Pith. sign in

REVIEW 4 major objections 8 minor 32 references

Research on Driving Scenario Technology Based on Multimodal Large Lauguage Model Optimization

T0 review · 4 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A small 7B vision-language model can nearly match a 72B teacher on driving-scene perception when its prompts, training data, and quantization are jointly optimized.

desk verdict A plausible VLM-driving engineering recipe whose headline accuracy gain is inflated by prompt tuning on the test set; the numbers need a clean holdout and artifacts before they can be trusted. read the letter →

arxiv 2506.02014 v1 pith:ZQUCVYJU submitted 2025-05-28 cs.CV cs.AI

Wang Mengjie , Zhu Huiping , Li Jian , Shi Wenxiu , Zhang Song This is my paper
classification cs.CVcs.AI
keywords MultimodalLargeLanguageModelText-to-imageGenerationPromptEngineeringDistillationFine-tuningAutonomousDrivingQuantizationSyntheticData
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

This paper argues that a small 7-billion-parameter vision-language model can be made accurate enough for driving-scene perception by combining four optimizations: prompts that adapt to each image's content, a training set that mixes mined real data with synthetic images, distillation from a larger teacher with low-rank fine-tuning, and activation-aware quantization. On the authors' test set, the average accuracy on cone detection, traffic light status, speed limit recommendation, and intersection alerts rises from 0.542 for the open-source quantized model to 0.894 for the full pipeline, at the same inference speed. If the result transfers, autonomous-driving perception tasks could run on edge hardware at near-teacher accuracy, without the cost of a large model. The authors also report that adding synthetic data after real-data training improves the average by about 0.022.

What carries the argument

The load-bearing mechanism is a closed loop that starts with a dynamic prompt optimizer: the model's outputs on test inputs are compared against expected results, and prompts are revised iteratively until performance criteria are met, so the final prompt is matched to the task and image content. This feeds a dataset pipeline in which real frames are pseudo-labeled by a vision-language model with spatiotemporal consistency checks, and synthetic data is produced both by a self-improving text-to-image model (fine-tuned from Kolors) and by procedural Blender rendering with traffic layouts. The training step distills knowledge from a 72B teacher into a 7B student using a hybrid soft-label and hard-label loss, with LoRA (low-rank adaptation) fine-tuning of key layers, and the resulting model is quantized with AWQ (activation-aware weight quantization), which chooses weight quantization levels from activation distributions. Each component is intended to preserve accuracy while cutting storage and compute.

What would settle it

Re-run the full pipeline with a strict train/validation/test split, tuning prompts and hyperparameters on the validation set only, and measure accuracy on the held-out test set; if the gap to the open-source quantized model shrinks well below the reported 0.352 or the ordering changes, the central claim is not robust. A second check is to apply the final pipeline, without retuning, to an independent public driving-scene dataset and see whether the accuracy level persists.

Watch

Extended reading notes

Core claim

The central claim is that the accuracy gap between a 7B quantized open-source model and its 72B teacher on four driving-scene tasks can be almost closed by a deployment-oriented pipeline. The pipeline's components are a dynamic prompt selected from a prompt library based on image content, a hybrid dataset combining VLM-pseudo-labeled real data with synthetic images from a fine-tuned text-to-image model and a Blender rendering engine, a hybrid of knowledge distillation and LoRA fine-tuning, and AWQ quantization. The reported numbers are an average accuracy of 0.894 for the full pipeline versus 0.542 for the open-source quantized 7B model, and 0.944 for the 72B teacher, with FPS of 11 versus 9 for the 7B model. The authors interpret this as showing that a small model can reach near-teacher accuracy while remaining efficient enough for end-side deployment.

Load-bearing premise

The reported accuracy gains assume that the test set was never used to select the prompts, LoRA settings, or data filters; the paper never describes a separate holdout, and its prompt-optimization section says the authors fed 'a series of test data' into the model and adjusted prompts until the output 'meets the predetermined performance criteria,' so part of the improvement could be an artifact of fitting that test set.

Editorial extensions

If this is right

  • A 7B quantized model with the full pipeline matches the 72B teacher's accuracy within about 0.05 on these tasks, making near-teacher performance attainable on edge hardware.
  • The same pipeline is portable to other vision-language verticals where a small model must replace a large one under compute limits.
  • Synthetic data from text-to-image and rendering together contribute roughly 0.022 average accuracy and specifically improve intersection and traffic-light F1, reducing the need for manual annotation.
  • Dynamic prompt optimization alone lifts the 7B model's average from 0.484 to 0.580, indicating prompt design is a high-leverage, low-cost optimization step.

