United States

Naval operators must execute missions characterized by heterogeneous platforms, contested environments, uncertain sensing, and tightly coupled policy and doctrinal constraints. In contrast, contemporary machine learning ecosystems often presume expert users, well-structured inputs, and narrowly scoped tasks, assumptions that often break down in operational naval contexts. 

This talk introduces LEXICON, an adaptive multi-agent orchestration framework that enables operators to express tasking in natural language. LEXICON leverages a constellation of specialized large language model (LLM) agents to convert these high-level intents into deterministic, verifiable, and auditable workflows that integrate seamlessly with existing planning tools, simulation backends, and retrieval-augmented knowledge sources. The result is a science-driven bridge between unstructured human intent and rigorous algorithmic execution, enabling mission planning pipelines that are transparent, reproducible, and compatible with the Navy’s emerging autonomy and decision-support architectures. 

Key elements include: 
-Multi-agent orchestration with clearly scoped roles wired in a GraphFlow that enforces ordering and termination. 
-A strict data contract and metadata-driven Execution Coordinator that turns REQUIRED_TOOLS lists into deterministic tool runs, with pre- and post-execution gates checking inputs, outputs, and coverage against tool metadata. 
-A retrieval-augmented context layer: document ingestion and embedding build a tagged index; a deterministic DocRetriever agent uses it to enrich context and close gaps, with “doc overrides” merged under explicit rules and checked by the same context-integrity and gating mechanisms as user-provided context. 
-Guardrails including a tool allowlist with an inline JSON to function-call shim that converts agent-emitted call blocks into actual tool invocations, strict versus adaptive execution modes, and a context-integrity policy that prevents agents or RAG from fabricating or silently dropping critical information. 
-Telemetry, tickets, and bounded loopbacks, where detected defects (for example missing geometry or context drift) generate targeted repair requests back to the responsible agent rather than silently continuing. 

We illustrate the approach with Navy-relevant proxy scenarios, including hotspot search and patrol routing with no-fly zones and scan policies, using run logs and state traces to show both successful missions and common failure modes. The talk is organized in two parts: a system overview covering the multi-agent plus RAG architecture and guardrail mechanisms, followed by results and lessons learned from these scenarios—focusing on how to wrap LLM-based agents around legacy tools without violating policy and safety constraints, how much structure data contracts and guardrails need to be useful in naval settings, and what open issues remain for deploying similar orchestration layers in operational environments.

NAML 2026

From Intent to Action: Multi-Agent LLMs Coordinating Naval Planning Tools

retrieval‑augmented generation

tool orchestration

naval operations

guardrails

llm agents

mission planning

#### *Message from the General Chair, Aline Villavicencio*
I’m delighted and honoured to welcome you to the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2026), taking place in the beautiful city of Rabat, in Morocco, in March 24-29, 2026. EACL is the flagship European conference of the Association and EACL 2026 proudly continues our field’s tradition of excellence in scholarship, innovation, and inclusivity. I am deeply grateful to the many volunteers whose dedication, generosity, and tireless efforts have made this conference possible.
For the first time EACL is being hosted in the African continent. This is an important milestone for our community, and we are grateful to our Moroccan hosts for enabling this historic moment by bringing this edition of EACL to Rabat. We are also delighted that the Second Arabic NLP School is co-located with EACL. We hope attendees enjoy this wonderful opportunity to strengthen ties with the Computational Linguistics communities across the African continent. *[Read full message](https://drive.google.com/file/d/14NlmHvwM6fPJuMmOvVh7K0vtQbEyv3SZ/view?usp=sharing)*<br><br>

<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://underline.io/events/522/reception','_blank')">Go to Workshops and Tutorials Program</button></html>
<br><br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>Access credentials will be sent to your email from Underline -  subject line "Welcome to EACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.eacl.org/registration/) first.

