Pith. sign in

REVIEW 3 major objections 6 minor 25 references

SAFLITE: Fuzzing Autonomous Systems via Large Language Models

T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read SaFliTe uses an LLM to rank mutants against a natural-language 'interestingness' definition, lifting bug-triggering selections by 93.1% and safety-violating test cases by up to 234.5% across three drone fuzzers.

desk verdict New idea with a plausible mechanism, but the headline efficiency gain comes from an in-sample classification proxy rather than direct fuzzing; the paper deserves a serious referee and a strong revision. read the letter →

arxiv 2412.18727 v1 pith:3GWV5TLR submitted 2024-12-25 cs.SE cs.AIcs.SYeess.SY

Taohong Zhu , Adrians Skapars , Fardeen Mackenzie , Declan Kehoe , William Newton , Suzanne Embury , Youcheng Sun This is my paper · ORCID
classification cs.SEcs.AIcs.SYeess.SY
keywords fuzztestingautonomoussystemsunmannedaerialvehicleslargelanguagemodelstestcaserankingmutation-basedfuzzingsafetypolicyPX4
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

The paper proposes SaFliTe, a predictor that sits between a fuzzer's mutation step and the simulator, using a large language model to score each mutant against a natural-language definition of interestingness drawn from the safety policy under test. The central claim is that this ranking makes fuzzing of autonomous drone control systems more efficient: compared with PGFuzz's random selection, SaFliTe raises the per-iteration chance of selecting a bug-triggering operation by an average of 93.1%, and when integrated into DeepHyperion-UAV, CAMBA, and TUMB it increases the number of safety-violating test cases generated by 234.5%, 33.3%, and 17.8%, respectively. The paper also reports two newly discovered bugs in PX4 found during 24-hour runs, including one violation absent from PGFuzz's bug list. A sympathetic reader would care because AS fuzzing suffers from enormous search spaces, and the paper offers a way to focus mutation effort using an LLM's understanding of safety rules instead of hand-written heuristics for each system.

What carries the argument

The central object is SaFliTe itself: an LLM-based predictor that converts qualitative safety knowledge into a quantitative mutant ranking before any expensive simulation runs. Its load-bearing piece is a prompt template that takes the definition of interestingness, the current system state, and the mutants, and asks the LLM via chain-of-thought reasoning to score each mutant from 0 to 10. That ranked list is then filtered by a selection threshold, so only mutants the LLM judges likely to violate the safety property are executed. The work this machinery does is to replace the comparatively blind mutation-selection step of existing fuzzers with a knowledge-driven filter, without modifying the fuzzers' seed management or mutation internals.

What would settle it

Repeat the RQ3 campaigns many times with fresh random seeds and compare the distribution of valid test cases between the original and SaFliTe-integrated tools; if the 234.5% advantage, or the 33.3% and 17.8% gains, collapses into overlap with the baseline, the headline gains are run-to-run noise. Separately, rerun the RQ2 classification on bug reports that were not used to build the definition of interestingness; if the bug-causing parameter is not ranked interesting under this held-out protocol, the 93.1% figure reflects information leakage from the bug report itself.

Watch

Extended reading notes

Core claim

SaFliTe's claim, stated on the paper's own terms, is that an LLM can act as a test-case relevance predictor for autonomous-system fuzzing. Given three text inputs—a definition of interestingness (the safety policy or test objective written in natural language), the current system state, and a structured description of the mutants—the LLM reasons about each mutant, emits a chain-of-thought explanation, and assigns a score from 0 to 10; SaFliTe then selects the top-ranked mutants for execution. The paper evaluates this predictor in three stages: flight-log classification (GPT-3.5 reaches 68.6% accuracy and F1, Mistral-7B 62.9% accuracy), bug-oriented ranking on PGFuzz's repository (all eight bug-causing test cases ranked at least mid-interesting, with an average 93.1% increase in per-round selection probability), and full integration into three competition tools, where valid-test-case counts rose by 234.5%, 33.3%, and 17.8%. The paper further reports that the integrated PGFuzz found a new PX4 Hold-mode violation in 126 fuzzing rounds versus 269 for the baseline.

Load-bearing premise

The efficiency claims stand on the assumption that SaFliTe's ranking quality, measured in controlled settings, carries over to live fuzzing loops: the 93.1% improvement is computed from classifying a known bug-causing parameter name as interesting using a definition and system state taken from the same bug report, and the integration numbers in Table II come from single runs without reported variance or significance tests.

