We have developed and tested a new way of coupling bolometric light detectors to scintillating crystal bolometers based upon simply resting the light detector on the crystal surface, held in position only by gravity. This straightforward mounting results in three important improvements: (1) it decreases the amount of non-active materials needed to assemble the detector, (2) it substantially increases the light collection efficiency by minimizing the light losses induced by the mounting structure, and (3) it enhances the thermal signal induced in the light detector thanks to the extremely weak thermal link to the thermal bath.We tested this new technique with a 16 cm 2 Ge light detector with thermistor readout sitting on the surface of a large TeO 2 bolometer. The light collection efficiency was increased by greater than 50% compared to previously tested alternative mountings. We obtained a baseline energy resolution on the light detector of 20 eV RMS that, together with increased light collection, enabled us to obtain the best α vs β/γ discrimination ever obtained with massive TeO 2 crystals. At the same time we achieved rise and decay times of 0.8 and 1.6 ms, respectively. This superb performance meets all of the requirements for the CUPID (CUORE Upgrade with Particle IDentification) experiment, which is a 1-ton scintillating bolometer follow up to CUORE. tons, producing heat (phonons) that is measured by a suitable thermometer. The main difference among the various Bolometric Light Detector (BLD) instruments currently in use is the choice of the thermometer element, e.g. Transition Edge Sensors (TES) [12], Neutron Transmutation Doped (NTD) thermistors [13] or Micro Magnetic Calorimeters (MMC) [14].The work presented here was performed within the CU-PID framework [15,16], the future follow up of CUORE [17] that represents the largest world-wide bolometric experiment to date. The aim was to develop NTD-based BLDs with improved performance in terms of sensitivity, time response and simplified packaging for large arrays. Using the tiny Cherenkov light emission of TeO 2 [18,19] to decrease by two order of magnitude the α-induced background, requires a BLD with a S/N ratio of the order of ∼5 [16]: this corresponds to a RMS baseline resolution of the BLD of the order of ∼20 eV being the Cherenkov light signal of the order of 100 eV. Actually one can work towards the optimization of the light collection [20] and/or towards the energy resolution of the BLD or -as we made in this work-both. Additionally, in case of 100 Mo-based compounds, beside the same need to suppress the surface α-induced background, a fast time response of the BLD (≤ 1 ms) is mandatory to suppress the background induced the pile-up of the 2ν DBD [21]: also in this case the S/N ratio will play an important