Aero-LLM: A Distributed Framework for Secure UAV Communication and Intelligent Decision-Making

Anyi Liu; Balakrishnan Dharmalingam; Brett Piggott; Guohuan Feng; Rajdeep Mukherjee

REVIEW 5 major objections 6 minor 24 references

Aero-LLM: A Distributed Framework for Secure UAV Communication and Intelligent Decision-Making

T0 review · 5 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Aero-LLM claims that a distributed team of specialized small language models, placed on the drone, at the edge, and in the cloud, can secure UAV communications by predicting packets, detecting sensor anomalies, and forecasting readings.

desk verdict Plausible system design, but the security evaluation never tests the attacks it claims to defend against. read the letter →

arxiv 2502.05220 v1 pith:UK52TSLO submitted 2025-02-05 cs.CR cs.AI

Balakrishnan Dharmalingam , Rajdeep Mukherjee , Brett Piggott , Guohuan Feng , Anyi Liu This is my paper · ORCID

classification cs.CRcs.AI

keywords UAVcommunicationlargelanguagemodelsanomalydetectiontimeseriesforecastingdistributededge-cloudarchitecturecyberthreatsupervisedfine-tuningreinforcementlearningfromhumanfeedback

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Aero-LLM sets out to show that a single all-purpose large language model is the wrong unit for securing UAV missions: a team of much smaller, task-specialized models, distributed across onboard, edge, and cloud computing, can do the job with less latency and lower memory. The paper fine-tunes OPT-350M and OPT-125M to predict and validate network packet fields, TimesNet to flag anomalous sensor readings, and Time-LLM to forecast future sensor values. The reported results include near-perfect prediction of several packet fields, an anomaly-detection accuracy around 0.84, and forecasting accuracy above 82 percent. If these results hold, the framework offers a practical path to real-time cyber defense for UAVs without requiring large models on the aircraft.

What carries the argument

The carrying mechanism is the three-tier deployment of specialized LLMs: small models on the UAV for low-latency packet-level inference, medium models at the edge for aggregated analysis and intermediate anomaly detection, and large models in the cloud for forecasting and deep analysis. The packet-prediction subtask uses a prompt format that presents previous packets as context and labels the next packet as 'chosen' (ground truth) or 'rejected' (a perturbed key-value pair), training the model to prefer the valid packet. Anomaly detection computes the mean squared error between predicted and ground-truth sensor values and sets the anomaly threshold as the percentile of the training loss corresponding to the desired anomaly ratio. Forecasting uses a time-series LLM that reprograms a language model backbone without altering its pretrained weights.

What would settle it

Replay genuine attack traffic through the same fine-tuned models: for example, run a live software-in-the-loop mission, have an adversary alter MAVLink packets in transit, and feed that traffic to the packet validator and anomaly detector. If precision or recall on those real attacks falls materially below the reported values (accuracy 0.8438, precision 0.8542), the central security claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that task specialization plus distribution is what makes LLM-based UAV security work. Fine-tuned small models on board predict the next network packet in a session, and the ground truth packet is paired with a deliberately perturbed 'rejected' packet to train the model, through supervised fine-tuning and iterative reinforcement learning from human feedback, to distinguish valid from corrupted messages. TimesNet, fine-tuned on PX4 sensor telemetry, detects anomalies by thresholding the reconstruction loss at a percentile of the training loss distribution. Time-LLM, built on a Llama-2-7B backbone, forecasts future sensor readings from historical telemetry. On the evaluation data, packet-field accuracy reaches 100 percent for source and destination ports, around 98-99 percent for flags, sequence number, and length, but only 48-54 percent for the acknowledgment number; the best anomaly-detection run reports accuracy 0.8438, precision 0.8542, recall 0.8438, and F-score 0.8050; and forecasting reports test loss 0.1068001, MAE loss 0.2587003, and accuracy above 82 percent.

Load-bearing premise

The security claim assumes that the artificially injected anomalies in the test dataset behave like real jamming, spoofing, man-in-the-middle, and sensor-manipulation attacks.

