AviationLLM: Aviation-Specific Language Models

• AviationLLM is a suite of large language and multimodal models specialized for processing complex aviation data and supporting operational safety.
• It employs domain adaptation, multimodal fusion, and retrieval-grounded reasoning to integrate diverse inputs like cockpit imagery, voice, and technical manuals.
• Applications include cockpit advisory, flight training, maintenance analytics, and UAV operations, with robust metrics proving its practical efficacy.

AviationLLM denotes an emerging class of aviation-specialized LLM and large multimodal model systems that are adapted to aviation data, procedures, and operational constraints rather than treated as generic conversational models. In current literature, the label spans both concrete systems—such as the Qwen2.5-based training platform in "AviationLLM: An LLM-based Knowledge System for Aviation Training" (Wan et al., 17 Jun 2025)—and broader architectural blueprints such as "AviationLMM: A Large Multimodal Foundation Model for Civil Aviation" (Li et al., 14 Jan 2026). Adjacent systems including AviationGPT (Wang et al., 2023), the Virtual Co-Pilot for single-pilot operations (Li et al., 2024), LeRAAT (Schlichting et al., 5 Mar 2025), FRAMe (Tabrizian et al., 8 Jul 2026), and knowledge-grounded safety frameworks (Iyengar et al., 10 Apr 2026) collectively define the technical scope of the field: multimodal ingestion, retrieval-grounded reasoning, operational decision support, formal verification, and domain-specific evaluation.

1. Problem setting and conceptual scope

AviationLLM research is motivated by a mismatch between the heterogeneity of aviation data and the narrowness of most existing aviation AI pipelines. Civil aviation produces voice communications, radar tracks, on-board telemetry, video, structured texts, maintenance logs, and regulatory documents, yet many deployed AI systems remain siloed by task or modality and therefore provide limited cross-modal situational awareness (Li et al., 14 Jan 2026). The same literature identifies fragmentation, limited generalization, trust and safety, missing modalities, deployment constraints, and the absence of a domain-tailored any-to-any I/O system embedding aviation knowledge as central gaps in the current landscape (Li et al., 14 Jan 2026).

A second driver is the operational shift toward more demanding human-machine configurations. In single-pilot and remote operations, the workload and risk profile intensify, which motivates a Virtual Co-Pilot capable of integrating pilot instruction and real-time cockpit instrumental data to retrieve applicable manuals and procedures (Li et al., 2024). In low-altitude operations, the rapid growth of the low-altitude economy introduces additional requirements for intent recognition, routing, communication control, and natural-language navigation for UAVs (Lei et al., 8 Sep 2025). In training and business operations, generic web answers are described as insufficiently accurate for professional aviation theory or regulated operational knowledge, prompting work on aligned, retrieval-grounded aviation assistants and dedicated benchmarks (Wan et al., 17 Jun 2025).

The domain also has a data-structure problem. AviationGPT explicitly frames aviation as a domain with abundant complex, unstructured text, extensive technical jargon, and scarce labeled data, and presents domain adaptation of LLaMA-2 and Mistral as a response to that gap (Wang et al., 2023). This suggests that AviationLLM is best understood not as a single model family, but as a research program centered on domain adaptation, multimodal fusion, grounding, and verification under aviation-specific reliability constraints.

2. Architectural patterns

The most expansive architectural formulation appears in AviationLMM, which defines an Encode $\rightarrow$ Align $\rightarrow$ Fuse $\rightarrow$ Decode pipeline for civil aviation (Li et al., 14 Jan 2026). Its encoder bank includes WavLM-style audio encoding, Vision-Transformer video encoding with masked-video pretraining, a DINOv2 image backbone, an instruction-tuned LLM text encoder, a Transformer-based time-series encoder for FDR/HUMS and surveillance streams, and a 3D-CNN for volumetric radar or LiDAR. Alignment is performed through temporal-spatial synchronization, ImageBind-style joint embedding, ontology-guided alignment, and reliability-aware alignment; fusion is implemented through cross-attention Transformer blocks, graph-based fusion, hierarchical fusion, contrastive fusion, and mixed-query fusion; output modules span audio, video, image, text, series, and other sensor reconstructions (Li et al., 14 Jan 2026).

