Despite major advances in Earth system modeling, substantial gaps persist between observed and simulated carbon fluxes across both land and ocean domains. Bridging these gaps requires improved process understanding, tighter integration of observations, and targeted model innovation. First, I examine the strengths and limitations of current process-based models in representing net CO₂ fluxes, highlighting where, and why, they diverge from observational constraints. Second, I propose a strategic framework for advancing models to enable more robust climate predictions, emphasizing improved representation of critical processes, guided by remote sensing data and refined benchmarking practices. Finally, I present new results from the CliMA Land model demonstrating how observation-informed refinements to model complexity, specifically through the integration of global leaf trait variability into land surface radiative transfer schemes, substantially enhance simulations of surface albedo, energy balance, and carbon uptake. Together, these efforts underscore how advancing the representation of biophysical and biogeochemical processes, guided by observations, can accelerate innovation in Earth system modeling and strengthen the reliability of future climate projections.