Editorial extensions

If this is right

Onboard models can validate each incoming or outgoing packet in near real time, so corrupted messages can be dropped before they affect flight decisions.
The edge tier can run anomaly detection on aggregated telemetry without waiting for cloud round trips, keeping detection latency low.
Forecasting future sensor readings gives the ground station a reference baseline, so deviations from expected values become visible early.
Because models are fine-tuned per task and per data source, new mission profiles can be handled by retraining only the affected specialist model.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The modest ack-field accuracy (48-54 percent) suggests the current packet validator would frequently mislabel valid acknowledgment-number changes; a real deployment would likely need per-field confidence thresholds or a dedicated model for that field.
The security claim is only as strong as the anomaly injection procedure; replaying actual jamming, spoofing, or man-in-the-middle traffic through the same pipeline is a direct test the paper does not run.
The same pattern of task-specialized small models across device, edge, and cloud tiers could transfer to other cyber-physical systems, such as autonomous vehicles or industrial controllers, that emit time-series telemetry over vulnerable links.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

Quick take: the paper describes a three-tier distributed LLM framework for UAV communication and decision-making, and that part is described clearly. It fine-tunes OPT variants, TimesNet, and Time-LLM on sensor and network data from SITL/HITL setups, which is legitimate engineering. But the paper's headline claim—robust defense against cyber threats—is not measured anywhere. The evaluation only injects synthetic point anomalies into scalar sensor channels (every nth record, random alterations, variance changes, Poisson interspersion). None of the threat model's attack classes (jamming, spoofing, MITM, sensor manipulation, software exploits, insider threats) is instantiated in the test data. So the abstract's 'robust defense against cyber threats' is a categorical leap from the reported ~0.84 accuracy on synthetic perturbations.

There are also internal inconsistencies. Algorithm 1 defines alpha as TP/(TP+FP), which is precision, not accuracy. The 'weighted harmonic mean' for F-score is undefined. The batch-size table reports identical accuracy and recall, and the paper even notes they are not distinguishable—which is odd because they measure different things. Forecasting accuracy of >82% is asserted from test loss and MAE; there is no direct accuracy calculation or error bars. No baselines are compared against.

Credit where due: the multi-tier architecture with task-specialized LLMs is a sensible direction, and the data collection from a real autopilot stack (PX4, ArduPilot) is a step beyond pure simulation. The fine-tuning procedure is clearly laid out. But these are components, not evidence for the security claim.

If the paper were recast as an anomaly detection framework for synthetic sensor perturbations, the evaluation would be borderline acceptable for a workshop. As a security contribution, it needs a redo with actual attack data or at least a clearly scoped threat model.

I'd send it to peer review because the topic is timely and the system is real, but I'd expect major revisions: proper attack instantiation, corrected metrics, baselines, and honest claims. My own verdict would be reject in the current form.

Recommendation: engage with it as a cautionary example of evaluation/security mismatch, not as a trusted result.

Referee Report

5 major / 6 minor

Summary. The paper presents Aero-LLM, a multi-tier framework that distributes specialized LLMs (OPT-350M/1.3B for packet inference, TimesNet for anomaly detection, Time-LLM for forecasting) across onboard, edge, and cloud servers to secure UAV-GCS communications. The manuscript reports high accuracy for packet-field prediction, anomaly detection accuracy around 0.84 on synthetic sensor perturbations, and a forecasting 'accuracy' above 82% inferred from test loss, and concludes that the framework provides robust defense against cyber threats.

