GURU: A Hierarchical Neuro-Symbolic Multi-Agent System for Verification-Aware Automation of Mission-Critical Digital Engineering and Operations

Allan Grosvenor, Anton Zemlyansky, Abdul Wahab, Kyrylo Bohachov, Aras Dogan, Ryland Adams, Dwyer DeighanarXiv 2026 · arXiv

TL;DR

GURU is a hierarchical neuro-symbolic multi-agent system designed for mission-critical operations where AI speed must be balanced with physical safety. By partitioning authority—letting learned models propose actions while a symbolic rules engine and validated physics retain final veto power—GURU achieves high-performance automation in fields like hypersonic flight and nuclear engineering that require rigorous verification.

What Problem Does This Paper Address?

In mission-critical domains like aviation, space operations, and nuclear power, the cost of failure is catastrophic. While modern AI (particularly Large Language Models and deep neural networks) can find solutions faster than humans, they are often "black boxes" that struggle with hard physical constraints and "hallucinate" unsafe actions.

Current regulatory bars from the FAA and NRC demand that AI be reliable, explainable, and controllable. The challenge is that purely symbolic systems (traditional code) are too brittle for complex, "open-world" data, while purely data-driven models are too unpredictable to serve as a safety basis. GURU addresses this by creating a "hybrid intelligence" that uses AI for what it's good at (prediction and search) without ever letting it have the final word on safety.

Methods

GURU utilizes a three-layer agent hierarchy built on a Common Internal Language (CIL)—a shared, neuro-symbolic representation that ensures all agents, whether learned or symbolic, speak the same "unit-aware" language.

Figure 1: The GURU hierarchy. The Planning layer performs long-horizon forecasting and planning with JEPA world models and model predictive control. The Action layer turns plans into bounded decisions through optimization, parameter prediction, and anomaly detection. The Navigation layer executes and monitors operations across software and physical systems. Feedback flows bidirectionally, so plans constrain low-level action while observations update the planner. The middle layer, labeled Discrete actions in the figure, is the Action layer.

The architecture consists of:

  • Planning Layer: Uses Joint Embedding Predictive Architecture (JEPA) world models and model predictive control for long-horizon forecasting.
  • Action Layer: Translates plans into discrete, bounded decisions using optimization and anomaly detection.
  • Navigation Layer: Handles the "boots on the ground" execution, monitoring software and physical systems in real-time.
  • Symbolic Verification Layer: This is the "load-bearing" safety gate. It checks every proposed action against hard physical limits (e.g., "will this move melt the reactor core?") before execution.

The system is supported by a Training Factory that extracts procedural knowledge from manuals and videos into a massive knowledge graph, allowing agents to be trained and validated in a reproducible pipeline.

Figure 2: The GURU platform. The Learning Engine trains Navigation, Action, and Planning agents. The rules engine validates schemas, ontologies, constraints, and execution state. The agent society coordinates specialized agents through message pools and a shared workspace for auditable execution. Inputs span sensor streams, simulation outputs, documents, and direct user interaction, and the same platform drives applications in digital engineering, space operations, and reactor operations.

Datasets

The authors evaluated GURU across several heterogeneous datasets and benchmarks:

  • Computational Fluid Dynamics (CFD): A 15,000-case Latin-hypercube campaign using the SU2 solver and a corpus of 312 aircraft surfaces.
  • Space Domain Awareness: Public TLE (Two-Line Element) orbital records and a "virtual range" of 200,396 high-fidelity GMAT simulations for maneuver detection.
  • Software Control: A ParaView knowledge graph containing 105,125 nodes extracted from 64 source packages (PDFs, videos, and APIs).
  • Airspace: Recorded flight traffic from NASA’s Sherlock archive.

