Port-LLM: Maritime & Wireless Insights

• Port-LLM is a research label for LLM-grounded frameworks that integrate structured operational data with natural language explanations in both maritime and wireless contexts.
• It leverages advanced models such as Temporal Graph Attention Networks and LoRA-adapted GPT architectures to generate accurate, auditable predictions from AIS data and channel forecasts.
• Applications include day-ahead port congestion forecasting with 99.6% directional consistency and fluid antenna port selection achieving NMSE below -25 dB, highlighting its practical operational impact.

“Port-LLM” is not a single standardized model name in the recent literature. In maritime AI, it most specifically denotes an LLM-grounded framework for interpretable port congestion prediction in which AIS-derived daily spatial graphs are processed by a Temporal Graph Attention Network and then translated into faithful natural-language risk reports by a constrained LLM module (Xue et al., 5 Mar 2026). The same label, or a closely related “Port-LLM” perspective, also appears in work on container throughput forecasting, autonomous maritime inspection, and other port-operation workflows (Kim et al., 24 Feb 2026, Din et al., 19 Jan 2026). In a separate line of wireless-communications research, “Port-LLM” refers to fluid-antenna port prediction rather than maritime ports, using pretrained LLMs to forecast future channel tables and choose movable antenna ports (Zhang et al., 14 Feb 2025).

1. Terminology and scope

The term has been used across multiple technical contexts. This suggests that “Port-LLM” functions more as a research label than as a single architecture or benchmark.

Usage in literature	Representative system	Core task
Interpretable maritime congestion prediction	AIS-TGNN (Xue et al., 5 Mar 2026)	Daily cell-level congestion-escalation prediction and faithful explanation
Port-logistics forecasting	PK-TimeLLM (Kim et al., 24 Feb 2026)	Container throughput forecasting with contextual prompting
Autonomous maritime inspection	LLM–VLM fusion framework (Din et al., 19 Jan 2026)	Symbolic planning and semantic inspection with UAV–USV teams
Fluid antenna port prediction	GPT-2/LoRA Port-LLM (Zhang et al., 14 Feb 2025)	Forecast future channel tables and select FA ports

Within maritime logistics, the most technically developed use of the exact label is the AIS-TGNN framework for explainable congestion monitoring. There, Port-LLM is explicitly a dual-objective system: it predicts congestion escalation at the daily cell level and generates faithful, auditable explanations grounded in model-internal evidence rather than free-form narration (Xue et al., 5 Mar 2026).

Other maritime papers use the name more broadly to denote LLM-centered port-intelligence systems. In that broader sense, Port-LLM spans symbolic mission planning, contextual forecasting, and decision-support agents, but these systems do not share a single model backbone, data modality, or operational target (Din et al., 19 Jan 2026, Kim et al., 24 Feb 2026).

2. AIS-TGNN: Port-LLM as interpretable congestion prediction

In the congestion-prediction formulation, the operational problem is day-ahead identification of congestion escalation at major maritime hubs such as Los Angeles and Long Beach, where congestion propagates across berths, anchorages, and inland networks. The stated need is not only forecasting accuracy but also cell-level explanations that non-technical stakeholders can trust, because traditional systems either treat locations independently or rely on opaque graph models, while post-hoc attribution methods remain numeric and unguided LLMs are not necessarily faithful to model evidence (Xue et al., 5 Mar 2026).

The data source is NOAA Marine Cadastre AIS broadcasts from January–June 2023 over the San Pedro Bay anchorage and the Port of Los Angeles/Long Beach, with the region discretized into a regular $0.1^\circ \times 0.1^\circ$ grid, approximately $11\text{ km} \times 11\text{ km}$ per cell. The system processes 89 valid daily snapshots $G_1,\dots,G_{89}$, amounting to approximately $3.02\times 10^4$ node–day samples. For each day $t$, active nodes $V_t$ are grid cells with at least one AIS observation, node features $X_t \in \mathbb{R}^{|V_t|\times 10}$ are aggregated per cell and z-score normalized snapshot-wise across active nodes, and edges $E_t$ are induced by a fixed-topology $k$-nearest-neighbor graph with $k=8$ based on centroid Euclidean distance. The binary label is congestion escalation at node $11\text{ km} \times 11\text{ km}$0 for day $11\text{ km} \times 11\text{ km}$1, defined as $11\text{ km} \times 11\text{ km}$2 if $11\text{ km} \times 11\text{ km}$3, else 0, with positive rate approximately $11\text{ km} \times 11\text{ km}$4 (Xue et al., 5 Mar 2026).

