Reverse link (or uplink) performance of cellular systems is becoming increasingly important with the emergence of new uplink-bandwidth intensive applications such as Video Share , where end users upload video clips captured through their mobile devices. In particular, it is important to design the system to provide good user throughput in most of the coverage area, including at the cell edge. Soft fractional frequency reuse (FFR) is one of the techniques for mitigating inter-cell interference in cellular systems, leading to overall spectral efficiency enhancements and/or cell edge throughput improvements. We propose a novel algorithm that dynamically creates efficient soft FFR patterns on the uplink of orthogonal frequency division multiple access (OFDMA) based cellular systems; this allows the system to " automatically" adapt to user traffic distribution and system layout. Our algorithm is based on systematically ascending towards a local maximum of the system-wide sum of user utilities, which depend on user throughputs. We show that this can be done in a semi-autonomous fashion: each sector does its resource allocation independently, with only an infrequent periodic exchange of interference costs between neighboring sectors. The proposed algorithm, called Multi-sector Gradient for Uplink (MGR-UL), allocates in-sector resources (power, frequency, time-slots to each user) in a way that simultaneously takes into account both the benefit to its "own" users' utility and the cost of creating interference to neighboring sectors; along with that each sector estimates the cost of interference to itself. Extensive simulation results show that significant performance benefits (up to 69% in total throughput in some typical scenarios) can be achieved with respect to a baseline approach. Simulations also show the automatic formation of soft FFR patterns.