Reading between the lines

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

  • A retest with a proper holdout would likely shrink the reported gain, but the prompt-optimization effect may remain substantial because the paper's Table 3 ablation also shows large gains from prompts alone.
  • Because the same model family acts as pseudo-labeler, teacher, and student, the pseudo-labels inherit the teacher's biases; the reported accuracy may partly reflect self-confirmation rather than independent ground truth.
  • A cheap testable extension is to run the dynamic-prompt loop on simulated scenes with known ground truth to separate prompt-selection gains from model-capacity gains.
  • The paper's data suggests pure rendering gives larger accuracy gains than text-to-image generation (2.23% vs 1.16% when added separately), so a natural next experiment is to scale rendering diversity and measure whether the gap persists.
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

4 major / 8 minor

Summary. The paper proposes a comprehensive pipeline for adapting multimodal large language models (specifically Qwen2.5-VL) to driving-scenario tasks: cone detection, traffic light recognition, speed limit recommendation, and intersection alerts. The pipeline combines dynamic prompt optimization (adjusting prompts based on image content), construction of a hybrid real/synthetic dataset (via self-collected data mining, text-to-image generation, and Blender-based rendering), model training with knowledge distillation and LoRA fine-tuning, and AWQ quantization for deployment. Experiments on self-collected and synthetic test sets report large average-accuracy gains over a quantized 7B baseline (e.g., from 0.542 to 0.894 in Table 2) and additional small gains when synthetic data is used for fine-tuning (Table 4). The paper claims the method significantly improves task accuracy while maintaining efficient resource utilization.

Significance. If the reported results are reliable, this paper provides a practical, systems-level recipe for deploying compact multimodal models in driving scenarios, with a useful combination of dynamic prompting, synthetic data, and efficient training/quantization. The paper covers an application domain of current interest and brings together several existing techniques in a coherent pipeline. However, the current evidence is weakened by the lack of a clear holdout protocol for prompt tuning, the absence of error bars or repeated runs, and arithmetic inconsistencies in the headline numbers. The paper does not provide code or data, so the reproducibility of the quantitative claims cannot be assessed. These issues limit the evidential value of the results as presented.

major comments (4)
  1. [Section 3.1 and Tables 2–4] The dynamic prompt optimization procedure is described as an iterative loop in which the authors "input a series of test data into the large language model" and adjust prompts until "the model's output meets the predetermined performance criteria." The paper never states that the test sets used for the evaluations in Tables 2–4 were held out from this prompt-selection loop. Since Table 3 shows that prompt optimization alone increases average accuracy from 0.484 to 0.580 (Δ = 0.096), and Table 2's headline gain is measured against a baseline that appears to use the original prompt, a substantial portion of the reported improvement may be due to test-set-specific prompt tuning. The authors must either describe a strict separation between development data and evaluation data or re-run the evaluation on a genuinely unseen holdout set.
  2. [Section 4.2, Tables 2 and 4] The reported improvements do not match the table arithmetic. Section 4.2 states the average accuracy increased by 0.322 compared to the open-source quantized model, but Table 2 shows 0.894 − 0.542 = 0.352. Similarly, Section 4.2 says that fine-tuning with synthetic data increased average accuracy by "about 0.022," but Table 4 shows 0.894 − 0.864 = 0.030. The authors must reconcile these discrepancies; if the tables are correct, the text numbers should be corrected, and if the text is correct, the table values need to be updated.
  3. [Tables 2–4] All quantitative results are point estimates from a single test set with no error bars, confidence intervals, or multiple runs. Given that Table 4 reports improvements as small as 0.011 (Traffic Lights P) and 0.005 (Obstacles R), the claim that synthetic data improves performance cannot be distinguished from noise without variance estimates. The authors should provide repeated-run statistics or error bars for the main metrics.
  4. [Sections 3.3, 3.4, 4.1] The experimental description is insufficient for reproducibility. The paper does not specify the LoRA rank and alpha, the distillation loss weighting or temperature, the AWQ calibration set or bit-width, the contents of the prompt library and selection rules, or the training hyperparameters (learning rate, epochs, optimizer). For a paper whose contribution is a "comprehensive method," this level of detail is necessary for the scientific record.
minor comments (8)
  1. [Title] The title contains a typo: "Lauguage" should be "Language."
  2. [Table 4] The column header "Pure Renderinf" should read "Pure Rendering."
  3. [Section 2.6] The parenthetical citation "(Jain et al., 2021)" appears in the text but has no corresponding entry in the reference list; either add the citation or remove the parenthetical.
  4. [Throughout] The acronym "A WQ" is typeset with a space throughout the manuscript; it should be written consistently as "AWQ."
  5. [Section 4.4.1] The phrase "an average precision increase of about 10%" is ambiguous. Table 3 shows an absolute increase of 0.096 (from 0.484 to 0.580), which is a relative improvement of 19.8%. Please clarify which quantity is meant.
  6. [Tables 2–4] The "Average" row is not defined; specify whether it is the mean of the F1 scores for the three classification tasks and R² for speed regression, or some other combination, so that readers can interpret the headline numbers correctly.
  7. [Table 1 and Section 4.2] Table 1 lists separate test sets for real, text-to-image, and pure rendering data, but the paper does not explicitly state which test sets are used in each of Tables 2–4. Clarify the relationship between the test splits and the prompt-optimization loop.
  8. [Section 4.2] The phrase "without reducing the inference speed of the quantized model" is confusing because Table 2 shows that FPS increases from 9 to 11 for the authors' strategy; please rephrase to "while maintaining or improving inference speed."

