United States

Large Language Model (LLM) agents increasingly rely on external tools to accomplish complex tasks, and how they scaffold tool invocations can influence both efficiency and reliability of the agent. 

The classic Model Context Protocol (MCP) requires an LLM agent to emit a structured JSON “tool‑call” that is parsed by an intermediary before execution. Code‑Mode, by contrast, lets the LLM write short programs that call typed MCP bindings inside a sandboxed runtime; because models are trained on extensive code corpora, this approach may allow agents to compose multiple tool calls more naturally and with fewer tokens. 

A second dimension—inter‑agent communication—has received comparatively little systematic study. In hierarchical or multi‑agent settings, an agent often delegates subtasks to subordinate agents that in turn invoke tools. We examine three inter‑agent communication modalities: 

- Direct: Code‑Mode calls between agents’ MCP bindings. 
- Classic: MCP messages that encode tool requests. 
- Natural‑language: queries that the receiving agent must interpret before acting. 

To compare these modalities, we developed a controlled evaluation measuring three criteria: 
1. Token efficiency — tokens consumed per completed workflow step. 
2. Reliability and error rate — frequency of malformed calls or execution failures. 
3. Coordination overhead — number of round‑trips and latency incurred by the communication method. 

This poster will present a systematic benchmark for assessing whether Code‑Mode’s theoretical advantages translate into practical gains in token efficiency and inter‑agent coordination. Both Code‑Mode and MCP are publicly documented; researchers can replicate our framework to evaluate alternative tool‑invocation strategies in their own hierarchical LLM agents.

NAML 2026

Code‑Mode vs MCP: Token Efficiency and Inter‑Agent Communication in Hierarchical LLM Agents

hierarchical llm agents

inter‑agent communication

code‑mode

model context protocol (mcp)

It is my great pleasure and honor to welcome you to the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), held in San Diego, California, USA, July 2–7, 2026. ACL 2026 continues our fields tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.

It is exciting to see the many developments in the field reflected at our conference, such as the advancement of large language models (born out of our community) into many application directions. This is not without challenges, partly due to unsolved problems, and partly due to misuse of the technology. All this is reflected well at ACL 2026. The conference also reflects that the field is not only about advancing technology, but also the science of fundamentally understanding language and how it is use by many different speaker communities. 
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) 
<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://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing')">Conference Handbook</button></html>

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

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.aclweb.org/registration/) first.

Registration Required

ACL 2026

The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) will take place in San Diego, California, from July 2 - 7, 2026.

poster

### 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

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. 



Defense operations increasingly demand effective human-AI teaming under conditions that severely strain cognition: simultaneous information streams, time-critical decisions with incomplete data, and sustained operations inducing fatigue and stress. Current AI systems face a critical limitation: they model operator behavior without accounting for how cognitive state affects that behavior. When an intelligence analyst revisits the same data repeatedly or a UAV operator exhibits degraded response times, existing systems cannot distinguish intentional strategy shifts from cognitive overload. This ambiguity undermines AI assistance—systems may present additional information when operators lack capacity to process it, or fail to intervene when fatigue compromises decision quality.

We have developed HiVAE, a hierarchical variational autoencoder inspired by the belief-desire-intention structure of human Theory of Mind (accepted to AAAI 2026 Theory of Mind Workshop). By structuring latent variables to mirror the causal dependencies in cognitive ToM models—where beliefs inform desires, which inform intentions—our three-level BDI hierarchy substantially outperforms non-hierarchical baselines on goal inference tasks, achieving significantly lower prediction error and maintaining robust performance when agent preferences shift. Ablation studies confirm this performance advantage is directly attributable to the hierarchical BDI organization itself. This work demonstrates that embedding ToM-motivated causal structure into deep architectures yields measurable gains in behavioral prediction. Our ongoing work extends this foundation by integrating physiological monitoring to capture how affective states modulate the mental states we can now infer.

This capability enables a critical extension for operational contexts: Affective Theory of Mind (AToM), which integrates continuous physiological monitoring to model how stress, arousal, and fatigue modulate inferred mental states. By incorporating biosensor data (heart rate variability, electrodermal activity, movement patterns) alongside behavioral observations, AToM would distinguish behaviors driven by mental state changes from cognitive resource depletion. We propose a hybrid architecture where deep learning models process physiological streams to generate affect estimates, which condition large language model inference of mental state hypotheses for Bayesian inverse planning. This would enable AI teammates that adapt support to operator cognitive capacity—recognizing when narrowed analyst focus reflects overload rather than strategy, or when degraded UAV operator performance signals need for handoff rather than increased autonomy.


Affective Theory of Mind: Physiologically-Informed Mental State Inference for Human-Machine Teaming

This project develops a system for remotely launching unmanned aerial vehicles (UAVs) and autonomously executing waypoint-based flight plans around a moving target. As the target object translates or rotates, the UAV continuously updates its trajectory by shifting each planned waypoint based on real-time target motion data. Target tracking is achieved either through GPS stations mounted on the target or via an in-house computer vision model capable of estimating target pose. The UAVs were custom-built in the RESTORE laboratory and the autonomy was implemented using ROS 2 with pymavlink communication. The GPS-based system has been successfully executed as two UAVs were remotely launched from a location hundreds of miles away and autonomously circumnavigated a designated target while capturing imagery at defined intervals and transmitting it to a remote server in real time. The computer vision–based tracker has been validated independently and is currently being integrated into the full pipeline.