**Online Registration Form**: https://acl.swoogo.com/eacl2026

Registration Required

EACL 2026

Welcome to the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL). EACL 2026 will be held in Rabat, Morocco, from March 24–29, 2026. 

technical paper

### Welcome to the 10th annual workshop on Naval Applications of Machine Learning!

NAML is the premier annual event showcasing current machine learning research relevant to Naval applications. We welcome participation from the defense community, industry, and academia. The NAML workshop is an opportunity for researchers from Navy research and development centers and our partner organizations to exchange technical ideas and to come together as a team. 
Naval Information Warfare Center Pacific is partnering with AFCEA San Diego chapter to host the workshop.
#### On-Demand Access

* This is an on-demand site and all the recorded content from the workshop is available here. 
* Access to the recordings is exclusively available to registered participants.
* Each registrant has received an email notification, with instructions on how to access the content.

[![](https://assets.underline.io/markdown_image/1/image/60cd12aa376c191d3467ea2ee60944ea.png)](https://naml2026.org/)

The goals of NAML are:

* To spread awareness of current machine learning research relevant to Naval applications
* To connect machine learning researchers with experts in military requirements
* To build and strengthen collaborations within the defense research community

#### *"What we do as a Navy is a human business. Don’t pass up an opportunity to come together as a team. Don’t pass up leadership opportunities, and don’t outsource them to a PowerPoint. If you have an opportunity to come together as a team, to increase belonging, do it."*
Admiral John M. Richardson, 31st Chief of Naval Operations
<br><br>

This is an on-demand site. All recorded content from NAML 2026 is available here and access to it is exclusively available to registered participants.

**Please log in using the email address you registered with.** If you have any issues logging in, please contact us at <naml2026@underline.io>.

Let’s celebrate the past decade by investing in the next. 🚀


Please log in to explore this event.

This is an on-demand site. All recorded content from NAML 2026 is available here and access to it is exclusively available to registered participants.

**Explore the latest advancements in machine learning tailored to Naval applications 🔎**

With on-demand access to NAML 2026, you can watch all lectures & browse poster presentations at your convenience, while also engaging directly with presenters via the Underline platform - ask questions, exchange ideas, and stay connected.

**If you still haven’t registered, you can do so now.**
👇
**Take advantage of a special one-time offer: full access for just $99, with complimentary registration for students, government, and military employees.**

**Get access here: https://naml2026-ondemand.paperform.co/**

Please check your spam folder, or contact our support team at **<naml2026@underline.io>** if you registered, but still haven't received an access email.

Let’s celebrate the past decade by investing in the next. 🚀

Please register!

The 10th Annual Workshop on Naval Applications of Machine Learning (NAML 2026) was held in San Diego, California, March 2-5, 2026. NAML is a forum for interdisciplinary experts from government, industry, and academia to discuss results of ongoing research on theory and applications of machine learning to address some of the most challenging problems faced by our naval forces. 



The Marine Core Planning Process (MCPP) is a cyclical and disciplined process for military planners to develop, test, and ultimately transition Courses of Action (COAs) into orders. However, many steps within MCPP are manual, time intensive, and resource constrained. Under the COA Optimization, Analysis, and Comparison Harness (COACH) project, a research effort funded under the Office of Naval Research (ONR) Code 34 Science of AI Program, we aim to address these challenges of developing COAs by building Artificial Intelligence (AI) enabled decision-making tools and integrating them into MCPP. Specifically, we see two main challenges: 1) there are limitations in existing AI training frameworks that hinder the development and ultimate adoption of AI technologies for military decision-making processes, and 2) the data required to train such models is severely limited. To overcome these challenges, we will present Framework for Intelligent Simulation Command with Hierarchically Embedded Reinforcement Learning (FISCHER), a scheme for training AI decision-making models, and a newly developed, high-speed, low-fidelity military simulator to train COA generation models. FISCHER is an adapter between AI agent code and battlespace data streams, whether they be simulations of military operations or live real-world operations, to enable rapid prototyping of new AI models or to plug in existing models into new environments. We will present our current progress in training Reinforcement Learning (RL) agents, a type of Machine Learning (ML), in COA generation tasks. These developed tools under COACH will be integrated into the Higher Echelon Mission Planner (HEMP) and the Small Unit Mission Planner (SUMP), planning tools being developed under ONR Code 34’s Artificial Intelligence To Operations program.

