CATA: Continual Machine Unlearning via Conflict-Averse Task Arithmetic

2026-05-18 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

The paper introduces CATA (Conflict-Averse Task Arithmetic), a novel method for continual machine unlearning in Vision-Language Models (VLMs). VLMs, despite their multimodal capabilities, face privacy and copyright issues due to large training datasets, necessitating unlearning. Unlike existing single-shot unlearning methods, CATA addresses the practical scenario of sequential removal requests over time. The authors identify three key challenges in this continual setting: effective knowledge removal, preservation of model utility, and prevention of knowledge re-emergence. CATA represents each unlearning request as a task vector and uses sign-aware conflict-averse aggregation of historical task vectors to suppress updates that could weaken prior forgetting effects. Experiments in both single-shot and continual settings demonstrate CATA's superior performance in forgetting effectiveness, model fidelity, and forgetting persistence compared to baseline methods.

Key takeaway

For research scientists developing or deploying Vision-Language Models, CATA offers a robust approach to managing sequential unlearning requests. If your VLM applications require continuous data removal due to privacy or copyright concerns, implementing CATA can ensure effective knowledge removal and model fidelity while preventing re-emergence of forgotten information. Consider integrating CATA's task arithmetic framework to maintain compliance and model integrity over time.

Key insights

CATA enables continual machine unlearning in VLMs by aggregating unlearning task vectors to prevent re-emergence of forgotten knowledge.

Principles

• Unlearning requests can be modeled as task vectors.
• Aggregating task vectors can preserve prior unlearning effects.

Method

CATA represents each forget request as an unlearning task vector, then aggregates historical task vectors using sign-aware conflict-averse aggregation to suppress conflicting update components that might weaken previous forgetting effects.

In practice

• Apply CATA for sequential data removal in VLMs.
• Use task vectors to manage unlearning requests.

Topics

• Continual Machine Unlearning
• Vision-Language Models
• CATA Method
• Task Arithmetic
• Knowledge Forgetting

Best for: Computer Vision Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, AI Ethicist

Related on AIssential

• CATA: Continual Machine Unlearning via Conflict-Averse Task Arithmetic — CATA enables continual machine unlearning for VLMs by aggregating unlearning task vectors to prevent knowledge re-emergence.
• KARLA: Knowledge-base Augmented Retrieval for Language Models — KARLA enables LLMs to dynamically retrieve and integrate external factual knowledge via special tokens, enhancing accuracy and traceability.
• Understanding Cross-Modal Contributions in Continual Vision-Language Models: A Theoretical Perspective — A new theoretical perspective clarifies cross-modal contributions in continual vision-language models, addressing catastrophic forgetting.
• Mechanism-Guided Selective Unlearning for RLVR-Induced Reasoning — Selective unlearning via mechanism-aligned targeting reduces collateral damage in RLVR-induced reasoning models.
• New CHEEM framework helps AI learn new tasks without forgetting old ones — CHEEM enables AI models to continually learn new tasks and adapt computation without forgetting old knowledge or sacrificing efficiency.
• RepSelect: Robust LLM Unlearning via Representation Selectivity — RepSelect achieves robust LLM unlearning by selectively targeting forget-set-specific representations, preventing easy reversal.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.