Significance. The architecture idea is topical: using small, task-specific LLMs in a distributed UAV setting is worth studying, and the paper clearly describes the system tiers and fine-tuning pipeline. However, the evaluation does not measure cybersecurity performance as claimed. The only anomaly test data are synthetic point perturbations of normal sensor records, which do not instantiate the threat model's attacks (jamming, spoofing, MITM, software exploits, insider threats). The metric definitions in Algorithm 1 contain a load-bearing error (accuracy computed as precision), and the anomaly threshold in Eq. (2) is fitted to the training loss, making the reported detection performance circular. The forecasting claim is made from a loss value without a defined accuracy metric, baselines, or error bars. The paper provides no code, no data, and no reproducibility artifacts, so the claims cannot be independently checked.

major comments (5)

[Section VII.1 vs Section III] The anomaly detection evaluation only uses synthetic perturbations of normal sensor records (every n-th record set anomalous, random alteration, variance changes, Poisson injection). None of the threat-model attack classes (jamming, spoofing, man-in-the-middle, sensor manipulation, software exploits, insider threats) is instantiated in the test data. Consequently, the Abstract's claim of 'robust defense against cyber threats' and Section VIII's 'ensuring security' are not supported by the measurements. The evaluation should be redone with actual attack traces or a clearly defined realistic attack injection model for each threat vector.
[Algorithm 1, line 17] The formula α = TP/(TP+FP) defines precision, not accuracy. Accuracy should be (TP+TN)/(TP+TN+FP+FN). This is not a notation nit: the reported numbers labeled 'accuracy' in Table V and Figure 4 are actually precision values, and Section VII.1c's statement that 'accuracy and recall lines are not distinguishable as the accuracy and recall data are the same' is inconsistent with the definitions in Algorithm 1, where recall = TP/(TP+FN) and precision = TP/(TP+FP). The anomaly detection results must be recomputed with the correct metric.
[Eq. (2), Section VII.1] The anomaly threshold is defined as the (100 − anomaly_ratio) percentile of the training loss, so anomalies are flagged when the test loss exceeds a quantile of the training loss distribution. Since the test anomalies are generated by the same perturbation procedures used to create the threshold, the reported accuracy/precision/recall reflect the threshold calibration rather than the model's ability to detect genuinely novel attacks. This circularity means the numbers in Figures 4 and 5 do not provide independent evidence of detection capability.
[Section VII.2] The forecasting evaluation asserts 'the accuracy is > 82%' from a test loss of 0.1068001 and MAE of 0.2587003, but no definition of accuracy for time-series forecasting is given (MAE/MSE are error measures, not accuracy), and there are no baselines (persistence, ARIMA, or existing time-series models) or confidence intervals. The claim is unverifiable and should be replaced with standard forecasting metrics such as MSE, MAE, and a comparison against at least one baseline.
[Table V and Figure 6] The batch-size study reports identical accuracy (0.8438), precision (0.8542), recall (0.8438), and F-score (0.8050) for all batch sizes, while the text claims batch size has a 'significant role in sensitivity.' Identical metrics across batch sizes is implausible and suggests the evaluation either does not vary the model or the metric computation is flawed; this inconsistency further undermines the reported anomaly detection results.

minor comments (6)

[Section IV heading] The heading reads 'Areo-LLM uses a multi-tiered architecture'; this should be 'Aero-LLM'.
[Section VI.B.1] The sentence 'Accuracy and precision are measured by calculating MAE and MSE' is confusing because MAE and MSE are error metrics, not accuracy or precision; this should be rephrased.
[Section VI.A.1] The acronym 'SIL' is used for 'System-in-the-Loop' and 'SITL' for 'Software-in-the-Loop'; these should be defined and used consistently.
[Section VII.1] Figure 4 is mentioned in the text but never referenced; the anomaly metrics are said to be 'illustrated in Figure 5' twice.
[Table V] The table is missing an explicit row header for the metric values (accuracy, precision, recall, F1); the reader has to infer which rows correspond to which metric.
[Abstract and Introduction] The paper claims 'low memory footprints in terms of VRAM usage' but does not report any VRAM measurements; either report the data or remove the claim.