Editorial extensions

If this is right

  • Existing UAV fuzzers can be upgraded to LLM-ranked selection by inserting one module before execution; the paper demonstrates this for PGFuzz, DeepHyperion-UAV, CAMBA, and TUMB with no change to their mutation engines.
  • Local 7B models such as Mistral-7B come close to GPT-3.5's performance, so SaFliTe can run on-site without sending system states to a proprietary API.
  • The definition of interestingness is the main tuning knob: broad definitions limit accuracy, while bug-specific definitions sharply improve ranking, so testers should write precise natural-language objectives.
  • The approach can surface new bugs, not just reproduce known ones: the paper reports a PX4 violation discovered in 126 rounds that was not in PGFuzz's bug list.
  • Because the framework abstracts the fuzzing workflow, the same LLM-ranking pattern should transfer to other autonomous-system domains whenever a natural-language safety property and a state description are available.

Reading between the lines

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

  • Running each integrated campaign many times with fresh random seeds and reporting the distribution of valid-test-case counts is the natural next experiment, since the reported integration numbers come from single runs.
  • The RQ2 evaluation gave SaFliTe parameter names but not the mutated values PGFuzz assigns; an end-to-end variant that feeds value-level mutants to the LLM would measure the ranking's effect on full fuzzing loops rather than parameter-name selection.
  • Fine-tuning a local 7B model on labelled PX4 flight logs is a plausible next step suggested by the RQ1 results, potentially closing the gap to GPT-3.5 while keeping the system on-site.
  • The same prompt pattern could reduce simulation cost in other cyber-physical domains such as ground vehicles or marine robots wherever a natural-language safety objective and state description exist.
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.

Referee Report

3 major / 6 minor

Summary. The paper proposes SaFliTe, a component that uses an LLM to score and rank mutant test cases before execution, in order to improve the efficiency of mutation-based fuzzing for autonomous drone systems. The framework is described as universal, and SaFliTe is instantiated with GPT-3.5, Mistral-7B, and Llama-2-7B. Evaluation covers three research questions: (RQ1) LLM prediction of flight-log interestingness, where GPT-3.5 reaches 68.6% accuracy and F1; (RQ2) integration with PGFuzz on 8 known bugs, reporting a 93.1% average increase in the likelihood of selecting a bug-causing test case per fuzzing round, plus two newly observed bugs; and (RQ3) integration with DeepHyperion-UAV, CAMBA, and TUMB on the SBFT UAV Testing Competition benchmark, reporting increases in valid test case counts of 234.5%, 33.3%, and 17.8%, respectively.

Significance. If the central claim holds, SaFliTe would be a promising, tool-agnostic way to inject LLM-based relevance reasoning into existing AS fuzzing pipelines, with locally deployable models (Mistral-7B) performing comparably to GPT-3.5. The paper's strengths include the release of the tool, the breadth of LLMs and fuzzers examined, and the use of a public competition benchmark. However, the headline efficiency numbers are not yet backed by direct, repeated fuzzing measurements: RQ2 is an in-sample parameter-name classification proxy, and RQ3 consists of single-run counts without statistical support. The contribution is therefore plausible and interesting, but the current evidence is not strong enough to support the abstract's quantitative claims as stated.