Custom UAVs for Remote Autonomous Launch and Moving Target Tracking

As large language model (LLM)-powered software is given greater freedom to act autonomously via tool calling and agentic frameworks, the need to monitor its behavior in order to minimize unintended actions continues to grow. The Multimodal Sensemaking in SeaVision (MUSES) project at NIWC LANT has used several open-source tools to test and evaluate LLM's and associated tools that are part of our software. MUSES is also researching using a separate LLM framework to continuously monitor and analyze T&E results to enable autonomous self-correction when our systems fail or otherwise behave unexpectedly. This poster presents an overview of our T&E work using several publicly releasable case studies.

Open-source tools for LLM T&E in a maritime awareness domain

Autonomous systems need environmental cues for orientation, navigation, and specialized tasks. Interpreting the layout of a space can be done by detecting objects and reading signs, which lead to enhanced scene understanding like identifying and assessing equipment in large spaces. We propose a pipeline for detecting, reading, and documenting relevant signs within a ship to aid in shipboard navigation and inventory tasks. This pipeline distinguishes between important Navy shipboard signage, including placards, exit signs, configuration labels, and bullseyes. Once the signs have been detected and classified, they are fed into specialized optical character recognition algorithms to read and decipher text according to the automated task. Finally, the sign classes and transcriptions are corrected via post-processing and mapped into a corresponding point cloud to form an inventory of the target space. This unique, proposed system comprises algorithms to be used in-situ for a set of inventorying and navigation tasks.

Shipboard Optical Character Detection And Recognition For Scene Understanding

In the shallow ocean, the seabed significantly influences sound propagation. Sound reaches and interacts with the seafloor, where seabed properties determine the amount of sound that is reflected or absorbed. Proper modeling of the properties of the seabed, then, is crucial for accurate source localization. While passive sonar inference of seabed parameters has often been accomplished using Bayesian optimizations, a deep learning approach may be an effective approach that simplifies the process in the long term. To begin evaluating the accuracy of this strategy, a 1D CNN is trained to perform the task of estimating seabed properties. The input data are on single frequency sound levels as a function of range from a towed source. Synthetic training data generated by a range-independent normal mode model for ocean environments corresponding to a wide range of seabed properties are used to train the CNN, which is then applied to measured data collected northeast of Scotland. A towed source produced the signal, which was recorded on a bottom-moored hydrophone. The recorded signal was then processed to obtain single-frequency power spectral density levels as a function of range. These data are input to the trained CNN and to a Bayesian optimization, which returns the posterior probability distributions (PPDs) for these parameters. A comparison between the CNN results and the PPDs is used to assess efficacy of the CNN in this generalization test and point to potential improvements in the deep learning approach, including future application of Bayesian networks. This search for deep learning approaches for the estimation of seabed properties has important implications for source detection, ranging, and tracking in the shallow ocean.

Evaluating Deep Learning for Seabed Properties Using Bayesian Optimization

Modern Navy maritime operations increasingly rely on advanced decision-aid tools and large-scale data analytics within Mission Operations Centers (MOC). However, a significant accessibility gap remains between these complex tools and MOC watchstanders who may lack domain-specific technical training. This work introduces an Artificial Intelligence (AI)-driven Control Flow framework that enables watchstanders to readily generate maritime network topology visualization and analysis using a custom-built network simulation tool named TOMO-Sim, through natural language, chat-style interactions. This project implements an agentic AI assistant designed to bridge the usability gap by allowing intuitive, plain-language communication with TOMO-Sim. Leveraging a state-of-the-art Large Language Model (LLM) orchestrated through the LangGraph python Library, the AI assistant interprets the user intent, gathers required parameters for network configuration, and autonomously interacts with the TOMOSim APIs to produce network topology visualizations. This approach providing the MOC watchstanders with simplified, context-aware insights, and maintaining a conversational workflow that enables intuitive, iterative refinement of the generated network models. The resulting workflow also offers a modular blueprint for integrating more advanced, future technical decision aids within an agentic AI framework.

Agentic AI Assistant for Maritime Network Topology Simulation

The OLIVE multi-media processing framework is in wide use across military and intelligence community applications. Capabilities include speech recognition, voice and face biometrics, translation, video and imagery summarization, language detection, deep fake audio and video detection and multiple other special purpose technologies. This presentation will focus on the integration of large language models (LLM) as both a source of natural language processing to extract information from, for example, automatically generated speech transcripts and visual captions to provide tools such as summarization, named entity recognition, sentiment, etc. and to post-process raw metadata on a massive scale to enable big-picture understanding of complex data streams. We will discuss the challenges and benefits of using LLMs in the processing of ISR and other multi-media data types, including the difficulties inherent in producing a reliable textual representation of non-textual data, strategies for prompting, limiting and constraining outputs, difficulties with completeness of coverage, etc. We also explore LLMs’ translation capabilities, especially mixed language input translations, which have long posed a problem for traditional AI systems. Finally, we look at the future of massively multi-modal models that heretofore have not yielded operationally-viable results but which clearly reflect the future of AI in multi-media processing.