Course of Action Optimization, Analysis, and Comparison Harness (COACH)

Analysis of civil considerations is the topic that may receive the least amount of time in the
planning course at the Expeditionary Warfare School (EWS) but is the step that can
potentially take the most time during operational planning. The Marine Corps Planning
Process (MCPP) calls for planners to evaluate the civil terrain by analyzing six
characteristics of the Area of Operations (AO): Areas, Structures, Capabilities,
Organizations, People, and Events (ASCOPE). For each of these ASCOPE categories, six
operational variables are to be considered: People, Military, Economic, Social,
Informational, and Infrastructure (PMESII).

Crosswalking the ASCOPE categories with the PMESII variables produces thirty-six
categories of civil considerations that each must be analyzed individually. The data that supports this analysis is typically complex, includes a large amount of unstructured text,
and is unique to the Area of Operations (AO). It is typically not possible to re-use previous
analyses or even data sources when performing MCPP in a new region.

Under sponsorship of ONR Code 34’s Artificial Intelligence To Operations program, the
Civil-Considerations thrust of our COA Optimization, Analysis, and Comparison Harness
(COACH) project uses the ASCOPE/PMESII conceptual structure to automate the process
of identifying, retrieving, and organizing civil considerations information.

COACH’s Civil Considerations Service (CCS) uses a Large Language Model (LLM) to identify
information relevant to each of the thirty-six crosswalk categories in a corpus of
unstructured text. Depending on the specific LLM used, the CCS can process non-English
data and produce English-language translations of the identified information. The CCS
produces a draft of the PMESII/ASCOPE crosswalk that users can then review and modify
to produce the final MCPP planning product.

The CCS is currently being integrated with the Higher Echelon Mission Planner (HEMP) and
the Small Unit Mission Planner (SUMP) tools, both of which are also AI2OPS projects.

Automation to Address Civil Considerations in the Marine Corps Planning Process

Ocean surface robotic swarms offer significant potential as reconfigurable landing platforms, ocean monitoring and communication relays. Drawing inspiration from self-assembled chiral crystals formed by starfish embryos, our lab has developed a magnetomechanically actuated spinning robot swarm platform (MASbotS) that self-assembles at the water surface and undergoes controllable phase changes between rigid and fluid states through tuning of interbot repulsion. To explore decentralized control, we apply a reinforcement learning framework in simulation using centralized training with decentralized execution (CTDE) and show that a single shared policy allows a swarm model to navigate to targets with robustness to changes in swarm size, since each agent relies only on local observations. The same approach succeeds in simulated environments with obstacles. To encourage adaptive behavior, we are beginning to coevolve the policy with controlled changes in the simulated environment. Next steps include transferring learned policies to physical MASbotS robots with real-time actuation and sensing, and incorporating feedback to tune swarm interactions for robust operation in real marine settings.

Decentralized Reinforcement Learning for Self-Assembling Ocean Surface Robot Swarms

Cyber-physical systems (CPS)—composed of interdependent IT and OT components—underpin critical infrastructure, industrial control systems, and modern enterprises. Yet CPS security remains fragmented: traditional IT tools often disrupt fragile OT processes, while OT security practices rely heavily on isolation and obscurity. This gap leaves CPS vulnerable to increasingly sophisticated threats mapped across the MITRE ATT&CK framework and demonstrated in real-world incidents such as ransomware-induced disruption of industrial operations. To address this persistent risk, this dissertation proposes an Industry 4.0–aligned cybersecurity framework that integrates high-fidelity digital twins (DT), specialized artificial intelligence security agents, and continuous integration/continuous deployment (CI/CD) principles to evaluate, remediate, and strengthen CPS security without jeopardizing live operations.

