LOGIGEN: 3-Agent Synthesis Pipeline Generates 20K Verifiable Agentic Training Tasks, Boosts τ²-Bench from 40.7% to 79.5%

arXiv·medium signalXBlueskyLinkedInCopy link

LOGIGEN addresses the data scarcity problem in training agents for complex stateful tasks by using a three-agent system: an Architect converts natural language policies into database constraints, a Set Designer creates boundary-adjacent states triggering policy conflicts, and an Explorer searches for causal solution paths. Combined with RL on state-based rewards, LOGIGEN-32B achieves 79.5% on τ²-Bench versus 40.7% for the base model—a 38.8-point improvement using 20K synthesized tasks across 8 domains. The approach demonstrates that logic-driven synthesis with exact-state verification is more effective than traditional trajectory collection for training causally valid agents.