major comments (3)
  1. [Section III-B, Appendix A/B] The 93.1% (and 94.7%) improvement is computed from an in-sample parameter-name proxy, not from fuzzing runs. For each of the 8 bugs, SaFliTe is given the bug's own current state and a definition of interestingness translated from the same policy violation, then scores parameter names from PGFuzz's parameter list. The baseline PGFuzz mutates values, and those values are not shown to the LLM, as the paper itself concedes for Bugs 24 and 25 in Appendix B. This measures whether the known bug-causing parameter name is semantically related to the known policy, not whether an interesting-ranked concrete test case (parameter plus value) is more likely to trigger a violation in an actual fuzzing iteration. The two new bugs in Section III-B(c) are single observations with no repeated runs, so the 126-vs-269 rounds comparison also lacks support. A value-aware, repeated evaluation comparing PGFuzz vs. PGFuzz+SaFliTe is needed to substantiate the headline claim.
  2. [Section III-C, Table II] Table II reports one count per tool/mission/LLM cell, with no repetitions, confidence intervals, or significance tests. Several cells show no improvement (CAMBA Mission 6: 6 vs. 6; TUMB Missions 3 and 5: 15 vs. 15 and 7 vs. 7), and the paper itself attributes cross-mission variability to task complexity in Section III-C(b). The aggregate improvements (234.5%, 33.3%, 17.8%) are therefore not established as stable or statistically significant. It is also unclear whether the baseline rows were re-run in the same environment or taken from the competition results; if the latter, the comparison is confounded by environment and seed selection. Please provide multiple independent runs per configuration, report the baselines from the same environment, and include an appropriate significance test or effect sizes with confidence intervals.
  3. [Section II-B(a), Section III-B(a)] The interestingness thresholds (score 4 and 7) and the LLM temperature are free parameters, and no sensitivity analysis is reported. The definitions of interestingness in Appendix A are hand-written from the very bug reports used to evaluate RQ2, which creates a risk that the 93.1% figure is tuned to the selected bugs. Please report results across a grid of thresholds and temperature values, and ideally use independently written definitions (e.g., by a second annotator) to bound the influence of definition choice.
minor comments (6)
  1. [Algorithm 2] The line 'rank ist ← LLMAGENT(prompt)' contains a typo; it should read 'rank list'.
  2. [Section III-C(c)] The sentence 'considering it is a local LLM that has not undergone fine-tuning.' is duplicated verbatim.
  3. [Section II-B] The word 'Moverover' should be 'Moreover'.
  4. [Section III-C(a), Table II] The abstract reports percentage improvements that are not directly derivable from Table II; please add a column or explicit derivation so that 234.5%, 33.3%, and 17.8% can be reproduced from the table.
  5. [Section III-B(a), Figure 4] Figure 4 is not clearly mapped to the 93.1% calculation; please show how the pie-chart proportions for each bug combine into the reported average likelihood increase.
  6. [Section I and Data Availability] The paper states that SaFliTe is released at [8], but no versioned artifact or minimal reproduction script for the RQ2/RQ3 experiments is described; adding such details would strengthen reproducibility.

Circularity Check

2 steps flagged · score 5.0 of 10

RQ2's headline 93.1% gain is an in-sample semantic-match proxy; RQ1/RQ3 supply independent evidence.

  1. self definitional [Section III-B Experimental Setup; Appendix A (Table IV)]
    "If a test case has the potential to cause the policy violation or drone crash, then this test case is considered an interesting test case."

    Interestingness is defined as the potential to cause the very policy violation that the known bug triggered, and the state supplied is the state at that bug's occurrence. Thus, the target label ('interesting') is a paraphrase of the input definition; classifying the known bug-causing parameter as interesting is a consequence of how the definition was constructed, not an independent prediction.

  2. fitted input called prediction [Section III-B Results]
    "Therefore, compared to PGFuzz's random selection of all possible test cases, using SAFLITE to select only from the interesting category increases the likelihood of selecting a test case that leads to a policy violation by 93.1% in each fuzzing round."

    This number is derived from the in-sample classification of known bug-causing parameter names under definitions and states taken from the same bug reports; it is not a measurement of real fuzzing iterations. PGFuzz's baseline mutates values, but SaFliTe scored parameter names without showing mutated values (the paper itself notes GPT-3.5 lacked knowledge of specific values for Bugs 24/25), so the 93.1% figure is a semantic-match proxy. The claimed per-iteration gain is therefore an artifact of the input construction rather than a direct fuzzing result.