Circularity Check

1 steps flagged · score 6.0 of 10

Anomaly threshold is calibrated to the same anomaly_ratio used to generate the test anomalies, so the reported security metrics are a self-generated calibration rather than an independent prediction.

fitted input called prediction [Section VII.1, Eq. (2), item a)]
"To enrich our dataset with anomalous data, we employ several techniques to transform normal records. First, we designate every n-th record as anomalous. ... The anomaly threshold is calculated as: threshold = P ercentile(loss, 100 − anomaly_ratio) (2)"

Eq. (2) sets the detection threshold at the (100 − anomaly_ratio) percentile of training loss, so the detector's operating point is determined by the input anomaly_ratio. The test anomalies are then generated with the same ratio: choosing n=5 makes every fifth record anomalous, i.e., 20% of the test data is injected, and the other perturbation techniques use the same ratio parameter. Because these transformations are applied to the same dataset before the normal/anomalous split, training and test anomalies share the exact generative mechanism. The model is therefore calibrated to flag approximately anomaly_ratio of the points, and the labels contain approximately the same fraction by construction.

full rationale

The only exhibitable reduction is in the anomaly-detection evaluation. Eq. (2) uses the anomaly_ratio to set the threshold as a percentile of training loss, and Section VII.1.a generates test anomalies with the same ratio (n=5, i.e., 20% anomalous) using the same transformation routines. At the aggregate level, the number of predicted positives and the number of labeled positives are both approximately anomaly_ratio of the data, so the confusion matrix is matched by construction rather than measured against an independent benchmark. This makes the central security claim circular. The forecasting component is not circular: it reports test loss and MAE on a normal mission split, although its translation to 'accuracy > 82%' is unsupported. Algorithm 1 also mislabels precision as accuracy (alpha = TP/(TP+FP)) and sets alpha equal to pi by definition, but these are reporting defects rather than circularity. No load-bearing self-citations appear in the paper. The score is 6 because the paper's headline security claim reduces to a self-calibrated synthetic-anomaly evaluation, while the rest of the framework content is independent.

Assumptions & free parameters 2 free parameters · 4 assumptions · 1 invented entities

The central claim depends on standard ML components and simulation data; no physical constants or new entities are introduced. The main unverified assumptions are the representativeness of the simulated data and the validity of the chosen anomaly detection setup for real cyber attacks. The free parameters are the hand-chosen anomaly ratio and the periodic anomaly injection period, which directly set the operating point of the detection metric.

free parameters (2)

anomaly_ratio = not stated (used in Eq. 2)
The anomaly detection threshold is the (100 minus anomaly_ratio) percentile of training loss; the value is chosen by hand and directly controls the detection operating point.
n-th record anomaly period = 5
Section VII.1a sets every 5th record to anomalous; this periodic injection is an arbitrary choice that makes anomalies trivially detectable and is not a realistic attack pattern.

assumptions (4)

domain assumption TimesNet and Time-LLM, with their published architectures and pretrained weights, can be fine-tuned on the collected UAV sensor data and achieve the reported performance.
The paper relies on these models as core components but provides no comparison to a baseline or independent verification of the fine-tuned behavior.
domain assumption The SITL/HITL simulations from PixHawk and ArduPilot produce network and sensor data representative of real UAV communication channels and adversarial perturbations.
Section VI-A collects data from simulation; the threat model in Section III assumes real attack vectors, but the connection between simulated normal data and real attacks is not established.
domain assumption The MSE-loss-plus-percentile-threshold anomaly detection method is a valid proxy for detecting the listed cyber attacks.
Section VII uses synthetic perturbations; no attack-specific ground truth is used, so the validation does not cover jamming, spoofing, MITM, or insider threats.
domain assumption The three-stage DeepSpeed SFT/RLHF pipeline is applied as described and produces the reported packet-prediction models.
Section VI-B1 claims SFT, critic, and iterative RLHF, but no human feedback dataset, critic training data, or reward model details are provided.

invented entities (1)