Circularity Check

1 steps flagged · score 4.0 of 10

Prompt optimization is tuned on data called 'test data' (Sec. 3.1), so the Table 3 before/after gain is partly a fitting artifact; the rest of the empirical comparison is self-contained.

  1. fitted input called prediction [Section 3.1 (Dynamic Prompt Optimization); Figure 2; Table 3 in Section 4.4.1]
    "Initially, based on the preliminary prompts, we input a series of test data into the large language model and carefully observe the outputs. By comparing the differences between the expected goals and actual outputs, we make targeted adjustments to the prompts, such as clarifying task requirements, refining descriptions, or adding example guidance. Then, we conduct experiments with the updated prompts and repeat the process until the model's output meets the predetermined performance criteria."

    The dynamic prompt optimization loop is a fitting process: prompt text is adjusted iteratively until outputs on data explicitly labeled 'test data' meet predetermined performance targets. The gain attributed to prompt optimization in Table 3 (average 0.580 after optimization vs. 0.484 for the original prompt) is then reported on the same data used for that adjustment. The paper never states that Tables 2–4 were evaluated on a holdout split not seen during prompt selection, and Table 1 lists only one 'Test dataset'. Thus the prompt-ablation improvement is partly an in-sample fit, not an out-of-sample prediction. The main pipeline comparison (Table 2) retains independent external baselines, so the circularity is limited to this ablation rather than the entire system.

full rationale

This is primarily an empirical engineering paper, not a formal derivation, so most claimed components are self-contained: the baselines are official open-source Qwen2.5-VL models, and the final comparison in Table 2 uses an external quantized 7B model as the reference point. No load-bearing self-citation, imported uniqueness theorem, or ansatz-smuggling pattern was found. The one concrete reduction is the dynamic-prompt ablation: prompts are selected by iterating on data called 'test data' until performance criteria are met, and then the same data are used to report the before/after accuracy gain in Table 3. This makes that specific gain of about 0.096 average accuracy partly a fitting artifact. Because this contamination touches a component-level claim rather than the entire external comparison, the overall circularity score is moderate-low at 4/10.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim is an empirical engineering result, so the ledger records unreported hand-chosen configuration items and assumptions about data quality and evaluation integrity. No new theoretical entities are introduced.