A more concrete multimodal cockpit architecture is given by the Virtual Co-Pilot. That system uses GPT-4 with a curated aviation knowledge base and exemplar dialogues, ingesting cockpit images, pilot voice transcriptions, and preloaded manuals such as checklists and the Quick Reference Handbook. The multimodal fusion step is written as

$z = W_t \cdot x_t + W_v \cdot x_v + b,$

after which the fused context is used in a retrieval-augmented prompt that points GPT-4 to indexed aviation manuals (Li et al., 2024). Retrieval proceeds by cosine similarity over a precomputed embedding index, with optional reranking and vector-database support such as FAISS (Li et al., 2024).

AviationLLM systems are not confined to a single centralized backbone. Aero-LLM distributes specialized models across an on-board, edge, and cloud hierarchy: ultra-lightweight on-board OPT-125M and OPT-350M models for immediate packet-level inferencing and anomaly flags, mid-sized edge models including OPT-1.3B and OPT-6.7B plus TimesNet, and cloud-side Llama-2-7B, Llama-2-13B, and Time-LLM for long-horizon analytics and retraining (Dharmalingam et al., 5 Feb 2025). A distinct hierarchical design appears in the integrated terrestrial and non-terrestrial network setting, where a HAPS-deployed LLM handles meta-control and on-board UAV LLMs handle motion planning, telecommunication decisions, and XML-style action emission such as <meta_action>Offload{4}</meta_action> or <tran_action>FASTER</tran_action> (Yan et al., 6 Jun 2025).

Several systems further decompose architecture by function rather than modality. FRAMe combines a planner LLM, a RAG-based memory, and a multi-modal coach agent for end-to-end flight planning (Tabrizian et al., 8 Jul 2026). The low-altitude natural-language navigation framework combines an NL$\rightarrow$STL translator, an STL-constrained MILP planner, and a specification repair module (Ping et al., 29 Mar 2026). Across these variants, the recurring pattern is modular composition: perception and encoding, retrieval or memory, LLM reasoning, and an external mechanism for verification, optimization, or repair.

3. Grounding, alignment, and formal reasoning

AviationLLM systems rely heavily on grounding because parametric knowledge alone is treated as insufficient for aviation reliability. The training-oriented AviationLLM system based on Qwen2.5-14B-Instruct uses Retrieval-Augmented LLM Alignment via Direct Preference Optimization, or RALA-DPO. At inference, a retrieval encoder computes a query embedding, cosine similarity is used against a vector database of knowledge fragments, and top-$k$ retrieved segments are prepended to the user query in a late-fusion prompt (Wan et al., 17 Jun 2025). The alignment stage uses DPO on a preference dataset of 9,740 expert-annotated pairs, with loss

$$L_{DPO}(\theta) = - \sum_{(x,y_w,y_l)\in D}\log \sigma [ \beta \Delta(x; y_w,y_l) ],$$

where $\Delta$ is defined by log-ratio differences between the trainable policy and a frozen reference policy (Wan et al., 17 Jun 2025). The same paper treats zero-cost knowledge updates as a retrieval-layer operation: new regulations or manuals are segmented, embedded, and added to the index without retraining (Wan et al., 17 Jun 2025).

Knowledge grounding is also pursued through graph-centric methods. The framework in "Building Trust in the Skies" constructs an Aviation Safety Knowledge Graph from incident reports, flight-data logs, maintenance manuals, and regulatory PDFs, then uses a GraphRAG phase in which an LLM translates user queries into Cypher, retrieves a subgraph from Neo4j, and conditions a generator on those retrieved facts (Iyengar et al., 10 Apr 2026). Retrieval is scored as

$$\mathrm{score}(q,SG) = \alpha\;\mathrm{cosine}(e_q,e_{SG}) + \beta\;\mathrm{conf}(SG),$$

and every generated assertion is linked to supporting triple IDs, after which a validator re-executes cited triples to check grounding (Iyengar et al., 10 Apr 2026).