Intelligent orchestrator/moderator agent
purpose: Coordination and moderation among the multiple specialized LLMs (mentioned in Section I).
The introduction says the distributed architecture 'potentially introduces intelligent agents' that orchestrate the LLM team, but no design, implementation, or evaluation of such an agent appears in the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Aero-LLM: A Distributed Framework for Secure UAV Communication and Intelligent Decision-Making." pith.science (2026). https://pith.science/paper/UK52TSLO

@misc{pith2026250205220,
  author       = {Pith},
  title        = {Pith review of: Aero-LLM: A Distributed Framework for Secure UAV Communication and Intelligent Decision-Making},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UK52TSLO}},
  note         = {Machine review of arXiv:2502.05220}
}

read the original abstract

Increased utilization of unmanned aerial vehicles (UAVs) in critical operations necessitates secure and reliable communication with Ground Control Stations (GCS). This paper introduces Aero-LLM, a framework integrating multiple Large Language Models (LLMs) to enhance UAV mission security and operational efficiency. Unlike conventional singular LLMs, Aero-LLM leverages multiple specialized LLMs for various tasks, such as inferencing, anomaly detection, and forecasting, deployed across onboard systems, edge, and cloud servers. This dynamic, distributed architecture reduces performance bottleneck and increases security capabilities. Aero-LLM's evaluation demonstrates outstanding task-specific metrics and robust defense against cyber threats, significantly enhancing UAV decision-making and operational capabilities and security resilience against cyber attacks, setting a new standard for secure, intelligent UAV operations.

Figures

Figures reproduced from arXiv: 2502.05220 by the authors.

**Figure 1.** System Architecture of Aero-LLM in real-time, optimizing performance and responsiveness for mission-critical applications. A. LLM Fine-tuning and Deployment [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** illustrates the interaction between different components, including the GCS, the UAV, the cloud, and edge servers. The cloud server performs offline processes, including fine-tuning the LLM for enhanced inferencing, ensuring the LLM’s responsiveness is precisely calibrated for anomaly detection. During active missions, the UAV transmits real-time sensor data to the edge server for immediate processing—an online pro… view at source ↗

**Figure 3.** The Format of Data Under the fine-tuning. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Metrics From TimesNet Anomaly Detection [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Anomalous Data vs Metrics b) Change variance of anomalous data: Sensor reading accelerometer_m_s2_2 is arbitrarily changed to simulate variances and the anomaly detection script is run for each variance and the data is collected. The metrics from changing the variance…

**Figure 6.** Figure 6: Comparative analysis of TimesNet performance [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 16 canonical work pages

[1]

ChatGPT for Software Security: Exploring the Strengths and Limitations of ChatGPT in the Security Applications,

Z. Wang, L. Zhang, and P. Liu, “ChatGPT for Software Security: Exploring the Strengths and Limitations of ChatGPT in the Security Applications,” 2023

work page 2023
[2]

Machine-generated text: A comprehensive survey of threat models and detection methods,

E. N. Crothers, N. Japkowicz, and H. L. Viktor, “Machine-generated text: A comprehensive survey of threat models and detection methods,” IEEE Access, vol. 11, 2023

work page 2023
[3]

Pop quiz! can a large language model help with reverse engineering?

H. Pearce, B. Tan, P. Krishnamurthy, F. Khorrami, R. Karri, and B. Dolan-Gavitt, “Pop quiz! can a large language model help with reverse engineering?” CoRR, vol. abs/2202.01142, 2022

arXiv 2022
[4]

Jigsaw: Large language models meet program synthesis,

N. Jain, S. Vaidyanath, A. Iyer, N. Natarajan, S. Parthasarathy, S. Ra- jamani, and R. Sharma, “Jigsaw: Large language models meet program synthesis,” in Proceedings of the 44th International Conference on Software Engineering, ser. ICSE ’22. New York, NY , USA: Association for Computing Machinery, 2022, p. 1219–1231. [Online]. Available: https://doi-org....

work page arXiv 2022
[5]

Revolutionizing cyber threat detection with large language models,

M. A. Ferrag, M. Ndhlovu, N. Tihanyi, L. C. Cordeiro, M. Debbah, and T. Lestable, “Revolutionizing cyber threat detection with large language models,” 2023

work page 2023
[6]

