Steering MoE LLMs via Expert (De)Activation
This addresses the need for efficient behavior control in MoE LLMs without retraining, but it is incremental as it builds on existing MoE routing methods.
The paper tackled the problem of controlling behaviors like faithfulness and safety in Mixture-of-Experts Large Language Models by detecting and steering behavior-linked experts, resulting in improvements of up to +20% in safety and +27% in faithfulness, and demonstrating vulnerabilities with safety drops of up to -100% in adversarial settings.
Mixture-of-Experts (MoE) in Large Language Models (LLMs) routes each token through a subset of specialized Feed-Forward Networks (FFN), known as experts. We present SteerMoE, a framework for steering MoE models by detecting and controlling behavior-linked experts. Our detection method identifies experts with distinct activation patterns across paired inputs exhibiting contrasting behaviors. By selectively (de)activating such experts during inference, we control behaviors like faithfulness and safety without retraining or modifying weights. Across 11 benchmarks and 6 LLMs, our steering raises safety by up to +20% and faithfulness by +27%. In adversarial attack mode, it drops safety by -41% alone, and -100% when combined with existing jailbreak methods, bypassing all safety guardrails and exposing a new dimension of alignment faking hidden within experts.