Other systems implement grounding as operational memory or as formalization. LeRAAT uses a relay server that fuses real-time X-Plane aircraft state, ECAM text, weather, nearby airport information, and top-10 FAISS-retrieved manual chunks before querying GPT-4o (Schlichting et al., 5 Mar 2025). FRAMe stores scenario geometry identifiers, preference embeddings, candidate plans, validity and alignment verdicts, optional feedback, and planner reasoning in a RAG-based memory, then retrieves preference-nearest historical plans for the same geometry before the next planning call (Tabrizian et al., 8 Jul 2026). In the low-altitude UAV navigation framework, grounding takes a more formal form: a reasoning-enhanced LLM translates free-form natural language into Signal Temporal Logic, GRPO improves syntactic validity and semantic fidelity, and a MILP-based repair mechanism selectively relaxes non-safety-critical predicates or time bounds when the original specification is infeasible (Ping et al., 29 Mar 2026).

These designs show two distinct but complementary notions of grounding. One grounds generation in external evidence stores, whether vector indexes or knowledge graphs. The other grounds generation in executable formalisms such as STL and MILP. A plausible implication is that AviationLLM research increasingly treats trustworthy output as a property of the entire pipeline rather than of the LLM alone.

4. Major application domains

Cockpit advisory and quick-access procedure support form one of the earliest explicit AviationLLM use cases. The Virtual Co-Pilot targets automated quick procedures searching for single-pilot operations by combining pilot instruction with cockpit imagery and indexed manuals (Li et al., 2024). LeRAAT extends this pattern into simulator-integrated, real-time assistance, using X-Plane flight state, ECAM messages, weather, airport ranking, and RAG over Airbus A320NEO manuals, emergency procedures, and FAA directives to generate concise pilot-style advisories for either on-screen or VR display (Schlichting et al., 5 Mar 2025).

Training and aviation knowledge access constitute a second major cluster. The Qwen2.5-based AviationLLM system is designed for aviation theory training, where it answers professional questions using DPO alignment and RAG over regulatory manuals, theory texts, and ICAO documents (Wan et al., 17 Jun 2025). AviationGPT pursues a broader natural-language processing agenda across question-answering, summarization, document writing, information extraction, report querying, data cleaning, and interactive data exploration after continued pre-training and instruction fine-tuning on aviation corpora (Wang et al., 2023). Complementing these systems, Pre-Flight evaluates whether general and open-weight LLMs can reason safely and correctly about aviation operational knowledge across international ground operations, ICAO and FAA regulations, aviation general knowledge, and complex operational scenarios (Brooker et al., 2 Jul 2026).

Safety analytics, maintenance, and structured knowledge extraction form another axis. LogSyn, introduced by Agarwal et al., converts unstructured general aviation maintenance logs into JSON records with fields such as summary_problem, summary_action, failed_component, and category, using few-shot in-context learning and Controlled Abstraction Generation over 6,169 records (Agarwal et al., 24 Nov 2025). The knowledge-grounded aviation safety framework builds an Aviation Safety Knowledge Graph for explainable safety querying over NTSB CAROL data, maintenance bulletins, and FAA Airworthiness Directives (Iyengar et al., 10 Apr 2026).

Trajectory analysis and flight planning have produced both exploratory and end-to-end systems. One line of work fine-tunes LLaMA 2-7B with LoRA to reconstruct noisy ADS-B trajectories from tokenized sequences segmented into 60 s windows under a $\sim 2{,}048$-token limit (Zhang et al., 2024). Another reframes single-step and multi-step flight trajectory prediction as next-token generation over ADS-B waypoint prompts, fine-tuning native 7B-9B LLMs with LoRA and 4-bit quantization (Luo et al., 29 Jan 2025). FRAMe addresses eVTOL-style route planning by translating scenario geometry and human preferences into candidate waypoint sequences, then filtering them through rule-based validity checks and a multi-modal coach (Tabrizian et al., 8 Jul 2026).

Low-altitude and UAV applications represent the most diverse current expansion of the field. MLLM-enabled UAV intent recognition combines on-board RGB and mmWave radar perception with cloud-hosted multimodal reasoning over structured kinematic and payload records (Lei et al., 8 Sep 2025). LAURA uses an LLM as an intelligent crossover operator inside an evolutionary algorithm for Age-of-Information minimization in UAV-assisted wireless sensor networks (Wei et al., 29 Mar 2025). Other works use LLMs for hierarchical motion and communication control in aerial highways (Yan et al., 6 Jun 2025), secure distributed inferencing and anomaly detection across UAV, edge, and cloud tiers (Dharmalingam et al., 5 Feb 2025), natural-language-to-STL navigation and repair (Ping et al., 29 Mar 2026), and offline LLM-designed rewards for adaptive UAV trajectory optimization in onboard VLM inference settings (Li et al., 11 Oct 2025).