The research constructs a cloud-based DT of an operational CPS modeled on the Alabama A&M University Industry 4.0 laboratory—a representative segment of the Tennessee Valley Authority’s critical infrastructure. This DT replicates network-store-compute elements, OT device behavior, and interdependencies at sufficient fidelity to support destructive testing, iterative remediation, and performance measurement. Within a CI/CD-like pipeline, multiple specialized AI agents independently perform vulnerability scanning, OT-safe threat hunting, configuration assessment, and policy verification. Their outputs feed into a human-in-the-loop (HITL) supervisory process supported by a general-purpose LLM acting as an analytical integrator. This structure enables the identification of false positives, confirmation of actionable threats, ranking of remediation priorities, and comprehensive change-control documentation aligned with NIST and OMB requirements.

A quasi-experimental design evaluates two hypotheses: whether modern enterprise EDR alone is sufficient to secure CPS, and whether multifactor, DT-based AI testing yields superior defensive posture. Statistical analysis across repeated, controlled trials demonstrates that EDR-only approaches miss classes of OT-specific vulnerabilities and occasionally introduce operational risk. In contrast, the multifactor DT pipeline provides earlier detection of compound vulnerabilities, safer remediation pathways, and improved assurance of CPS resilience, all without disrupting production systems.

This work shows that DT-driven, AI-augmented CI/CD pipelines can transform CPS cybersecurity from reactive patching to proactive, model-verified resilience. The proposed framework is generalizable across energy, transportation, healthcare, defense, and other high-risk sectors where CPS integrity is mission-critical. Future extensions include LLM-based autonomous orchestration, broader OT coverage, and integration of continuous mission-impact modeling.

Enhancing Cyber-Physical System and Continuos Devsecops Using Specialized and General Purpose AI Within Digital Twins

Two central challenges in multi-agent reinforcement learning (MARL) are (i) effectively aggregating information from local observations through inter-agent communication, and (ii) improving communication efficiency by reducing unnecessary message exchanges. The solutions for these challenges are inherently conflicting to each other. While information aggregation is crucial for distributed agents to build global awareness and make effective decisions, indiscriminate aggregation often leads to excessive communication overhead. In this talk, we present a novel centralized training with de-centralized execution (CTDE) framework, BElief-based Predictive Auxiliary Learning (BEPAL) with communication reduction, to address both challenges. Compared to traditional reinforcement learning, agents in BEPAL maintain a richer and more accurate latent representations of environment dynamics with the help of the auxiliary learning task, which reconstructs a belief of the global environment from local observations and inter-agent communications. In addition, it uniquely incorporates an unlearning-based communication advantage baseline, enabling agents to learn when to suppress uninformative messages during training, thus improving communication efficiency. We evaluated BEPAL on four cooperative MARL benchmarks: Predator-Prey, Traffic Junction, Google Research Football, and Robotic Warehouse. In all environments, BEPAL consistently outperforms prior methods, achieving an average 31.33% improvement in performance and demonstrating more stable convergence. The communication unlearning stage effectively reduces communication by 22.25% with nearly no impact to the agent’s performance.

Communication Efficient Multi-Agent Systems using BEliefbased Predictive Auxiliary Learning

Genuine human understanding is the underpinning of small and large scale cooperating teams. Understanding is the bridge between information and meaning. It grounds words in perception, actions, and shared values. Without it, communication risks becoming hollow, and trust becomes fragile. While LLMs are evolving with reasoning, planning, tool use and agentic systems to become more valuable at thinking jobs, it is hard to overlook the fact that they do not truly understand like humans. Compounding the problem, human users anthropomorphize AI systems, leading to over trust in long-term collaborations.

