LOGIGEN: Logic-Driven Generation of Verifiable Agentic Tasks

The evolution of Large Language Models (LLMs) from static instruction-followers to autonomous agents necessitates operating within complex, stateful environments to achieve precise state-transition objectives. However, this paradigm is bottlenecked by data scarcity, as existing tool-centric reverse-synthesis pipelines fail to capture the rigorous logic of real-world applications. We introduce $\mathbf{\text{LOGIGEN}}$, a logic-driven framework that synthesizes verifiable training data based on three core pillars: $\mathbf{\text{Hard-Compiled}}\mathbf{\text{Policy}}\mathbf{\text{Grounding}}$, $\mathbf{\text{Logic-Driven}}\mathbf{\text{Forward}}\mathbf{\text{Synthesis}}$, and $\mathbf{\text{Deterministic}}\mathbf{\text{State}}\mathbf{\text{Verification}}$. Specifically, a Triple-Agent Orchestration is employed: the $\mathbf{\text{Architect}}$ compiles natural-language policy into database constraints to enforce hard rules; the $\mathbf{\text{Set}}\mathbf{\text{Designer}}$ initializes boundary-adjacent states to trigger critical policy conflicts; and the $\mathbf{\text{Explorer}}$ searches this environment to discover causal solution paths. This framework yields a dataset of 20,000 complex tasks across 8 domains, where validity is strictly guaranteed by checking exact state equivalence. Furthermore, we propose a verification-based training protocol where Supervised Fine-Tuning (SFT) on verifiable trajectories establishes compliance with hard-compiled policy, while Reinforcement Learning (RL) guided by dense state-rewards refines long-horizon goal achievement. On ${\tau }^2$-Bench, LOGIGEN-32B(RL) achieves a $\mathbf{\text{79.5\%}}\mathbf{\text{success}}\mathbf{\text{rate}}$, substantially outperforming the base model (40.7%). These results demonstrate that logic-driven synthesis combined with verification-based training effectively constructs the causally valid trajectories needed for next-generation agents.