5. Empirical results and evaluation practice

Reported performance is strongly task-specific. In the Virtual Co-Pilot case study over 200 A320 simulator scenarios, situational analysis accuracy was 90.5%, procedure retrieval accuracy was 86.5%, and index correction accuracy was 70.5%, with 95% binomial confidence intervals of $\rightarrow$0, $\rightarrow$1, and $\rightarrow$2, respectively; compared with a text-only baseline, situational analysis was approximately 82%, retrieval was approximately 60%, and improvements were statistically significant with $\rightarrow$3 under McNemar’s test (Li et al., 2024). The dual-engine flameout example reported an end-to-end latency of 1.2 s, broken down as 350 ms for vision encoding, 200 ms for retrieval, and 650 ms for the LLM response (Li et al., 2024).

In training and question answering, the DPO-aligned AviationLLM system reports a 57% win rate for Qwen-DPO versus 43% for Qwen-SFT on head-to-head ATDS evaluation, and expert total scores of 13.71 for DPO+RAG, 12.96 for SFT+RAG, 11.62 for DPO, and 10.92 for SFT over 100 QA pairs (Wan et al., 17 Jun 2025). AviationGPT reports 94% arrival-runway extraction accuracy and 96% departure-runway extraction accuracy for Aviation-Mistral-7B-v0.1 on DATIS runway extraction, compared with 39% and 50% for a rule-based baseline, which the paper describes as a performance gain of more than 40 percentage points in both categories (Wang et al., 2023).

For maintenance and safety structuring, LogSyn reports mean LLM-as-Judge scores of 4.7 for SummaryAccuracy, 4.5 for ComponentAccuracy, and 4.8 for CategoryRelevance. Its few-shot configuration achieves 0.9021 accuracy, 0.7455 macro-precision, 0.7779 macro-recall, and 0.7614 macro-F1, outperforming both zero-shot LLM and rule-based NER baselines in the reported table (Agarwal et al., 24 Nov 2025). The same analysis identifies "Powerplant – Sealing & Gaskets" as the largest class with 3,454 records and reports that Sankey flows reveal approximately 4,273 gasket failures leading to component replacements (Agarwal et al., 24 Nov 2025). The safety GraphRAG framework reports query precision, response accuracy, hallucination rate, and traceability of $\rightarrow$4 for the proposed LLM+ASKG system, versus $\rightarrow$5 for an LLM-only GPT-4 baseline (Iyengar et al., 10 Apr 2026).

Trajectory modeling results show both gains and latency costs. The exploratory reconstruction study reports validation loss decreasing from 0.3251 to 0.2533 over six epochs, horizontal errors of at most approximately 50 m and altitude errors of at most approximately 20 ft for linear climbs and descents when ADS-B data are present, and centerline errors within 100 m for curved turns even when an entire curved segment had missing ADS-B points (Zhang et al., 2024). The trajectory prediction study reports that LLaMA-3.1-8B achieved one-step MAE values of 0.0053 in longitude, 0.0046 in latitude, and 23.2201 m in altitude, versus 0.0059, 0.0055, and 26.0105 m for a Transformer baseline, but with inference latency of 1.0585 s versus 0.0007 s (Luo et al., 29 Jan 2025).

UAV and planning studies likewise combine strong task metrics with explicit constraints. The MLLM-enabled UAV intent recognition system reports 92% intent recognition accuracy, mean time from first detection to output of 0.35 s, and false alarm rate of 8% over 50 simulated runs (Lei et al., 8 Sep 2025). LAURA reports average maximum-AoI values of 107 s for $\rightarrow$6 and 122 s for $\rightarrow$7 with Deepseek-v3, compared with 128 s and 162 s for Greedy, corresponding to reductions of 16.6% and 24.7% (Wei et al., 29 Mar 2025). The dual-LLM aerial-highway controller reports convergence to total reward approximately 30 in about $\rightarrow$8 episodes, collision rate below 0.08 in density sweeps, and a 16.3% average total-reward gain over the best baseline (Yan et al., 6 Jun 2025). FRAMe reports that the full RAG+Coach system produces the highest validity for every planner, reaching 93.8% aggregate validity and 99.0% on Easy scenarios for DeepSeek-R1 (Tabrizian et al., 8 Jul 2026).

