Metraipolis › archive › dc_803f3669dfe4

Test-Time Unlearning via Sparse Autoencoder — arXiv:2609.16229

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.16229
License (as archived)
CC0-1.0
Archived
2026-10-15T19:23:23.059Z
SHA-256 of the archived text
bb534a9469552f2b9636492b2afd08229f46d5e9cacb15c5210acdbccb798025
Stable URL of this page
https://metraipolis.com/documento/dc_803f3669dfe4

How to cite

Test-Time Unlearning via Sparse Autoencoder — arXiv:2609.16229 — arXiv (en), https://arxiv.org/abs/2609.16229 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_803f3669dfe4 (Metraipolis archive dc_803f3669dfe4, sha256 bb534a9469552f2b9636492b2afd08229f46d5e9cacb15c5210acdbccb798025)
BibTeX
@misc{dc803f3669dfe4,
  title = {Test-Time Unlearning via Sparse Autoencoder — arXiv:2609.16229},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.16229},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_803f3669dfe4; sha256 bb534a9469552f2b9636492b2afd08229f46d5e9cacb15c5210acdbccb798025},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.16229. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

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.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-16 in cs.CL).
Categorie arXiv: Machine Learning (cs.LG) ; Computation and Language (cs.CL).
Pagina del lavoro: https://arxiv.org/abs/2609.16229 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-10-15T18:32:32.889Z dalla pagina https://arxiv.org/catchup/cs.CL/2026-09-16?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).