-level localization. Results show that current SOTA unlearning methods have imprecise results and are vulnerable to resurfacing attacks, but a simple gradient-based unlearning method can achieve strong erasure and resistance to these attacks when localization is successful.

What the Document Says

  • The researchers introduce LACUNA, a testbed for evaluating the precision of post hoc removal methods (unlearning) for LLMs.
  • LACUNA injects personal identifiable information into specific parameters of OLMo-based models via masked continual pretraining.
  • It enables direct evaluation of whether unlearning targets the weights responsible for knowledge storage.
  • The researchers use LACUNA to benchmark current SOTA unlearning methods and find that they are imprecise and vulnerable to resurfacing attacks.
  • Even a simple gradient-based unlearning method can achieve strong erasure and robustness to resurfacing attacks when localization is successful.
  • The researchers release LACUNA to drive further advances in robust, localization-based unlearning.

Key Concepts Mentioned

  • LLMs
  • unlearning
  • personal identifiable information (PII)
  • OLMo-based models
  • LACUNA testbed
  • ground-truth parameter-level localization
  • resurfacing attacks

Claims Made by the Document

  • Existing SOTA unlearning methods are imprecise and vulnerable to resurfacing attacks.
  • A simple gradient-based unlearning method can achieve strong erasure and robustness to resurfacing attacks when localization is successful.

Evidence Used

  • Results from using LACUNA to benchmark current SOTA unlearning methods

Methods / Approach

  • masked continual pretraining
  • unlearning
  • gradient-based unlearning

Limitations Stated by the Authors

  • The testbed only focuses on OLMo-based models.

Limitations Noted by the Agent

Agent interpretation — not from the source.

  • The study does not evaluate the efficiency or scalability of LACUNA.
  • The study assumes that the success of localization implies successful erasure, which may not always be the case.

Open Questions

  • Can the results be extended to other types of models?
  • What is the efficiency and scalability of LACUNA?

Notes for Future Q&A

May include interpretation. Labeled where it does.

  • The authors focus on evaluating the precision of unlearning methods, but it's important to consider other aspects such as efficiency and scalability.

Raw Abstract / Excerpt

LLMs memorize sensitive training data, including personally identifiable information (PII), creating a pressing need for reliable post hoc removal methods. Unlearning has emerged as a promising solution, with state-of-the-art(SOTA) methods often following a localize-first, unlearn-second paradigm that targets specific model parameters. However, existing benchmarks evaluate unlearning solely at the output level, leaving open the question of whether unlearning truly erases knowledge from a model's parameters or merely obfuscates it, a concern reinforced by the success of resurfacing attacks. To bridge this gap, we introduce LACUNA: the first unlearning testbed with ground-truth parameter-level localization. LACUNA injects PII of synthetic individuals into predefined parameters of 1B and 7B OLMo-based models via masked continual pretraining, enabling direct evaluation of whether unlearning targets the weights responsible for knowledge storage. We use LACUNA to benchmark current SOTA unlearning methods and find that, despite strong output-level performance, existing methods are highly imprecise and susceptible to resurfacing attacks. We further show that when localization is successful, even a simple gradient-based unlearning method achieves strong erasure and robustness to resurfacing attacks, highlighting the importance of precise unlearning. We release LACUNA to complement behavioral evaluations and drive further advances in robust, localization-based unlearning.