In this paper we examine ways in which LLMs can be evaluated for human like understanding even on tasks they complete satisfactorily. Based on that we postulate that the gap in human and LLM understanding engenders a critical barrier for long running human machine teams especially in high stakes domains. We argue that this gap cannot be bridged by improving LLM reasoning alone. 

In view of this, any human-machine long running system should be designed to explicitly account for this disconnect by: avoiding a blackbox paradigm (opaque understanding and uncertainties in LLMs), creating a calibrated trust design (such as using confidence scoring), formalizing human oversight and grounding (mandatory review gates) and creating designs for context aware interaction (that explicitly overcome misinterpretation).

We propose a framework of principles for high-stakes human machine team collaboration system design to help narrow the gap:
1. Mode - combine LLMs with causal reasoning systems to reduce hallucinations and improve consistency 
2. Interaction - leverage explainable AI techniques such as contrastive explanations and traceability
3. System - treat LLM as tools, not as teammates with precise roles

The immediate goal is not to build LLMs that "understand like humans," but to design systems that compensate for their limitations while leveraging their strengths. However, we do recognize that the LLMs that exhibit human like understanding might arrive in future.

The Understanding Gap: Managing Risk in Human-LLM Collaboration When Reasoning Is Not Enough

Background: Military operations are rapidly changing due to new robotic technologies. As evidenced in Ukraine, unmanned systems are becoming indispensable tools for situational awareness, strike, logistics, and force protection. Therefore, skillful deployment, operation, maintenance, and repair of unmanned systems are becoming essential for unit lethality and survivability. To address these needs, NAVADMIN 036/24 announced the establishment of the Robotics Warfare (RW) Specialist Rating. Currently, an introductory “A-school” curriculum has been developed and advanced “C-school” courses are planned. There is also a desire for just-in-time training (JITT), which consists of initial general training followed by shorter system-specific fine-tuning later.

Objectives: The goals of this work are to identify RW training gaps, methods, and performance metrics to:
(1) Update A-school
(2) Create C-school
(3) Enable JITT

Approach: Our approach focuses on RW Sailor mental models (MMs), i.e., cognitive representations of systems and how they work. In particular, we are investigating how to nurture MMs that enable Sailors to rapidly acquire and retain the knowledge, skills, and abilities (KSAs) needed to effectively operate and maintain unmanned machine-learned systems. We probe MM accuracy in interviews by obtaining trainee descriptions of system capabilities, limitations, control, troubleshooting, operations, and dependencies, which we grade by comparing to instructor documentation. Trainees also make performance predictions of system accuracy, networking, mechanics, controls, and problems. These are compared to performance outcomes at later live events. By probing MMs, we identify deficiencies in content and methods to enhance current A-school training, gaps in C-school training, and MMs that are similar across systems. This allows us to identify common concepts and KSAs to create scaffolding to enable JITT.

Results: Thus far, this effort has several findings. Current training efforts would benefit from more hands-on learning—trainees consistently desired this, and those who received no hands-on learning demonstrated poorer understanding and knowledge retention. Future C-school training could address identified KSA gaps in networking, machine-learned computer vision, and acoustics. Trainees consistently expressed frustration and inability to solve networking problems. Many did not understand common computer vision pitfalls, which affected trust and understanding of appropriate system use, and many repeatedly asked for guidance or demonstrated a lack of sound propagation and detection knowledge. For JITT, we identified a common scaffolding that includes KSAs related to setup, teardown, repair, and servicing; considerations such as endurance, terrain, and weather; operations such as controls, mission profiles, networking, and communications; and tactics, techniques, and procedures involving safety, troubleshooting, and survivability via detection avoidance.

Mental Models for Robotics Warfare Specialist Rating and Machine Learning Training

Compass

The adoption of agentic AI is a high-priority topic for the Department of War, offering the potential to dramatically accelerate the development of complex systems like Command and Control (C2). However, this potential is balanced by significant risk. The core question for Naval leadership is no longer if to use agentic AI, but how to control it. Un-governed agents introduce profound security, supply chain, and reliability risks, from model "hallucination" and prompt injection vulnerabilities to insecure code generation. These risks are unacceptable for mission-critical systems that demand the high levels of trust central to the NAML mission. 