free parameters (5)
  • Dynamic prompt template library and selection rules
    Section 3.1 and Figure 2: prompts are edited iteratively based on model outputs on test data until criteria are met; no final templates or criteria are reported.
  • LoRA rank, alpha, and target layers
    Section 3.3 mentions LoRA fine-tuning but does not report rank, alpha, or which layers are adapted; these choices directly affect task accuracy.
  • Distillation loss weighting and temperature
    Section 3.3 defines soft and hard label losses but no coefficients or temperature; the balance controls how much teacher knowledge transfers.
  • AWQ quantization bit-width and calibration set
    Section 3.4 selects AWQ after ablations but does not report the bit-width, group size, or calibration data used.
  • Pseudo-label filtering thresholds
    Section 3.2.1 uses a sliding time window and motion-state consistency checks without specifying the window size or consistency rules.
assumptions (5)
  • standard math KL divergence is a valid objective for transferring teacher probability distributions to the student (Eq. 1).
    Standard distillation foundation; invoked in Section 3.3 without proof.
  • domain assumption Qwen2.5-VL teacher pseudo-labels and expert detector outputs are accurate enough to serve as training targets.
    Section 3.2.1 builds the real-data training set from these pseudo-labels; if they are systematically wrong, the improved accuracy claim is unsupported.
  • domain assumption Synthetic data from Kolors text-to-image and Blender rendering is representative of real driving conditions and improves generalization.
    Section 3.2.2 and Table 4 attribute accuracy gains to synthetic data; no external validation shows the synthetic distribution matches deployment.
  • ad hoc to paper The data used to tune dynamic prompts is disjoint from the final evaluation set.
    Section 3.1 says test data is used for iterative prompt adjustment, but no holdout split is stated; this assumption is required for the reported gains to be unbiased.
  • domain assumption AWQ quantization preserves task accuracy at the chosen precision for this model and task suite.
    Section 3.4 relies on the cited AWQ result [32] and reports only a small drop; no calibration or bit-width details support this for the specific tasks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Research on Driving Scenario Technology Based on Multimodal Large Lauguage Model Optimization." pith.science (2026). https://pith.science/paper/ZQUCVYJU

@misc{pith2026250602014,
  author       = {Pith},
  title        = {Pith review of: Research on Driving Scenario Technology Based on Multimodal Large Lauguage Model Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZQUCVYJU}},
  note         = {Machine review of arXiv:2506.02014}
}
read the original abstract

With the advancement of autonomous and assisted driving technologies, higher demands are placed on the ability to understand complex driving scenarios. Multimodal general large models have emerged as a solution for this challenge. However, applying these models in vertical domains involves difficulties such as data collection, model training, and deployment optimization. This paper proposes a comprehensive method for optimizing multimodal models in driving scenarios, including cone detection, traffic light recognition, speed limit recommendation, and intersection alerts. The method covers key aspects such as dynamic prompt optimization, dataset construction, model training, and deployment. Specifically, the dynamic prompt optimization adjusts the prompts based on the input image content to focus on objects affecting the ego vehicle, enhancing the model's task-specific focus and judgment capabilities. The dataset is constructed by combining real and synthetic data to create a high-quality and diverse multimodal training dataset, improving the model's generalization in complex driving environments. In model training, advanced techniques like knowledge distillation, dynamic fine-tuning, and quantization are integrated to reduce storage and computational costs while boosting performance. Experimental results show that this systematic optimization method not only significantly improves the model's accuracy in key tasks but also achieves efficient resource utilization, providing strong support for the practical application of driving scenario perception technologies.

Figures

Figures reproduced from arXiv: 2506.02014 by the authors.

Figure 1
Figure 1. The overall architecture diagram proposed in this paper. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Dynamic Prompt Iterative Optimization Process. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. The model’s output results before and after dynamic Prompt opti [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: The flowchart of the self-Collected Real dataset construction process. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: The images generated using text prompts by text-to-image methods. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: The text-to-image generation process employed in this paper. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Images generated using pure rendering method. [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: The data synthesis process using pure rendering methods. [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Long-tail scenarios of traffic lights that are easily confused. [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 16 canonical work pages

  1. [1]

    A survey of au- tonomous driving: Common practices and emerging technologies,

    E. Yurtsever, J. Lambert, A. Carballo, and K. Takeda, “A survey of au- tonomous driving: Common practices and emerging technologies,” IEEE access, vol. 8, pp. 58 443–58 469, 2020

    work page 2020

  2. [2]

    End-to-end autonomous driving: Challenges and frontiers,

    L. Chen, P. Wu, K. Chitta, B. Jaeger, A. Geiger, and H. Li, “End-to-end autonomous driving: Challenges and frontiers,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2024

    2024

  3. [3]

    Multi-modal fusion transformer for end-to-end autonomous driving,

    A. Prakash, K. Chitta, and A. Geiger, “Multi-modal fusion transformer for end-to-end autonomous driving,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 7077–7087

    2021

  4. [4]

    Scene-adaptive and region-aware multi-modal prompt for open vocabulary object detection,

