ComplexConstraints: A Benchmark for Entangled Instruction Following

Surge AI introduces ComplexConstraints, a new instruction-following benchmark featuring 75 expert-crafted prompts with 1,559 evaluation rubrics that test interdependent constraints reflecting real professional work. Training a Qwen3-4B model on the accompanying dataset improved performance on ComplexConstraints itself while also generalizing to other benchmarks, boosting AdvancedIF scores by 8.45 percentage points and MultiChallenge scores by 10.1 percentage points.