RouteScan: A Non-Intrusive Approach to Auditing MoE LLMs Safety via Expert Routing Telemetry
The episode discusses 'RouteScan,' a method for non-intrusively auditing the safety of Mixture-of-Experts (MoE) LLMs. Hosts explain that the paper shifts AI safety from qualitative reporting to quantitative measurement by analyzing internal routing telemetry. Key proposals include using penalty functions to proactively guide model behavior and constrain potential harmful paths.