Get Started
Home
Topics
Search
Library
Research questionHow can mixture-of-experts language models preserve safety when adversaries manipulate sparse expert routing?Sparse routing makes a model’s safety behavior depend on which experts are activated. Jailbreaks, malicious fine-tuning, and pruning can disrupt those safety-critical pathways, allowing harmful behavior despite prior alignment.
AI
Alignment & Safety
LLM Pretraining & Post-training
Machine Learning
Natural Language Processing
Latest papersRecent research connected to this question, newest first.SEAL: Reinforcing Global Safety in Mixture-of-Experts through Shared Expert ALignmentApplies to mixture-of-experts language models, including hybrid architectures with always-active shared experts. The source examines harmful prompting, jailbreaks, malicious fine-tuning, and safety-critical neuron pruning, reporting training-time defense results across six attack scenarios.research paper · Sep 2, 2026
Related questions
How does sparse mixture-of-experts routing trade off approximation, learning error, and computation under misspecification and evolving experts?How should experts be pruned in over-dispersed MoE routing when router importance and perplexity mislead?When can factual recall in sparse MoE language models be attributed to one expert rather than an expert set?How can language models distinguish trusted instructions from untrusted text to resist prompt injection?