full rationale

The circularity is concentrated in RQ2. The evaluation constructs the 'definition of interestingness' by translating the very policy violation that a known bug triggered, and supplies the system state recorded at that same bug occurrence; it then scores the parameter names associated with that bug and turns the classification of the known bug-causing parameter into the 93.1% per-iteration improvement figure. Because the prompt already contains the violated policy and the pre-violation state, an LLM rating the violation-triggering parameter as interesting is a semantic re-description of the input rather than an independent prediction of fuzzing performance. The result is also computed on parameter names only, while PGFuzz's baseline randomly selects a parameter and mutates its value; the paper itself acknowledges that for Bugs 24 and 25, GPT-3.5 lacked knowledge of the specific values, confirming that value-level information was not part of the prediction. RQ1 and RQ3 are not circular in the same way: RQ1 uses independent labelled flight logs with a broad, pre-defined notion of interestingness, and RQ3 measures actual valid-test-case counts produced by full fuzzing runs before and after integration. The RQ3 single-run counts without repetitions are an empirical-rigor concern, not a definitional circularity. There is no load-bearing self-citation chain; the only self-reference is the authors' public repository link. Overall, one headline quantitative claim reduces by construction, but the central idea retains independent supporting evidence, giving a score of 5.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central method introduces no new physical entity or fitted scientific constant. The numerical claims rest on hand-chosen scoring thresholds, manually authored interestingness definitions, simulator-fidelity assumptions, and a proxy metric for fuzzing efficiency. No parameter is fit to maximize the reported percentages, but several evaluation choices directly shape them.

free parameters (2)
  • Interestingness score thresholds (4 and 7) = below 4 non-interesting, 5 to 7 mid-interesting, above 7 interesting
    Hand-chosen thresholds in RQ2 (Section III-B) define the categories that produce the reported 93.1% and 94.7% likelihood gains; no sensitivity analysis is provided.
  • LLM temperature setting = default value not reported; temperature 1 tested separately
    The paper shows GPT-3.5 at temperature 1 degrades accuracy sharply and recommends against high temperature, but it does not report the default temperature used in the main experiments, so the exact sampling behavior behind the scores is a free choice.
assumptions (4)
  • domain assumption Simulator executions are faithful proxies for real autonomous system behavior, so a valid test case or policy violation in simulation is a real defect.
    All evaluation results and both new-bug reports come from ArduPilot, PX4, and PX4-Avoidance simulators with no hardware confirmation (Sections III-B and III-C).
  • ad hoc to paper The hand-written definition of interestingness used in each RQ2 experiment captures the full safety-relevant semantics of the corresponding PGFuzz bug.
    Each definition is manually translated from the same bug list used for evaluation (Section III-B a), so the prompt itself encodes part of the expected answer.
  • domain assumption The 117 log files in RQ1 are correctly labelled as interesting or non-interesting.
    Labels come from PX4 GitHub issues, developer test cards, and submodule development logs (Section III-A); no independent verification of label quality is reported.
  • domain assumption PGFuzz selects parameter names uniformly at random, so increasing the probability of picking the known bug parameter by filtering categories equals improving fuzzing efficiency.
    The RQ2 probability calculation (Section III-B b) compares random selection over all parameters with selection from the interesting category; this ignores value-level mutation and actual sampling distributions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SAFLITE: Fuzzing Autonomous Systems via Large Language Models." pith.science (2026). https://pith.science/paper/3GWV5TLR

@misc{pith2026241218727,
  author       = {Pith},
  title        = {Pith review of: SAFLITE: Fuzzing Autonomous Systems via Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3GWV5TLR}},
  note         = {Machine review of arXiv:2412.18727}
}
read the original abstract

Fuzz testing effectively uncovers software vulnerabilities; however, it faces challenges with Autonomous Systems (AS) due to their vast search spaces and complex state spaces, which reflect the unpredictability and complexity of real-world environments. This paper presents a universal framework aimed at improving the efficiency of fuzz testing for AS. At its core is SaFliTe, a predictive component that evaluates whether a test case meets predefined safety criteria. By leveraging the large language model (LLM) with information about the test objective and the AS state, SaFliTe assesses the relevance of each test case. We evaluated SaFliTe by instantiating it with various LLMs, including GPT-3.5, Mistral-7B, and Llama2-7B, and integrating it into four fuzz testing tools: PGFuzz, DeepHyperion-UAV, CAMBA, and TUMB. These tools are designed specifically for testing autonomous drone control systems, such as ArduPilot, PX4, and PX4-Avoidance. The experimental results demonstrate that, compared to PGFuzz, SaFliTe increased the likelihood of selecting operations that triggered bug occurrences in each fuzzing iteration by an average of 93.1\%. Additionally, after integrating SaFliTe, the ability of DeepHyperion-UAV, CAMBA, and TUMB to generate test cases that caused system violations increased by 234.5\%, 33.3\%, and 17.8\%, respectively. The benchmark for this evaluation was sourced from a UAV Testing Competition.

Figures

Figures reproduced from arXiv: 2412.18727 by the authors.

