Test-Time Unlearning via Sparse Autoencoder
This research presents a method for improving the privacy and security of large language models by making them forget specific information without retraining, which is a critical concern for users and developers.
Researchers propose ARIA, a test-time unlearning method that uses a sparse autoencoder to detect when a model is about to generate unwanted knowledge.
ARIA leaves model weights intact and only intervenes during these states, improving the trade-off between forgetting and retaining utility compared to weight-based methods.
Empirical evaluations on TOFU, R-TOFU, and WMDP benchmarks show ARIA reduces forgetting accuracy while maintaining performance on MMLU, and it remains robust against post-unlearning adversarial attacks.
Machine unlearning aims to remove specific knowledge from a trained large language model (LLM) without retraining from scratch. Existing methods modify model weights via gradient ascent and its advances. While effective on certain benchmarks, these weight-based approaches exhibit a sharp forget-utility trade-off, where stronger forgetting of target knowledge can degrade model utility, and unlearned knowledge may reappear under post-unlearning fine-tuning or prompt attacks. We propose ARIA (autoencoder-gated inference-time unlearning), a test-time unlearning method that leaves model weights intact and gates access to unwanted knowledge only when generation enters a forget-related state. ARIA uses sparse autoencoder (SAE) latents to train a lightweight linear detector, then applies an interpretable intervention on triggered states with negligible test-time overhead. Empirical evaluations on TOFU, R-TOFU, and WMDP show that ARIA improves the forget-retain trade-off over weight-based baselines across both a thinking model (DeepSeek-R1-Distilled-Qwen-1.5B) and an instruction model (Gemma-3-1B-it), e.g., reducing WMDP-cyber forget-set accuracy significantly while keeping MMLU within 1% of the pre-unlearning model. We further introduce three post-unlearning adversarial attacks targeting weight-space and decoding-space recovery, and find that ARIA remains robust under all three, with forgetting changing by less than 1% under attack. A feature-level case study leveraging the interpretability of ARIA suggests that some retain degradation may reflect response styles underlying the unlearning data rather than leakage of the targeted knowledge itself, highlighting a potential source of bias in unlearning task construction.
This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Test-Time Unlearning, large language models, Gemma-3, DeepSeek-R1.
Primary source verification: published directly by arXiv: DeepSeek with first-party authority. Cross-source corroboration: 3 independent sources have covered this development (including 量子位, AI Will), confirming industry-wide signal strength.