The discovery of the Higgs Boson marks a milestone in the Standard Model of particle physics, but key questions remain open, including the nature of dark matter, the hierarchy problem and the structure of electroweak symmetry breaking. These open questions point toward new physics beyond the Standard Model, which despite extensive searches at the LHC, remains elusive. This motivates both the exploration of non-standard signatures as well as increasing the precision of measurements. Achieving this increasingly requires innovative approaches beyond traditional object reconstruction and identification, and will be especially important for the High-Luminosity LHC (HL-LHC) era, where CMS will operate in a much higher collision-rate environment and deploy highly granular detectors such as the HGCAL. Machine learning (ML) provides powerful tools to meet these challenges and to extract maximal information from the recorded data. In this talk, I will discuss how modern ML based reconstruction in CMS is extending new-physics sensitivity by learning directly from low-level detector information, particularly the exotic decays of the Higgs Boson to photons, and enabling more precise estimation of jet substructure. I will also discuss the reconstruction challenges posed by the future upgrade of the CMS experiment and the ML based strategies being developed both for increasing the sensitivity to rare unusual signatures of new physics as well as precision measurements.