Key Insights and Ideas

  • The "Physics Decides" Principle: Unlike "harnessed" LLMs where the model is at the center, GURU puts a validated physical reference at the center. The AI is a "proposer," not a "decider."
  • JEPA for Physical Stability: The paper finds that JEPA world models—which predict future states in a latent representation space rather than reconstructing raw pixels—are significantly more stable and accurate for physical tasks than fine-tuned language models.
  • Load-Bearing Symbolics: An ablation study showed that removing the symbolic constraints reduced boundary-layer capture in CFD grids by 13 percentage points. This proves that the symbolic layer isn't just "decoration" for safety; it materially improves the quality of the AI's output.
  • Authority Partitioning: The most critical insight is that trust in AI comes from diversity of implementation. By having a symbolic checker that is fundamentally different from the neural proposer, the system defends against "common-cause" failures where an AI might be "confidently wrong."
Table 1: A language model with a harness, compared with GURU.

Key Results

  • CFD Automation: GURU raised the converged-simulation rate on unseen aircraft geometries from under 1% to 74%. It successfully ran 54 live cases up to Mach 6 across three different supercomputing facilities.
  • Space Safety: In conjunction risk (collision) screening, GURU’s frozen JEPA model outperformed five other methods, including fully fine-tuned LLMs, while exhibiting several-times-lower seed variance.
  • Software Control: Using knowledge-graph retrieval and role adapters, GURU raised ParaView control success from 55% to 75%.
  • Extreme Scaling: The training factory demonstrated near-linear scaling, running on up to 8,000 nodes on the Aurora supercomputer.

Limitations and Future Work

The authors are transparent about the varying maturity of these applications. While the CFD and Space modules are "end-to-end" measured results, the Microreactor Digital Twin remains a "prospective" research target.

Current limitations include a reliance on synthetic data for training (which must be continuously validated against real-world observations) and the fact that in multi-aircraft control, separation is currently handled as a "cost" in the controller rather than a hard, verified constraint. Future work aims to map these results to formal aerospace and nuclear certification standards (like SAE ARP6983).

Summary updated

Abstract

An AI can find a safe action faster than any human, but it must never decide what counts as safe. We present GURU, a hierarchical neuro-symbolic multi-agent system for mission-critical digital engineering and operations, built on one partition of authority: learned models propose, rank, and accelerate; a symbolic rules engine checks every consequential action against hard limits before it executes; and validated physics confirms and decides. GURU runs a workflow as a closed control loop, with Navigation, Action, and Planning agents sharing one Common Internal Language and one training factory. Nuclear, aviation, and national-security guidance now requires this pattern; a harnessed language model keeps the model at the center, GURU puts a validated reference there. In computational fluid dynamics, the workflow raised the converged-simulation rate on unseen aircraft geometries from under 1 percent to 74 percent over 100 single-attempt cases, ran 54 live cases to Mach 6 on three facilities, and an ablation put the symbolic layer at 13 points of boundary-layer capture. Knowledge-graph retrieval and role adapters raised ParaView control success from 55 to 75 percent. In space domain awareness, where a language model cannot be trusted to decide, a frozen self-supervised world model under validated propagation scored highest of five methods at conjunction risk, including fully fine-tuned language models, with several-times-lower seed variance: the trustworthy design gives up nothing in accuracy, and the recipe transferred to reactor anomaly detection. A radar-to-avoidance debris pipeline was demonstrated in stages on a 2026 reentry; the hierarchy coordinates simulated teams of 200 aircraft. Training scaled near-linearly past 1,000 Aurora nodes and ran at 8,000. The approach extends to a control-grade microreactor digital twin with validated physics, not a learned model, as the safety basis.
BibTeX
@misc{grosvenor2026guru,
  author = {Allan Grosvenor and Anton Zemlyansky and Abdul Wahab and Kyrylo Bohachov and Aras Dogan and Ryland Adams and Dwyer Deighan},
  title = {{GURU: A Hierarchical Neuro-Symbolic Multi-Agent System for Verification-Aware Automation of Mission-Critical Digital Engineering and Operations}},
  year = {2026},
  eprint = {2611.01001},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.01001},
}
Figure