New ARIA method unlearns LLM knowledge without weight modification

Researchers have developed ARIA (autoencoder-gated inference-time unlearning), a novel method for removing specific knowledge from large language models without altering their core weights. Unlike traditional weight-modification approaches, ARIA operates at inference time, using a lightweight detector trained on sparse autoencoder latents to intervene when unwanted knowledge is accessed. This approach aims to mitigate the forget-utility trade-off and improve robustness against post-unlearning attacks. Empirical results on benchmarks like TOFU, R-TOFU, and WMDP demonstrate ARIA's effectiveness in reducing targeted knowledge recall while preserving model utility and resisting adversarial recovery attempts. AI

IMPACT This research could lead to more robust and efficient methods for controlling sensitive information within LLMs, improving safety and privacy.

RANK_REASON The cluster contains a research paper detailing a new method for machine unlearning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• ARIA
• arXiv
• DeepSeek-R1-Distilled-Qwen-1.5B
• Gemma-3-1B-it
• Hugging Face
• Massive Multitask Language Understanding
• R-TOFU
• Test-Time Unlearning via Sparse Autoencoder
• TOFU
• WMDP

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →