Successful reactive balance control requires coordinated modulation of hip, knee, and ankle torques. Stabilizing joint torques arise from neurally-mediated feedforward tonic muscle activation that modulates muscle short-range stiffness, which provides instantaneous "mechanical feedback" to the perturbation. In contrast, neural feedback pathways activate muscles in response to sensory input, generating joint torques after a delay. However, the specific contributions from feedforward and feedback pathways to the balance-correcting torque response are poorly understood. As feedforward- and feedback-mediated torque responses to balance perturbations act at different delays, we modified the sensorimotor response model (SRM), previously used to analyze the muscle activation response, to reconstruct joint torques using parallel feedback loops. Each loop is driven by the same information, center of mass (CoM) kinematics, but each loop has an independent delay. We evaluated whether a torque-SRM could decompose the reactive torques during balance-correcting responses to backward support surface translations at four magnitudes into the instantaneous "mechanical feedback" torque modulated by feedforward neural commands before the perturbation and neurally-delayed feedback components. The SRM accurately reconstructed torques at the hip, knee, and ankle, across all perturbation magnitudes (