Evaluation practice has also become a research topic in its own right. Pre-Flight defines an open-source 300-question multiple-choice benchmark and reports that the strongest evaluated model, GPT-5.5, reached 82.7% accuracy, whereas an informal expert reference from aviation professionals was approximately 95% (Brooker et al., 2 Jul 2026). FAA regulations were the hardest category with a 53.0% failure rate across model-question pairs, and the overall micro-average failure rate was 28.4% (Brooker et al., 2 Jul 2026). This benchmark result is central because it quantifies a persistent gap between current general-purpose LLM performance and expert-level aviation operational knowledge.

6. Limitations, controversies, and research directions

A consistent limitation across the literature is that high task performance does not by itself imply certification-grade trustworthiness. AviationLMM explicitly lists hallucination, privacy leakage, lack of uncertainty calibration, brittleness under missing modalities, edge-cloud deployment constraints, and robustness to sensor dropout as unresolved challenges (Li et al., 14 Jan 2026). Pre-Flight sharpens the point empirically: even after gradual improvement from roughly 75% in early 2025 to 82.7% in mid-2026, the top benchmarked model remains substantially below the approximately 95% informal expert reference, and the benchmark is framed as a necessary precondition for responsible deployment in non safety critical aviation operations (Brooker et al., 2 Jul 2026).

Individual systems document concrete failure modes. In the Virtual Co-Pilot, failure sources include vague pilot instructions, OCR errors on low-contrast instrument text, occasional latency spikes during peak traffic in the vector database, and nonexistent content when the knowledge base lacks rare emergency checklists; the proposed mitigations include fallback to the on-board paper QRH or a human ATC query when retrieval confidence drops below a threshold or latency exceeds a safe bound (Li et al., 2024). The trajectory reconstruction work identifies a practical token-length ceiling of approximately 2,048 tokens, limiting effective sequence length to roughly 60 s at 1 Hz and degrading accuracy beyond that horizon (Zhang et al., 2024). The UAV intent-recognition study reports that IR fails to identify small-payload pods in low light and that radar tracks can swap IDs when two UAVs fly in tight formation below 2 m separation (Lei et al., 8 Sep 2025).

Grounding mechanisms reduce but do not eliminate risk. The training-oriented AviationLLM paper states that retrieval recall hinges on index coverage and that out-of-index queries still risk fallback hallucination; it also notes that large-model inference remains nontrivial in bandwidth-constrained cockpits (Wan et al., 17 Jun 2025). The knowledge-graph safety framework reports that Text2Cypher may produce semantically incomplete queries when dialogue context is deep and that its current TF-IDF plus Logistic Regression relation extractor misses subtle causal relations such as “icing $\rightarrow$9 stall” (Iyengar et al., 10 Apr 2026). In LeRAAT, the finite 8 k token context window constrains the number of manual excerpts, and the absence of NOTAMs and real-time ATC constraints is explicitly noted as a limitation (Schlichting et al., 5 Mar 2025).

The forward research agenda is correspondingly hybrid. AviationLMM identifies eight high-impact directions: reliability-weighted, causal and ontology-grounded alignment and fusion; a public-private multimodal data fabric; compute- and data-efficient pretraining curricula; neuro-symbolic reasoning and tool-augmented inference; certification-grade trust pipelines; federated and split-learning deployments; graceful degradation under missing inputs; and domain-controlled scenario generators and digital twins (Li et al., 14 Jan 2026). Other papers align with this trajectory through domain-specific pretraining, reinforcement learning from pilot feedback, adaptive prompting with inferred flight phase, multimodal extensions to maintenance intelligence, hybrid vector-graph retrieval, and formal planner coupling through STL and MILP repair (Li et al., 2024). This suggests that the defining direction of AviationLLM research is not unrestricted text generation, but the progressive coupling of domain-adapted LLMs with multimodal evidence, external knowledge stores, optimization back-ends, and explicit verification layers.