We proposed and experimentally demonstrated a highly sensitive optical fiber torsion sensor based on a reflective Lyot filter for the first time to our knowledge. The reflective Lyot filter is constructed by inserting a section of polarization-maintaining fiber (PMF) between a fiber linear polarizer and a 3dB coupler based fiber loop reflector. Based on the intensity modulation, the proposed torsion sensor exhibits a high torsion sensitivity of up to 20.336 dB/rad, one order of magnitude higher than the achieved in state-of-the-art. In contrast, the temperature cross-sensitivity and strain cross-sensitivity of the proposed torsion sensor are low to -2.0 × 10 -4 rad/°C and -6.39 × 10 -6 rad/με, respectively, thus overcoming the cross-sensitivity problem resulting from temperature and strain. Moreover, we perform the theoretical simulation of the proposed torsion sensor, and the simulation result obtained agrees well with the experiment results, vividly confirming the viability of the fiber Lyot filter based torsion sensor. Such fiber Lyot filter may find potential applications of highly sensitive torsion sensors in the field of modern smart structure monitoring.