This presentation details a "dual-control" DevSecOps framework that provides the necessary governance to safely harness agentic AI. This pattern is aligned with the DoW DevSecOps Reference Design and lessons from initiatives like Platform One and NIWC's Overmatch Software Armory. First, we govern the agent's inputs by making it context-aware. The agent is constrained by ingesting a rich, domain-specific context from hardened, pre-configured application templates that programmatically codify DoW-mandated security controls (e.g., NIST 800-53). Second, we govern the agent's outputs by subjecting all AI-generated code to a rigorous, automated CI/CD pipeline. This pipeline validates the agent's output via policy-as-code enforcement (e.g., OPA/Kyverno), comprehensive security scanning, and GitOps promotion gates. 

The auditable result of this governed framework is a complete, machine-readable compliance package. The pipeline automatically generates: 1) A Software Bill of Materials (SBOM) , enabling continuous supply chain vulnerability monitoring per EO 14028. 2) A CI-generated Body of Evidence (BoE) , which provides the verifiable artifacts and evidentiary foundation required to rapidly author the formal RMF documents (SSP, SAR, POA&M). This approach is portable across Naval cloud environments and disconnected tactical edge targets like ships and autonomous platforms. 

This "dual-control" framework provides a tangible methodology for building trust in AI-generated systems. It directly addresses the RMF/ATO bottleneck as a byproduct of solving the more critical AI governance problem. This enables the technical foundation for Continuous Authorization to Operate (cATO) , allowing the DoN to move trusted C2 and AI-driven capabilities from "lab-to-fleet" at the speed of relevance.

Governing Agentic AI: A DevSecOps Control Framework for Trusted Naval C2 Systems

As naval operations increasingly rely on machine learning for combat decisions, sensor fusion, and autonomous systems, vendor lock-in threatens the Navy's ability to adapt ML capabilities across evolving threats and platforms. This presentation examines strategies for maintaining ML platform independence in Department of War procurement, addressing unique challenges in deploying machine learning systems for naval applications.

Machine learning platforms create dependencies through proprietary training frameworks, custom model formats, and specialized inference environments. Naval ML applications face amplified risks: classified training data, edge deployment on vessels with limited connectivity, real-time inference requirements for weapon systems, and mission-critical reliability constraints make vendor switching prohibitively expensive. When ML models drive targeting decisions, threat detection, and autonomous navigation, platform dependencies compromise operational flexibility and technological sovereignty.

We provide actionable strategies proven in naval ML deployments. Technical approaches include containerization for portable ML inference pipelines, ONNX for model interoperability across 1,700+ tools enabling seamless framework migration, MLflow for vendor-neutral experiment tracking across distributed teams, and Infrastructure-as-Code for reproducible ML training environments. Procurement reforms address modular contracting breaking ML projects into interoperable components, data rights ensuring government ownership of training datasets and model weights, and performance-based contracting focused on operational metrics rather than vendor-specific implementations. Implementation follows a phased roadmap validated at Naval Surface Warfare Centers.

Case studies demonstrate feasibility for naval ML systems. The Chief Digital and AI Office's Global Information Dominance Experiments deployed vendor-agnostic ML integration layers across all platforms. Naval Surface Warfare Centers successfully migrated computer vision models between cloud providers and edge devices. The Navy's "buying operational outcomes" approach prevents lock-in by procuring mission capabilities rather than specific ML frameworks.

We demonstrate that ML platform independence is both technically achievable and operationally essential for maintaining naval superiority in machine learning applications.

Premium content

Downloads

Next from NAML 2026

Course of Action Optimization, Analysis, and Comparison Harness (COACH)

Similar lecture

DART: Leveraging Multi-Agent Disagreement for Tool Recruitment in Multimodal Reasoning