VADAOrchestra: Neurosymbolic Orchestration of Adaptive Reasoning Workflows

How can organizations automate knowledge-intensive, dynamic workflows—such as financial risk assessment—that require both the flexibility of Large Language Models (LLMs) and the rigorous, auditable, and scalable guarantees of traditional Business Process Management (BPM) systems?

The authors introduce VADAOrchestra, a neurosymbolic framework that treats workflows as evolving reasoning processes. The system uses an LLM-based orchestrator to plan and adapt workflows at runtime. Crucially, these plans are encoded as logic programs in Vadalog (a fragment of Datalog+/-). By decoupling high-level orchestration from symbolic inference, the framework executes logical tasks via a dedicated engine, ensuring that every step is recorded in a verifiable, reproducible logical trace. The system also uses a dependency graph to enforce domain-specific constraints, preventing the hallucinations common in purely agentic architectures.

VADAOrchestra successfully bridges the gap between agentic flexibility and symbolic rigor. In financial use cases, such as assessing concentration risk, the system demonstrates that it can perform complex, multi-step investigations—like traversing corporate ownership chains to aggregate risk—that would overwhelm standard LLM context windows. Because the reasoning is grounded in a logic program, the system provides a complete, auditable trace of how it reached its conclusions, ensuring that the results are not just accurate, but also reproducible and explainable.

This work addresses the "black box" problem in agentic AI. By forcing the system to output a logical trace, VADAOrchestra makes automated decision-making suitable for high-stakes, regulated environments like banking. It also solves the scalability bottleneck of LLMs by using the orchestrator to selectively query data rather than loading entire datasets into the model's context window.