    X. Zhao, X. Liu, D. Wang, Y. Gao, and Z. Liu, “Scene-adaptive and region-aware multi-modal prompt for open vocabulary object detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 16 741–16 750

    work page 2024

  5. [5]

    Pseudoprop: Robust pseudo-label generation for semi-supervised object detection in autonomous driving systems,

    S. Hu, C.-H. Liu, J. Dutta, M.-C. Chang, S. Lyu, and N. Ramakrishnan, “Pseudoprop: Robust pseudo-label generation for semi-supervised object detection in autonomous driving systems,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Work- shops, June 2022, pp. 4390–4398

    work page 2022

  6. [6]

    Vlm-mpc: Vision language foundation model (vlm)-guided model predictive controller (mpc) for autonomous driving,

    K. Long, H. Shi, J. Liu, and X. Li, “Vlm-mpc: Vision language foundation model (vlm)-guided model predictive controller (mpc) for autonomous driving,” arXiv preprint arXiv:2408.04821 , 2024

    arXiv 2024

  7. [7]

    Contextvlm: Zero-shot and few-shot context understanding for autonomous driving using vision language models,

    S. Sural, R. R. Rajkumar et al. , “Contextvlm: Zero-shot and few-shot context understanding for autonomous driving using vision language models,” in 2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC) . IEEE, 2024, pp. 468–475

    work page 2024

  8. [8]

    Generative adversarial networks,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial networks,” Communi- cations of the ACM , vol. 63, no. 11, pp. 139–144, 2020. 19

    work page 2020

Show all 32 references
  1. [9]

    Dall-e: Creating images from text,

    M. D. M. Reddy, M. S. M. Basha, M. M. C. Hari, and M. N. Penchalaiah, “Dall-e: Creating images from text,” UGC Care Group I Journal , vol. 8, no. 14, pp. 71–75, 2021

    2021

  2. [10]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems , vol. 33, pp. 6840–6851, 2020

    2020

  3. [11]

    Diffusion models beat gans on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,” Advances in neural information processing systems , vol. 34, pp. 8780–8794, 2021

    2021

  4. [12]

    Photorealistic text-to-image diffusion models with deep language under- standing,

    C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. L. Denton, K. Ghasemipour, R. Gontijo Lopes, B. Karagol Ayan, T. Salimans et al., “Photorealistic text-to-image diffusion models with deep language under- standing,” Advances in neural information processing systems , vol. 35...

    2022

  5. [13]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 10 684–10 695

    2022

  6. [14]

    Kolors: Effective training of diffusion model for photorealistic text-to-image synthesis,

    K. Team, “Kolors: Effective training of diffusion model for photorealistic text-to-image synthesis,” arXiv preprint, 2024

    2024

  7. [15]

    Carla: An open urban driving simulator,

    A. Dosovitskiy, G. Ros, F. Codevilla, A. Lopez, and V. Koltun, “Carla: An open urban driving simulator,” in Conference on robot learning. PMLR, 2017, pp. 1–16

    2017

  8. [16]

    Airsim: High-fidelity visual and physical simulation for autonomous vehicles,

    S. Shah, D. Dey, C. Lovett, and A. Kapoor, “Airsim: High-fidelity visual and physical simulation for autonomous vehicles,” in Field and Service Robotics: Results of the 11th International Conference . Springer, 2018, pp. 621–635

    2018

  9. [17]

    Mars: An instance-aware, modular and realistic simulator for autonomous driving,

    Z. Wu, T. Liu, L. Luo, Z. Zhong, J. Chen, H. Xiao, C. Hou, H. Lou, Y. Chen, R. Yang et al. , “Mars: An instance-aware, modular and realistic simulator for autonomous driving,” in CAAI International Conference on Artificial Intelligence. Springer, 2023, pp. 3–15

    2023

  10. [18]

    Editable scene simulation for autonomous driving via collaborative llm- agents,

    Y. Wei, Z. Wang, Y. Lu, C. Xu, C. Liu, H. Zhao, S. Chen, and Y. Wang, “Editable scene simulation for autonomous driving via collaborative llm- agents,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 15 077–15 087

    2024

  11. [19]

    Bevcontrol: Accu- rately controlling street-view elements with multi-perspective consistency via bev sketch layout,

    K. Yang, E. Ma, J. Peng, Q. Guo, D. Lin, and K. Yu, “Bevcontrol: Accu- rately controlling street-view elements with multi-perspective consistency via bev sketch layout,” arXiv preprint arXiv:2308.01661 , 2023. 20

