Identifying the location and distribution of nonaqueous phase liquid (NAPL) in the subsurface constitutes a vital step in the design and implementation of aquifer remediation schemes. In recent years, partitioning interwell tracer tests (PITT) have gained increasing popularity as a means to characterize NAPL saturation distribution in situ. In this method a suite of conservative and partitioning tracers are injected into the contaminated site. The chromatographic separation between the conservative and the partitioning tracers can be used to infer NAPL saturation distribution. The conventional approach to the analysis of the tracer response uses a first-order method of moments to compute average NAPL saturation in the tracer swept regions and cannot provide detailed spatial distribution of the NAPL. We propose a computationally efficient streamline-based inverse method for analyzing partitioning interwell tracer tests to estimate three dimensional spatial variation of NAPL saturation in the subsurface. Our approach is based on an analogy between streamlines and seismic ray tracing and relies on an analytic sensitivity computation method that yields sensitivities of the partitioning tracer response to subsurface parameters such as porosity, hydraulic conductivity, and NAPL saturation in a single streamline simulation. The inversion of tracer response is carried out in a manner analogous to seismic waveform inversion whereby we first match the “first arrival” followed by matching the “amplitudes” of the tracer response. The power and utility of the method is illustrated using synthetic and field applications. The field example is from the Hill Airforce Base, Utah, where partitioning tracer tests were conducted in an isolated test cell with 4 injection wells, 3 extraction wells, and 12 multilevel samplers. Tracer responses from 51 sampling locations are analyzed to determine hydraulic conductivity variations and NAPL saturation distribution in the test cell. Finally, a performance comparison with simulated annealing shows that our proposed approach is faster by 3 orders of magnitude.