Lost at c: A user study on the security implications of large language model code assistants,

G. Sandoval, H. Pearce, T. Nys, R. Karri, S. Garg, and B. Dolan-Gavitt, “Lost at c: A user study on the security implications of large language model code assistants,” 2023

work page 2023
[7]

Llama 2: Enhancing large language models,

Meta AI, “Llama 2: Enhancing large language models,” Accessed: March 28, 2024. [Online]. Available: https://huggingface.co/blog/llama2

work page 2024
[8]

Gemini: A family of highly capable multimodal models,

Google DeepMind, “Gemini: A family of highly capable multimodal models,” Accessed: March 28, 2024. [Online]. Available: https: //huggingface.co/papers/2312.11805

arXiv 2024

Show all 24 references

[9]

Mistral: Generating training data for large language mod- els,

Mistral AI, “Mistral: Generating training data for large language mod- els,” Accessed: March 28, 2024

2024
[10]

Dbrx base, a mixture-of-experts (moe) large language model,

Databricks, “Dbrx base, a mixture-of-experts (moe) large language model,” Accessed: March 28, 2024. [Online]. Available: https: //huggingface.co/databricks/dbrx-base

2024
[11]

Opt: Open pre-trained transformer language models,

S. Zhang, S. Roller, N. Goyal, M. Artetxe, M. Chen, S. Chen, C. Dewan, M. Diab, X. Li, X. V . Lin, T. Mihaylov, M. Ott, S. Shleifer, K. Shuster, D. Simig, P. S. Koura, A. Sridhar, T. Wang, and L. Zettlemoyer, “Opt: Open pre-trained transformer language models,” 2022

2022
[12]

Edgeformer: Transformer- based natural language processing on edge devices,

H. Qiu, J. Qiu, Y . Li, Q. Li, and J. Li, “Edgeformer: Transformer- based natural language processing on edge devices,” arXiv preprint arXiv:2205.12487, 2022

2022 arXiv
[13]

Deflating pre-trained language models for efficient deployment on embedded systems,

J. Zhang, Z. Li, Y . Xue, C. Zhang, and D. Li, “Deflating pre-trained language models for efficient deployment on embedded systems,” arXiv preprint arXiv:2212.08015, 2022

2022
[14]

Globalpipeline: An efficient model parallelism via simple data- parallelism principles,

Z. Su, Y . T. Tan, Y . Liang, H. Choi, H. Park, S. Chen, X. Lin et al., “Globalpipeline: An efficient model parallelism via simple data- parallelism principles,” in Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP) , 2022

2022
[15]

Knowledge distillation for small-footprint efficient transformers,

D. Dai, C. Li, and B. Peng, “Knowledge distillation for small-footprint efficient transformers,” in Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (ACL) , 2022

2022
[16]

Making pre-trained language models better few-shot learners,

T. Gao, A. Fisch, and D. Chen, “Making pre-trained language models better few-shot learners,” arXiv preprint arXiv:2103.03511 , 2021

2021 arXiv
[17]

Deduplicating training data makes language models better,

K. Lee, D. Ippolito, E. Nouri, R. L. Bras, C. Callison-Burch, and J. Sedoc, “Deduplicating training data makes language models better,” arXiv preprint arXiv:2203.06642 , 2022

2022 arXiv
[18]

Learning to summarize from human feedback,

N. Stiennon, L. Ouyang, J. Ziegler, R. Byrne, A. Radford, L. Paull, A. Bachand, D. Kim, P. M. D. Moore et al. , “Learning to summarize from human feedback,” arXiv preprint arXiv:2009.01325 , 2020

2009 arXiv
[19]

Training language models to follow instructions with human feedback,

L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. L. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray et al. , “Training language models to follow instructions with human feedback,” arXiv preprint arXiv:2203.02155, 2022

2022 arXiv
[20]