The feature set consists of mean speed over ground, standard deviation of speed over ground, slow-vessel ratio, anchor ratio, vessel count, cargo ratio, tanker ratio, mean vessel length, mean draft, and circular variance of course over ground. These are chosen to represent localized vessel activity rather than exogenous operational signals. A stated limitation follows directly: the model is single-modality kinematics and does not incorporate weather, terminal schedules, labor actions, or trucking capacity (Xue et al., 5 Mar 2026).

The predictor is a Temporal Graph Attention Network that injects temporal information through prior embedding concatenation rather than an explicit recurrent module. For node $11\text{ km} \times 11\text{ km}$5, the initial state is

$11\text{ km} \times 11\text{ km}$6

where $11\text{ km} \times 11\text{ km}$7 is the node embedding produced at the previous day and is initialized to zeros at $11\text{ km} \times 11\text{ km}$8. Multi-head attention aggregates neighbor information on each daily snapshot, and a two-layer MLP maps the final node embedding to a congestion-escalation logit. For attention head $11\text{ km} \times 11\text{ km}$9,

$G_1,\dots,G_{89}$0

and the layer update is

$G_1,\dots,G_{89}$1

Training uses class-weighted binary cross-entropy with $G_1,\dots,G_{89}$2, hidden dimension $G_1,\dots,G_{89}$3, attention heads $G_1,\dots,G_{89}$4, dropout $G_1,\dots,G_{89}$5, Adam with learning rate $G_1,\dots,G_{89}$6 and weight decay $G_1,\dots,G_{89}$7, up to 50 epochs, early stopping by validation AUC, and a decision threshold $G_1,\dots,G_{89}$8 selected to maximize validation F1 (Xue et al., 5 Mar 2026).

Under a strict chronological split—train snapshots 1–62, validation 63–75, test 76–89—the TGAT model achieves test AUC $G_1,\dots,G_{89}$9, AP $3.02\times 10^4$0, F1 $3.02\times 10^4$1, and recall $3.02\times 10^4$2 on 4,531 node–day test samples. The LR baseline attains AUC $3.02\times 10^4$3, AP $3.02\times 10^4$4, and recall $3.02\times 10^4$5; the GCN baseline attains AUC $3.02\times 10^4$6, AP $3.02\times 10^4$7, and recall $3.02\times 10^4$8. TGAT reaches validation AUC $3.02\times 10^4$9 by epoch 12 and maintains an advantage over LR and GCN (Xue et al., 5 Mar 2026).

3. Grounded explanation, prompting constraints, and validation

The defining feature of Port-LLM in this interpretation is that the LLM is not the predictor. It is a constrained explainer that converts graph-model evidence into faithful narratives. For each node–day in the test set, the pipeline records predicted probability $t$0, raw attention logits $t$1, attention-proxy neighbor influence, feature z-scores, and dataset-level point-biserial correlation directions $t$2. Neighbor influence is constructed by summing logits across heads and applying a softmax over neighbors: $t$3 Feature normalization is snapshot-wise,

$t$4

with $t$5 and $t$6 computed across active nodes in the daily snapshot (Xue et al., 5 Mar 2026).

The prompting strategy is explicitly structured and guardrailed. The prompt contains a header with cell ID, date, and predicted probability; the top-5 features with names, z-scores, correlation directions, and $t$7 values; the top-2 neighbors with attention weights and each neighbor’s most prominent feature; and a strict JSON output schema with six sections: target feature drivers, neighbor influence, risk summary, counterfactual suggestions, confidence or uncertainty, and limitations. The system prompt instructs the model to “Use only the evidence provided; do not invent external facts; align stated directions with z-scores and correlation signs.” The LLM used for explanation is GPT-4o-mini (Xue et al., 5 Mar 2026).

This design is intended to eliminate the usual ambiguity between attribution and narration. The narrative surface is generated by an LLM, but the semantic content is bound to verifiable model outputs. Operationally, that means every explanation can be traced back to normalized features, correlation signs, and attention-weighted neighbor influence rather than external knowledge or latent speculation (Xue et al., 5 Mar 2026).

Explanatory reliability is measured by a directional-consistency protocol. For a set of claims $t$8 about features $t$9, each with stated direction $V_t$0, agreement is

$V_t$1

In the reported evaluation, 100 reports with 5 feature claims each yield 500 judgments, of which 498 are consistent, giving $V_t$2 directional consistency. The two inconsistencies involve near-zero-correlation features such as tanker ratio with $V_t$3, where hedged language confused the parser. The result is presented as evidence that the LLM adhered to evidence constraints rather than hallucinating (Xue et al., 5 Mar 2026).

