TLDR: The XCSP3 Competition 2025 evaluated constraint solvers on diverse Constraint Satisfaction and Optimization Problems. Organized by CRIL, it featured 33 novel problems, a range of solver technologies, and multiple competition tracks. The event provides a benchmark for progress in AI optimization, with detailed results available online.
The annual XCSP3 Competition recently concluded its 2025 edition, bringing together leading minds in constraint programming to evaluate the performance of various solvers on a diverse set of combinatorial problems. Organized by Gilles Audemard, Christophe Lecoutre, and Emmanuel Lonca from CRIL, University of Artois & CNRS, France, the competition serves as a crucial benchmark for advancements in artificial intelligence and optimization techniques.
The competition focuses on two main types of problems: Constraint Satisfaction Problems (CSP) and Constraint Optimization Problems (COP). These problems are represented using the XCSP3 format, specifically its core subset, which handles integer variables and a standard set of 24 global constraints. These constraints cover a wide range of logical and mathematical conditions, from ensuring all values are different (allDifferent) to managing resource allocation (cumulative, binPacking).
Problem Selection and Diversity
A significant aspect of the competition is the careful selection of problems. For 2025, 33 problems were chosen, emphasizing novelty to prevent solvers from overfitting to previously published challenges. The organizers also prioritized diversity in the types of constraints involved, aiming to represent all popular XCSP3-core constraints without over-representing any single one. Furthermore, the problems were designed to scale up in difficulty, allowing researchers to observe how solvers behave as instances become progressively harder.
Some of the intriguing problems featured in the CSP category included “Accordion,” a single-player card game, “Almost Magic,” which involves placing distinct positive integers into a grid to form almost magic squares, and “Chain Reaction,” where participants create a chain of integers with specific divisibility rules. In the COP category, problems like “Altered States” challenged solvers to find U.S. states within a grid using King’s moves to maximize a score, while “Roller Splat” involved navigating a ball to paint an entire grid in the fewest moves.
Participating Solvers and Methodologies
A variety of constraint solvers participated, each employing different underlying technologies and heuristics. Notable participants included ACE (which ran off-competition due to its author’s involvement in instance selection), BTD and miniBTD (tree-decomposition based), Choco (a Java library), CoSoCo (a C++ solver), and CPMpy-* (a Python library acting as a middleware to various backend solvers like OR-Tools CP-SAT, Chuffed, Gecode, Z3, Gurobi, and Exact). Other solvers like Exchequer, Fun-sCOP (SAT-based), Nacre (nogood recording), Picat (SAT-encoded), PyCSP3-ortools, RBO and miniRBO (restart-based optimizers), Sat4j-CSP-PB (pseudo-Boolean-based), and toulbar2 (cost function network solver) also showcased their capabilities.
These solvers utilize sophisticated techniques such as backtrack search, generalized arc consistency, various variable and value ordering heuristics, and restart policies to navigate complex search spaces. Many also incorporate conflict analysis and solution-based phase saving to enhance efficiency, especially for optimization tasks.
Also Read:
- Team Twente’s Hybrid Strategy for Integrated Healthcare Timetabling
- Enhancing Decision-Focused Learning with Dual-Guided Surrogates for Scalable AI
Competition Tracks and Results
The competition featured several tracks: CSP (sequential, 30 min timeout), COP (sequential, 30 min), Fast COP (sequential, 3 min), Parallel COP (parallel, 30 min), Mini CSP (sequential, 30 min), and Mini COP (sequential, 30 min). Solvers were allocated a CPU and 64 GiB of RAM. The scoring system for CSP rewarded solvers for correctly determining satisfiability, while for COP, points were awarded based on finding the best known solution and proving optimality.
The detailed results, including solver traces and rankings, are publicly available on the competition’s website. For instance, in the CSP track, Picat, PyCSP3_ortools, and CPMpy_ortools emerged as top performers. In the COP track, CPMpy_ortools, CoSoCo, and Picat demonstrated strong results. The competition highlights the ongoing innovation in constraint programming and the diverse approaches researchers are taking to tackle challenging combinatorial problems. For more in-depth information, you can refer to the official proceedings. Proceedings of the XCSP3 Competition 2025.