Px4 autopilot user guide,

PixHawk, “Px4 autopilot user guide,” Accessed: May 20, 2024. [Online]. Available: https://docs.px4.io/main/en/

2024
[21]

Ardupilot documentation,

ArduPilot, “Ardupilot documentation,” Accessed: May 20, 2024. [Online]. Available: https://ardupilot.org/ardupilot/index.html

2024
[22]

Time-LLM: Time series forecasting by reprogramming large language models,

M. Jin, S. Wang, L. Ma, Z. Chu, J. Y . Zhang, X. Shi, P.-Y . Chen, Y . Liang, Y .-F. Li, S. Pan, and Q. Wen, “Time-LLM: Time series forecasting by reprogramming large language models,” in International Conference on Learning Representations (ICLR) , 2024

2024
[23]

Timesnet: Temporal 2d-variation modeling for general time series analysis,

H. Wu, T. Hu, Y . Liu, H. Zhou, J. Wang, and M. Long, “Timesnet: Temporal 2d-variation modeling for general time series analysis,” in International Conference on Learning Representations , 2023

2023
[24]

Deep- speed inference: Enabling efficient inference of transformer models at unprecedented scale,

R. Yazdani Aminabadi, S. Rajbhandari, M. Zhang, A. A. Awan, C. Li, D. Li, E. Zheng, J. Rasley, S. Smith, O. Ruwase, and Y . He, “Deep- speed inference: Enabling efficient inference of transformer models at unprecedented scale,” arXiv preprint arXiv:2207.00032 , 2022

2022 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

ChatGPT for Software Security: Exploring the Strengths and Limitations of ChatGPT in the Security Applications,

Z. Wang, L. Zhang, and P. Liu, “ChatGPT for Software Security: Exploring the Strengths and Limitations of ChatGPT in the Security Applications,” 2023

work page 2023

[2] [2]

Machine-generated text: A comprehensive survey of threat models and detection methods,

E. N. Crothers, N. Japkowicz, and H. L. Viktor, “Machine-generated text: A comprehensive survey of threat models and detection methods,” IEEE Access, vol. 11, 2023

work page 2023

[3] [3]

Pop quiz! can a large language model help with reverse engineering?

H. Pearce, B. Tan, P. Krishnamurthy, F. Khorrami, R. Karri, and B. Dolan-Gavitt, “Pop quiz! can a large language model help with reverse engineering?” CoRR, vol. abs/2202.01142, 2022

arXiv 2022

[4] [4]

Jigsaw: Large language models meet program synthesis,

N. Jain, S. Vaidyanath, A. Iyer, N. Natarajan, S. Parthasarathy, S. Ra- jamani, and R. Sharma, “Jigsaw: Large language models meet program synthesis,” in Proceedings of the 44th International Conference on Software Engineering, ser. ICSE ’22. New York, NY , USA: Association for Computing Machinery, 2022, p. 1219–1231. [Online]. Available: https://doi-org....

work page arXiv 2022

[5] [5]

Revolutionizing cyber threat detection with large language models,

M. A. Ferrag, M. Ndhlovu, N. Tihanyi, L. C. Cordeiro, M. Debbah, and T. Lestable, “Revolutionizing cyber threat detection with large language models,” 2023

work page 2023

[6] [6]

Lost at c: A user study on the security implications of large language model code assistants,

G. Sandoval, H. Pearce, T. Nys, R. Karri, S. Garg, and B. Dolan-Gavitt, “Lost at c: A user study on the security implications of large language model code assistants,” 2023

work page 2023

[7] [7]

Llama 2: Enhancing large language models,

Meta AI, “Llama 2: Enhancing large language models,” Accessed: March 28, 2024. [Online]. Available: https://huggingface.co/blog/llama2

work page 2024

[8] [8]

Gemini: A family of highly capable multimodal models,

Google DeepMind, “Gemini: A family of highly capable multimodal models,” Accessed: March 28, 2024. [Online]. Available: https: //huggingface.co/papers/2312.11805

arXiv 2024