4. Port-LLM across maritime port operations

A broader maritime interpretation of Port-LLM appears in container throughput forecasting. In that work, Port-LLM refers to an LLM-centered forecasting paradigm instantiated as PK-TimeLLM, based on TimeLLM with a frozen Qwen-3B decoder, a cross-modal alignment layer, and a port-logistics knowledge prompt. The prompt includes Data Description, Task Description, Domain Information, dynamic Berth Schedule text, and Auxiliary Information such as holidays and weather. Across 18 settings on two years of Busan Port terminal data, PK-TimeLLM ranks first in 8 of 9 settings on Dataset 1 with average improvement of $V_t$4 over the second-best model, and first in all 9 settings on Dataset 2 with average improvement of $V_t$5 over the next-best model. The reported interpretation is that context supplied semantically through prompt text is more effective than flattening the same context into multivariate numeric inputs (Kim et al., 24 Feb 2026).

A second maritime use of the label denotes an LLM–VLM fusion framework for autonomous port inspection with heterogeneous UAV–USV teams. There, the LLM replaces traditional state-machine mission planners by mapping natural-language mission instructions and operational rules into symbolic plans with preconditions and dependencies, while the VLM performs semantic inspection and compliance assessment. In the reported planning benchmarks, GPT-4o attains average Correctness $V_t$6, Execution Success Rate $V_t$7, and Response Time $V_t$8, outperforming GPT-3.5-Turbo, GPT-4, Gemini, and LLaMA. For semantic inspection, Qwen2-VL reports semantic correctness approximately $V_t$9–$X_t \in \mathbb{R}^{|V_t|\times 10}$0 with inference time approximately $X_t \in \mathbb{R}^{|V_t|\times 10}$1, while Moondream2 reports approximately $X_t \in \mathbb{R}^{|V_t|\times 10}$2–$X_t \in \mathbb{R}^{|V_t|\times 10}$3 with approximately $X_t \in \mathbb{R}^{|V_t|\times 10}$4 inference time (Din et al., 19 Jan 2026).

A third port-operations variant is PortAgent, an LLM-driven vehicle dispatching agent for Automated Container Terminals. PortAgent uses a Virtual Expert Team with four roles—Knowledge Retriever, Modeler, Coder, and Debugger—to transfer Vehicle Dispatching Systems across terminals without port operations specialists. In evaluation on 45 test instances of Multi-AGV Path Planning, it reports CER $X_t \in \mathbb{R}^{|V_t|\times 10}$5, overall SSR $X_t \in \mathbb{R}^{|V_t|\times 10}$6 corresponding to 42 out of 45 successful cases, and average end-to-end deployment time approximately $X_t \in \mathbb{R}^{|V_t|\times 10}$7. Ablation results show that removing RAG reduces CER to $X_t \in \mathbb{R}^{|V_t|\times 10}$8 and SSR to $X_t \in \mathbb{R}^{|V_t|\times 10}$9, while removing self-correction yields CER $E_t$0 and SSR $E_t$1 (Hu et al., 16 Dec 2025).

Related maritime LLM systems broaden the same design space even where the exact name differs. VTS-LLM formulates risk-prone vessel identification as knowledge-augmented Text-to-SQL with NER-based relational reasoning, semantic algebra intermediate representation, and query rethink; it reports $E_t$2 on operational style and shows marked sensitivity to linguistic style variation (Sun et al., 2 May 2025). MSD-LLM treats Port State Control detention prediction as a structured-representation-plus-LLM ranking problem and reports AUC $E_t$3 on Singapore ports versus $E_t$4 for DSRAE, with the stated gain of more than 12 percentage points on that domain-shifted test set (2505.19568). Collectively, these works indicate that in maritime research the Port-LLM idea centers on grounding LLM outputs in structured operational context rather than using free-form language generation as the primary predictive mechanism.

5. Wireless-communications usage: fluid antenna port prediction

A distinct research line uses “Port-LLM” for fluid antenna systems, where “ports” are movable radiating positions on the UE side rather than maritime facilities. In this formulation, the task is to predict future channel tables spanning all fluid-antenna ports and then select the port whose predicted channel minimizes deviation from a known reference channel. The original Port-LLM model is built on the smallest GPT-2 variant with feature dimension 768, uses only the first $E_t$5 layers, and applies LoRA to the multi-head attention $E_t$6 and $E_t$7 projections with rank $E_t$8, while all other weights are frozen (Zhang et al., 14 Feb 2025).

