The Mixture-of-Tunable-Experts (MoTE) approach allows for precise control over LLM behavior during inference without requiring retraining or fine-tuning.

• Uses functional Token Resonance Imaging (fTRI) to analyze neural activation patterns in DeepSeek-R1
• Enables on-demand modification of model responses and refusal behaviors
• Provides granular control over model outputs by manipulating expert weighting
• Creates more predictable and customizable AI systems with existing models

From a security perspective, this research offers a significant advancement in controlling potentially harmful outputs and tailoring refusal rates for sensitive content, enhancing model safety without compromising capability.

Mixture of Tunable Experts -- Behavior Modification of DeepSeek-R1 at Inference Time