Vision-Language Models (VLMs) can memorize sensitive or harmful information during pretraining. While machine unlearning aims to remove this data, most existing research focuses on forgetting within a single modality (text-only or image-only). The paper "Does Forgetting Transfer Across Modalities? A Real-World Benchmark for Cross-Modal Knowledge Unlearning Evaluation" investigates whether unlearning a fact in one modality—such as text—successfully removes that knowledge when the model is queried through another modality, such as an image.
The UNLINK-VL Benchmark
To evaluate cross-modal knowledge transfer, the authors introduced UNLINK-VL, a benchmark consisting of visually identifiable real-world entities. The researchers linked these entities to factual knowledge from Wikidata, creating both text-only and text-image versions of the same information. The benchmark evaluates four key areas:
Direct forgetting: The ability to remove specific target facts.
Relational propagation: Whether forgetting extends to related two-hop facts.
Knowledge preservation: Ensuring unrelated, non-target information remains intact.
Robustness: Whether the model remains "forgetful" when faced with semantically equivalent queries or paraphrases.
Testing Cross-Modal Transfer
The researchers tested four unlearning methods—GA, DPO, NPO, and RT—across four experimental settings: text-to-text (T→T), image-to-image (M→M), text-to-image (T→M), and image-to-text (M→T). By keeping the target knowledge constant while varying the input modality, the authors sought to determine if unlearning is consistent across different ways of accessing information.
Key Findings on Asymmetry
The experiments revealed a significant asymmetry in how unlearning transfers across modalities. Multimodal unlearning (using text-image inputs) proved effective even when the model was evaluated using text-only queries. Conversely, text-only unlearning transferred poorly to visual and cross-modal scenarios. This suggests that a model might appear to have forgotten a fact when tested with text, but could still retrieve that same information when presented with a corresponding image. The study also found that these unlearning methods generally preserved the models' broader capabilities, such as mathematical reasoning and OCR performance.
Implications for AI Safety
The findings indicate that relying solely on intra-modal evaluation—specifically text-only testing—can lead to an overestimation of a VLM's unlearning effectiveness. Because VLMs process information across multiple modalities, the authors argue that effective knowledge removal requires both cross-modal unlearning and comprehensive evaluation that accounts for how information is accessed.
Franklin Analysis
The evidence provided by the authors suggests that the modality used for unlearning is a critical variable in VLM safety. The observed asymmetry—where text-only unlearning fails to cover visual retrieval paths—indicates that "forgetting" in a VLM is not a monolithic process. Because the study utilized a controlled pipeline to ensure that text and image queries were semantically identical, the results provide a strong basis for the conclusion that current text-centric evaluation methods are insufficient for modern multimodal models.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!