EvoSyn: Generalizable Evolutionary Data Synthesis for Verifiable Learning

Constructing generalizable synthetic verifiable data is hard: generation is hallucination-prone, and weak verification artifacts fail to separate strong from weak solutions. EvoSyn is an evolutionary, task-agnostic, strategy-guided data synthesis framework that, from minimal seed supervision, jointly synthesizes problems, diverse candidate solutions, and verification artifacts. It iteratively discovers filtering strategies via a consistency-based evaluator that enforces agreement between human-annotated and strategy-induced checks.

Training on EvoSyn-synthesized data yields significant improvements under both RLVR and model distillation, on LiveCodeBench and the AgentBench-OS agent task.