Gennesaret Kharistio Tjusila, Alexander Hoen, Nils-Christian Kempke, Gioni Mexi, Timo Berthold, Ambros Gleixner, Thorsten Koch, and Sebastian Pokutta won the 2026 Land-Doig MIP Competition, announced at the Mixed-Integer Programming Workshop 2026. The Land-Doig MIP Competition is held yearly in honor of Ailsa H. Land and Alison G. Harcourt (née Doig), the inventors of the branch-and-bound algorithm for mixed-integer programming (MIP). It aims to encourage the development of novel practical techniques for MIP solving. The theme of this year's Land-Doig MIP Competition was GPU-Accelerated Primal Heuristics for MIP, motivated by the increasing momentum behind GPU-based optimization methods such as PDLP, a first-order LP solver based on primal-dual hybrid gradient methods.
The ZIB winning submission — CHAP — is built on three main components: first, on the CPU side, various Fix-and-Propagate strategies along with a CPU-based tabu search and feasibility pump; second, on the GPU side, a GPU-native tabu search algorithm along with low-precision LP solving via cuPDLPx; and third, a coordination mechanism for the various heuristics implemented through a shared solution pool containing high-quality feasible and partial solutions.
On the 50 test instances given for the competition, the ZIB submission found feasible solutions for 47 instances, ahead of Gurobi in default mode (44) and NVIDIA cuOpt in heuristics-only mode (43). On the same instance set, the team further demonstrated that a combined CPU–GPU approach outperforms a pure CPU approach on both gap (8.84% vs. 19.57%) and primal integral (41.84 vs. 72.59). The photo shows, from left to right: Christian Tjandraatmadja, software engineer at Google Research's Operations Research team, and Gioni Mexi, who accepted the award on behalf of the team.
