arXiv:2606. 00651v1 Announce Type: cross Abstract: Mixture-of-Experts (MoE) architectures scale Large Language Models (LLMs) efficiently, enabling greater capacity with reduced computational cost by dynamically routing inputs to relevant experts, yet introduce a critical vulnerability: Safety Sparsity, where safety capabilities concentrate in few experts, making them susceptible to adversarial bypassing.
By Yitong Sun, Yao Huang, Teng Li, Ranjie Duan, Yichi Zhang, Xingjun Ma, Hui Xue, Xingxing Wei
arXiv:2608. 12821v1 Announce Type: new Abstract: Large language models (LLMs) remain vulnerable to harmful requests and jailbreak attacks.
By Fangzhou Chen, Shiji Zhao, Mengyang Wang, Qihui Zhu, Ranjie Duan, Maoxun Yuan, Xingxing Wei
arXiv:2606. 00686v1 Announce Type: new Abstract: The prevailing paradigm in large language model (LLM) alignment operates via erasure, filtering unsafe data or training models to strictly refuse harmful prompts.
By Maryam Hashemzadeh, Jerry Huang, Minseon Kim, Marc-Alexandre C\^ot\'e, Sarath Chandar
The paper introduces GRIP, an algorithm‑agnostic framework for machine unlearning in Mixture‑of‑Experts large language models. GRIP enforces hard geometric constraints on router updates, projecting gradient changes into the null space of the retain set’s routing matrix to prevent routing manipulation. Two variants—training‑time stochastic projection and post‑training analytical correction—show significant improvements in routing stability, retain accuracy, and resistance to white‑box adversarial recovery across two MoE models.
By Andy Zhu, Rongzhe Wei, Yupu Gu, Pan Li
The paper investigates why large language models lose safety alignment after benign fine‑tuning. It argues that safety alignment relies on a low‑rank, output‑routing geometry that becomes flatter during fine‑tuning, and that after only 100 benign examples this routing is sharpened in output‑side MLPs, leading to fragile safety while general performance remains relatively intact. Techniques like LoRA and ASAM can delay this collapse by reducing output‑side sharpness, but their effectiveness diminishes with larger fine‑tuning scales.
By Yitong Guo, Xiaoyi Chen, Siyuan Zhang, Xiaofeng Wang, Haixu Tang
The paper introduces SEAL, a training-time, parameter‑efficient defense that attaches a plug‑and‑play adapter to the shared expert component of Mixture‑of‑Experts models, and SEAL++, which adds an orthogonal constraint to preserve existing safety subspaces. By leveraging the always‑activated shared expert, SEAL mitigates the structural vulnerability of sparse routing to adversarial manipulation, reducing attack success rates by up to 60% with minimal impact on model capability. The approach is evaluated across six attack scenarios involving harmful prompting, jailbreaks, malicious fine‑tuning, and neuron pruning.
By Qingyu Meng, Yiwei Zha, Jiahuan Pei, Koen Hindriks, Herbert Bos, Min Chen