Figure 1
Figure 1. An Illustrative Workflow of an Autonomous System for UAVs [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Workflow of the universal AS fuzzing with LLMs [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. A concrete example of SAFLITE applied to a fuzzing tool testing a UAV obstacle avoidance system, PX4-Avoidance. The safety policy of system requires the UAV to maintain a distance of at least 1.5 metres from obstacles. The goal is to find test cases where the UAV either comes within 1.5 metres of an obstacle or collides with one. Each test case represents the number, size, and position of obstacles. In this example,… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: For each bug, SAFLITE categorizes all potential test cases generated by PGFuzz into three groups based on the corresponding Current State and Definition of Interestingness. The figure highlights which category the test case that triggered the bug was assigned to. accur…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 21 canonical work pages

  1. [1]

    Security modeling of au- tonomous systems: A survey,

    F. Jahan, W. Sun, Q. Niyaz, and M. Alam, “Security modeling of au- tonomous systems: A survey,” ACM Computing Surveys (CSUR), vol. 52, no. 5, pp. 1–34, 2019

    work page 2019

  2. [2]

    Survey on vanet security challenges and possible cryptographic solutions,

    M. N. Mejri, J. Ben-Othman, and M. Hamdi, “Survey on vanet security challenges and possible cryptographic solutions,” Vehicular Communica- tions, vol. 1, no. 2, pp. 53–66, 2014

    work page 2014

  3. [3]

    Uber’s self-driving cars were struggling before arizona crash.,

    D. Wakabayashi, “Uber’s self-driving cars were struggling before arizona crash.,” International New York Times , pp. NA–NA, 2018

    work page 2018

  4. [4]

    Multi-modal sensor registra- tion for vehicle perception via deep neural networks,

    M. Giering, V . Venugopalan, and K. Reddy, “Multi-modal sensor registra- tion for vehicle perception via deep neural networks,” in HPEC, pp. 1–6, IEEE, 2015

    work page 2015

  5. [5]

    Multi-modal detection and mapping of static and dynamic obstacles in agriculture for process evaluation,

    T. Korthals, M. Kragh, P. Christiansen, H. Karstoft, R. N. Jørgensen, and U. R ¨uckert, “Multi-modal detection and mapping of static and dynamic obstacles in agriculture for process evaluation,” Frontiers in Robotics and AI, vol. 5, p. 28, 2018

    work page 2018

  6. [6]

    Program-adaptive mutational fuzzing,

    S. K. Cha, M. Woo, and D. Brumley, “Program-adaptive mutational fuzzing,” in Symposium on Security and Privacy , pp. 725–741, IEEE, 2015

    work page 2015

  7. [7]

    Pgfuzz: Policy-guided fuzzing for robotic vehicles.,

    H. Kim, M. O. Ozmen, A. Bianchi, Z. B. Celik, and D. Xu, “Pgfuzz: Policy-guided fuzzing for robotic vehicles.,” in NDSS, 2021

    work page 2021

  8. [8]

    SAFLITE [Online]

    T. Zhu, “ SAFLITE [Online].” Available: https://github.com/ Trusted-AI-in-System-Test/SaFliTe-Test-Interestingness-Predictor, 2024

    work page 2024

Show all 25 references
  1. [9]

    DeepHyperion-UA V at the SBFT Tool Competition 2024-CPS-UA V Test Case Generation Track,

    T. Zohdinasab and A. Doreste, “DeepHyperion-UA V at the SBFT Tool Competition 2024-CPS-UA V Test Case Generation Track,” in Proceed- ings of the 17th International Workshop on SBFT , pp. 49–50, 2024

    2024

  2. [10]

    CAMBA CPS-UA V at the SBFT Tool Competition 2024: CAMBA: Cost-Aware Mutation-Based Test Case Generation for Unmanned Aerial Vehicles,

    M. De Liso and Z. W. Soi, “CAMBA CPS-UA V at the SBFT Tool Competition 2024: CAMBA: Cost-Aware Mutation-Based Test Case Generation for Unmanned Aerial Vehicles,” in Proceedings of the 17th International Workshop on SBFT , pp. 47–48, 2024

    2024

  3. [11]

