You gotta remember that the LQG is a combination of a Kalman filter (a proper one, none of this FKF stuff) and a linear quadratic regulator.
The Kalman filter estimates the system state and outputs a rate estimate, a torque estimate and the bias.
What the rate does is obvious. The torque is more or less a differential term, since it's an estimation of how fast it accelerates and decelerates, and the bias is an integration of prediction error (scaled to max. torque).
All that should sound familiar. So you essentially have your the terms that are sort of equivalent to the traditional P, I and D terms, except they're quasi separate from each other.
The LQR then optimizes for the rate and torque terms, to reach target rates with least effort according to the system model configured by the autotune values. Bias is in raw actuator values, it just gets clamped and added.
As far as filtering goes, the default gyro filter is a 55hz first order (PT1 per Betaflight terms), in front of the Kalman, and the Kalman filter does some more first order action on the rate part. So you get the cascaded PT1 the other camp has been raving about in the past. A dynamic bitch might be worthwhile to investigate, the opinions differ on that. I might eventually tinker with that.
--edit: the noise figure in the autotune results is some estimation error. I would disregard that one.