    2023 arXiv

  12. [20]

    Synscapes: A photorealistic synthetic dataset for street scene parsing,

    M. Wrenninge and J. Unger, “Synscapes: A photorealistic synthetic dataset for street scene parsing,” arXiv preprint arXiv:1810.08705 , 2018

    2018 arXiv

  13. [21]

    All for one, and one for all: Urbansyn dataset, the third musketeer of synthetic driving scenes,

    J. L. G´ omez, M. Silva, A. Seoane, A. Borr´ as, M. Noriega, G. Ros, J. A. Iglesias-Guitian, and A. M. L´ opez, “All for one, and one for all: Urbansyn dataset, the third musketeer of synthetic driving scenes,” arXiv preprint arXiv:2312.12176, 2023

    2023 arXiv

  14. [22]

    Prompt engineering for chatgpt: a quick guide to techniques, tips, and best practices,

    S. Ekin, “Prompt engineering for chatgpt: a quick guide to techniques, tips, and best practices,” Authorea Preprints, 2023

    2023

  15. [23]

    A sys- tematic survey of prompt engineering in large language models: Techniques and applications,

    P. Sahoo, A. K. Singh, S. Saha, V. Jain, S. Mondal, and A. Chadha, “A sys- tematic survey of prompt engineering in large language models: Techniques and applications,” arXiv preprint arXiv:2402.07927 , 2024

    2024 arXiv

  16. [24]

    Automatic prompt optimization with

    R. Pryzant, D. Iter, J. Li, Y. T. Lee, C. Zhu, and M. Zeng, “Automatic prompt optimization with” gradient descent” and beam search,” arXiv preprint arXiv:2305.03495, 2023

    2023 arXiv

  17. [25]

    Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes,

    C.-Y. Hsieh, C.-L. Li, C.-K. Yeh, H. Nakhost, Y. Fujii, A. Ratner, R. Krishna, C.-Y. Lee, and T. Pfister, “Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes,” arXiv preprint arXiv:2305.02301, 2023

    2023 arXiv

  18. [26]

    An empirical study on parameter-efficient fine-tuning for multimodal large language models,

    X. Zhou, J. He, Y. Ke, G. Zhu, V. Guti´ errez-Basulto, and J. Z. Pan, “An empirical study on parameter-efficient fine-tuning for multimodal large language models,” arXiv preprint arXiv:2406.05130 , 2024

    2024 arXiv

  19. [27]

    Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter,

    V. Sanh, L. Debut, J. Chaumond, and T. Wolf, “Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter,” arXiv preprint arXiv:1910.01108, 2019

    1910 arXiv

  20. [28]

    Survey on knowledge distillation for large language models: methods, eval- uation, and application,

    C. Yang, Y. Zhu, W. Lu, Y. Wang, Q. Chen, C. Gao, B. Yan, and Y. Chen, “Survey on knowledge distillation for large language models: methods, eval- uation, and application,” ACM Transactions on Intelligent Systems and Technology, 2024

    2024

  21. [29]

    Qlora: Efficient finetuning of quantized llms,

    T. Dettmers, A. Pagnoni, A. Holtzman, and L. Zettlemoyer, “Qlora: Efficient finetuning of quantized llms,” Advances in neural information processing systems, vol. 36, pp. 10 088–10 115, 2023

    2023

  22. [30]

    Gptq: Accurate post-training quantization for generative pre-trained transformers,

    E. Frantar, S. Ashkboos, T. Hoefler, and D. Alistarh, “Gptq: Accurate post-training quantization for generative pre-trained transformers,” arXiv preprint arXiv:2210.17323, 2022

    2022 arXiv

  23. [31]

    Overcoming forgetting catas- trophe in quantization-aware training,

    T.-A. Chen, D.-N. Yang, and M.-S. Chen, “Overcoming forgetting catas- trophe in quantization-aware training,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 17 358–17 367. 21

    2023

  24. [32]

    Awq: Activation-aware weight quantization for on-device llm compression and acceleration,

    J. Lin, J. Tang, H. Tang, S. Yang, W.-M. Chen, W.-C. Wang, G. Xiao, X. Dang, C. Gan, and S. Han, “Awq: Activation-aware weight quantization for on-device llm compression and acceleration,” Proceedings of Machine Learning and Systems , vol. 6, pp. 87–100, 2024. 22

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.