Block-Bench: A Framework for Controllable and Transparent Discrete Optimization Benchmarking

A team of researchers from Leiden University and the University of Adelaide has introduced Block-Bench, a groundbreaking framework for constructing highly controllable benchmarks in discrete optimization. Published on arXiv (cs.NE/2604.06973), the system builds problems from modular 'block functions'—each mapping a subset of variables to a value—combined with weight factors and an adjacency graph defining their dependencies. This architecture provides unprecedented fine-grained control over core problem properties like modularity, ruggedness, and variable interactions, moving beyond traditional black-box benchmarks.

The framework's key innovation is its transparency: by analyzing intermediate block values, researchers can dissect algorithm performance not just by final objective scores, but at the level of variable representations within solutions. This is particularly valuable for analyzing heuristics on large-scale, multi-modal problems where understanding *how* an algorithm succeeds or fails is as important as the result. The authors demonstrate its utility in studying self-adaptation and diversity control in evolutionary algorithms.

Block-Bench is designed to support broader research domains, including dynamic algorithm configuration and multi-objective optimization, by enabling the systematic creation of benchmark families with known, tunable characteristics. This addresses a critical need in optimization research for reproducible, transparent, and purpose-built testing environments that can keep pace with increasingly complex algorithms.