Before Forgetting, Learn to Remember: Revisiting Foundational Learning Failures in LVLM Unlearning Benchmarks

Abstract

While Large Vision-Language Models (LVLMs) offer powerful capabilities, they pose privacy risks by unintentionally memorizing sensitive personal information. Current unlearning benchmarks attempt to mitigate this using fictitious identities but overlook a critical stage 1 failure: models fail to effectively memorize target information initially, rendering subsequent unlearning evaluations unreliable. Diagnosing under-memorization and the multi-hop curse as root causes, we introduce ReMem, a Reliable Multi-hop and Multi-image Memorization Benchmark. ReMem ensures robust foundational learning through principled data scaling, reasoning-aware QA pairs, and diverse visual contexts. Additionally, we propose a novel Exposure metric to quantify the depth of information erasure from the model's internal probability distribution. Extensive experiments demonstrate that ReMem provides a rigorous and trustworthy framework for diagnosing both learning and unlearning behaviors in LVLMs.

Quick Access

• Huggingface Dataset: herbwood27/Remem
• Paper: Accepted to Findings of ACL 2026 (arXiv:2605.03759)

Dataset Structure

Available Splits

This dataset consists of 7 splits designed to evaluate both the memorization (Stage 1) and the unlearning (Stage 2) phases:

Split	Examples	Description
finetune	2,000	The full set for foundational memorization of target identities.
forget1	100	Target subset for unlearning (5% of the train set).
forget2	200	Target subset for unlearning (10% of the train set).
forget3	300	Target subset for unlearning (15% of the train set).
forget4	400	Target subset for unlearning (20% of the train set).
retain	560	Evaluation set for utility preservation of non-target information.
test	560	Held-out set for assessing unlearning robustness across diverse contexts.

Data Fields

• image: The personal identification image (PIL.Image).
• question: Reasoning-aware VQA question regarding sensitive info.
• answer: Ground-truth answer containing personal information.
• keywords: Key entity or value for Exact Match (EM) evaluation.
• question_type: Reasoning depth (e.g., 1-hop).
• qa_category: Category of the information (e.g., personal_information).
• attribute: Specific PII type (e.g., email, date_of_birth).
• cloze_prompt: Prompt for measuring internal probability/exposure.
• image_path: Original image file mapping.

Quick Start

from datasets import load_dataset

# Load the full memorization set
ds_full = load_dataset("herbwood27/Remem", split="finetune")

# Load a specific forget split for unlearning
ds_forget = load_dataset("herbwood27/Remem", split="forget1")

Citation

This paper has been accepted to Findings of ACL 2026. If you find our work useful in your research, please consider citing our preprint:

@article{kwon2026before,
  title={Before Forgetting, Learn to Remember: Revisiting Foundational Learning Failures in LVLM Unlearning Benchmarks},
  author={JuneHyoung Kwon and MiHyeon Kim and Eunju Lee and JungMin Yun and Byeonggeuk Lim and YoungBin Kim},
  journal={arXiv preprint arXiv:2605.03759},
  year={2026}
}