Integration of Large Language Model Capabilities in a Media and ISR Processing Framework.

Artificial intelligence (AI) holds great promise in accelerating information processing in ways that can potentially help users make high-consequence decisions rapidly in operational environments. A key practical challenge is determining how to embed AI into command and control (C2) systems, especially those that are disconnected and have limited processing power. While large language models (LLMs) have demonstrated strong performance across various reasoning tasks, their size and computational requirements limit their suitability for deployment at the edge. The rapid growth of open-source small language models (SLMs) offers new opportunities for edge AI, but selecting the most appropriate model remains challenging, as standard benchmarks often fail to capture domain-specific tasks. This work reported here examines the viability of SLMs as support tools that enable users to make rapid decisions at the edge. As part of an Office of Naval Research-sponsored effort, we developed a novel evaluation framework combining insights from interviews with subject matter experts, workflow analyses, and guidance from naval doctrine to construct synthetic datasets and benchmarks representative of operational needs. 

We applied this framework to decision-making processes within the domain of operational law to evaluate the performance of SLMs in helping deployed commanders adhere to domestic and international regulations. While legal experts maintain the ultimate responsibility and authority in reviewing SLM outputs and making decisions, our results reveal benefits of SLM support for legal information retrieval and rules of engagement evaluation. 

These findings demonstrate that, with appropriate guardrails, SLMs provide a practical means to enhance decision support in resource-constrained C2 systems, improving both speed and reliability of critical actions. Finally, the proposed framework is generalizable to other mission domains and supports the development of AI-enabled C2 capabilities at the tactical edge. The poster will include illustrative workflow examples and SLM-assisted compliance checks to highlight practical operational applications.

Evaluating the Efficacy of Small Language Models as Decision Support Tools at the Edge

Sound propagation in the ocean depends on the parameterization of the seabed. The physical seabed properties are correlated and impact the data in different ways. To quantify the information content about the seabed properties in measured data, three tools for information geometry are used. These three methods are illustrated using bottom loss. Bottom loss represents the reflection coefficient of the seabed on a logarithmic scale and depends on both frequency and grazing angle. An analytical, multi-layered bottom loss model is used where the seabed layers are parameterized by their sound speed and density. Some combinations of layer properties produce similar bottom loss results, even if the specific values of their parameters are different. This ambiguity is investigated by applying information geometry to the model manifold: bottom loss as a function of frequency and grazing angle over a range of seabed properties. The resulting 2d model manifold is embedded in a data space with dimensions equal to the number of frequencies by the number of grazing angles. The edges of the model manifold correspond to bottom loss at the extreme values of one of the seabed parameters. To visualize this high dimensional data space, three information geometry techniques are used. For the first method, we use the Euclidean length of the edges of the manifold to indicate how much a parameter contributes to the overall model, which is referred to as “sloppy” or “stiff” parameters. The second method is to perform a Principal Component Analysis on the model manifold. Generally the first two principal components capture more than 90% of the model’s information and provide a simple way to evaluate the main sources of variation on the model manifold. The final method is to examine the Fisher Information Matrix (FIM). The FIM gives a quantitative measure of how much information a model carries about each parameter. To find the FIM, we perform automatic differentiation on the model for specific seabed properties. Using the FIM, we can find the conditional variance, Cramer-Rao bounds, and parameter correlations which can be examined to find how much information a parameter carries about the model. These information geometry techniques can be applied broadly to many types of models to quantify information content.

Information Geometry and Seafloor Bottom Loss

In this work, we present a multi-modal autonomous docking framework designed to operate robustly under unknown and changing dynamic conditions. The system supports three high-level operating modes: 1) human-in-the-loop supervision, 2) shared or assisted control, and 3) full autonomy, allowing seamless transitions across levels of operator involvement to support maritime operations 

Our approach integrates a deep-learning-based Koopman operator to obtain an accurate higher-dimensional linear representation of the vessel dynamics, enabling predictive control through an adaptive MPC formulation. Model mismatches and dynamic changes are handled using recursive least squares updates. Safety is enforced through a differentiable optimization–based Control Barrier Function (CBF) filter, which ensures real-time safe velocity commands across all modes of operation. 

Dock detection and localization are performed using multiple AprilTags, providing fast and reliable estimation of the docking structure and nearby obstacles. A simulation environment built in ROS and Gazebo enables systematic and scalable evaluation under a wide range of conditions. 

The proposed multi-modal control architecture is validated in simulation and through field experiments. Results indicate clear quantitative and qualitative improvements compared to baseline human drivers and alternative control strategies, demonstrating effective integration of differentiable CBFs with deep Koopman-based adaptive MPC for safe maritime docking.

Premium content

Next from NAML 2026

Affective Theory of Mind: Physiologically-Informed Mental State Inference for Human-Machine Teaming

Similar lecture

MCP-Flow: Facilitating LLM Agents to Master Real-World, Diverse and Scaling MCP Tools