    TUMB at the SBFT 2024 Tool Competition-CPS-UA V Test Case Generation Track,

    S. Tang, Z. Zhang, A. Cetinkaya, and P. Arcaini, “TUMB at the SBFT 2024 Tool Competition-CPS-UA V Test Case Generation Track,” in Proceedings of the 17th International Workshop on SBFT , pp. 53–54, 2024

    2024

  4. [12]

    SBFT tool competition 2024: CPS-UA V test case generation track,

    S. Khatiri, P. Saurabh, T. Zimmermann, C. Munasinghe, C. Birchler, and S. Panichella, “SBFT tool competition 2024: CPS-UA V test case generation track,” in 17th International Workshop on SBFT , ZHAW Z¨urcher Hochschule f ¨ur Angewandte Wissenschaften, 2024

    2024

  5. [13]

    PX4 autopilot [Online]

    P. D. Team, “ PX4 autopilot [Online].” Available: https://px4.io/, 2012

    2012

  6. [14]

    Training language models to follow instructions with human feedback,

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, et al., “Training language models to follow instructions with human feedback,” Advances in NeurIPS , vol. 35, pp. 27730–27744, 2022

    2022

  7. [15]

    Mistral 7b,

    A. Q. Jiang, A. Sablayrolles, A. Mensch, C. Bamford, D. S. Chaplot, D. d. l. Casas, F. Bressand, G. Lengyel, G. Lample, L. Saulnier, et al. , “Mistral 7b,” arXiv preprint arXiv:2310.06825 , 2023

    2023 arXiv

  8. [16]

    Llama 2: Open foundation and fine-tuned chat models,

    H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y . Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale, et al., “Llama 2: Open foundation and fine-tuned chat models,”arXiv preprint arXiv:2307.09288, 2023

    2023 arXiv

  9. [17]

    ArduPilot [Online]

    A. D. Team, “ ArduPilot [Online].” Available: https://ardupilot.org/, 2009

    2009

  10. [18]

    Coverage-based scene fuzzing for virtual autonomous driving testing,

    Z. Hu, S. Guo, Z. Zhong, and K. Li, “Coverage-based scene fuzzing for virtual autonomous driving testing,” arXiv preprint arXiv:2106.00873, 2021

    2021 arXiv

  11. [19]

    Adaptive test gener- ation for unmanned aerial vehicles using WOGAN-UA V,

    J. Winsten, V . Soloviev, J. Peltom ¨aki, and I. Porres, “Adaptive test gener- ation for unmanned aerial vehicles using WOGAN-UA V,” inProceedings of the 17th ACM/IEEE International Workshop on SBFT, pp. 43–44, 2024

    2024

  12. [20]

    Wasserstein generative adver- sarial networks,

    M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein generative adver- sarial networks,” in ICML, pp. 214–223, PMLR, 2017

    2017

  13. [21]

    Neural network guided evolutionary fuzzing for finding traffic violations of autonomous vehicles,

    Z. Zhong, G. Kaiser, and B. Ray, “Neural network guided evolutionary fuzzing for finding traffic violations of autonomous vehicles,” IEEE Transactions on Software Engineering , 2022

    2022

  14. [22]

    On the difficulty of training recurrent neural networks,

    R. Pascanu, T. Mikolov, and Y . Bengio, “On the difficulty of training recurrent neural networks,” in ICML, pp. 1310–1318, Pmlr, 2013

    2013

  15. [23]

    Reality bites: Assessing the realism of driving scenarios with large language models,

    J. Wu, C. Lu, A. Arrieta, T. Yue, and S. Ali, “Reality bites: Assessing the realism of driving scenarios with large language models,” arXiv preprint arXiv:2403.09906, 2024

    2024 arXiv

  16. [24]

    Language prompt for autonomous driving,

    D. Wu, W. Han, T. Wang, Y . Liu, X. Zhang, and J. Shen, “Language prompt for autonomous driving,” arXiv preprint arXiv:2309.04379, 2023

    2023 arXiv

  17. [25]

    Hilm-d: Towards high-resolution understanding in multimodal large language models for autonomous driving,

    X. Ding, J. Han, H. Xu, W. Zhang, and X. Li, “Hilm-d: Towards high-resolution understanding in multimodal large language models for autonomous driving,” arXiv preprint arXiv:2309.05186 , 2023. APPENDIX A. Bugs Information Provided to SAFLITE We selected information related to ...

    2023 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.