The architecture normalizes complex channel tables, splits them into real and imaginary parts, projects them to a 768-dimensional space, applies an 8-head temporal attention module, maps the result to the forecast horizon, and feeds it through the LoRA-adapted GPT-2 backbone. Training uses Adam, a warm-up-aided cosine learning-rate schedule, 500 epochs, history length $E_t$9, forecast horizon $k$0, and a dataset of 54,300 samples generated from a 3GPP TR 38.901 CDL-D channel. Evaluation spans UE speeds of 90, 120, and $k$1, with SISO training and MISO testing on $k$2, $k$3, and $k$4 BS arrays. The reported results are strong generalization from SISO to MISO without retraining, channel-table prediction NMSE below $k$5, validation NMSE approximately $k$6, and spectral-efficiency curves superior to Vec Prony and MPMP in medium and high mobility (Zhang et al., 14 Feb 2025). A later presentation of the same approach emphasizes that exact numerical NMSE values are not reported in that version but retains the same GPT-2-plus-LoRA design and the same qualitative superiority claims (Zhang et al., 1 Sep 2025).

Another wireless use of the term shifts from prediction to algorithm design. In “LLM-Enabled Automated Algorithm Design for Multiuser Fluid Antenna Communications,” LLM4AD/EoH is used to evolve heuristic search procedures for selecting active ports in a BS-FAS downlink under a max–min SINR objective. Two strategies are studied: LLM-optimized genetic-algorithm operators and AutoPort, a GRASP-like heuristic designed from scratch by the LLM. For $k$7, AutoPort achieves normalized performance $k$8 across $k$9–$k=8$0, matching exhaustive search; for $k=8$1, AutoPort reports normalized gains $k=8$2–$k=8$3 over Basic GA, and GA-CM reports $k=8$4–$k=8$5, while runtimes at $k=8$6 are $k=8$7 for AutoPort and $k=8$8 for GA-CM (Zheng et al., 14 May 2026). Here, “Port-LLM” no longer denotes sequence forecasting but offline LLM-driven heuristic synthesis.

6. Limitations, assumptions, and open directions

In the AIS-TGNN congestion framework, the main assumptions are sufficient AIS coverage, a fixed $k=8$9 grid as appropriate spatial granularity, snapshot-wise feature stationarity, and the sufficiency of short-term temporal conditioning via prior embeddings. The stated limitations are potential spurious correlations, domain drift across seasons or atypical events, label noise from slow-ratio day-over-day flips, limited temporal memory beyond one day, and the absence of exogenous signals such as weather, tides, ETAs, labor actions, and drayage capacity (Xue et al., 5 Mar 2026). The natural continuation is therefore multi-port transfer learning, domain adaptation across seasons, real-time streaming inference, multi-step horizons, richer causality-aware prompts, and human-in-the-loop evaluation (Xue et al., 5 Mar 2026).

Comparable limitations recur across the broader maritime Port-LLM ecosystem. PK-TimeLLM is computationally expensive, with substantial parameter cost and lower inference throughput than conventional time-series forecasters, while its advantage narrows at longer horizons because berth schedules and weather become more uncertain (Kim et al., 24 Feb 2026). The autonomous inspection framework remains sensitive to domain shift, adverse conditions, occlusions, and small-object detection; its own mitigation strategies are stronger VLMs, confidence thresholds, re-inspection, and closer vantage points (Din et al., 19 Jan 2026). PortAgent’s failures are primarily semantic misinterpretations of natural-language requirements, even when generated code is executable, which is why it couples role prompting with AST checks, sandbox execution, and a Reflexion-inspired self-correction loop (Hu et al., 16 Dec 2025).

A general misconception would be to treat Port-LLM as a single benchmark or product family. The literature instead shows several technically distinct patterns. In maritime ports, the recurrent theme is LLM grounding by structured operational evidence, whether through AIS-derived graph features, berth schedules, RAG-retrieved domain knowledge, or symbolic task descriptions (Xue et al., 5 Mar 2026, Kim et al., 24 Feb 2026). In wireless communications, the term refers to antenna-port prediction or LLM-designed port-selection heuristics rather than maritime decision support (Zhang et al., 14 Feb 2025, Zheng et al., 14 May 2026). A plausible implication is that future consolidation, if it occurs, will depend less on the shared name than on whether these systems converge on common principles of grounding, validation, and auditable deployment.