Pith. sign in

REVIEW 4 major objections 6 minor 40 references

VLAD: A VLM-Augmented Autonomous Driving Framework with Hierarchical Planning and Interpretable Decision Process

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A fine-tuned VLM that issues high-level driving commands cuts a state-of-the-art planner's average collision rate by about 32 percent on nuScenes.

desk verdict A plausible VLM+end-to-end integration with a headline safety gain that is not yet causal on the evidence presented. read the letter →

arxiv 2507.01284 v1 pith:LV22MLTP submitted 2025-07-02 cs.RO cs.AIcs.CVcs.ETcs.LG

Cristian Gariboldi , Hayato Tokida , Ken Kinjo , Yuki Asada , Alexander Carballo This is my paper · ORCID
classification cs.ROcs.AIcs.CVcs.ETcs.LG
keywords autonomousdrivingvision-languagemodelend-to-endplanningcollisionrateexplainableAInuScenesmeta-actionsfine-tuning
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 VLAD, a hybrid autonomous-driving framework that pairs a fine-tuned vision-language model (VLM) with VAD, a transformer-based end-to-end planner. The VLM looks at surround-view camera images and outputs a high-level navigational command plus a natural-language justification; VAD then turns that command and the images into a trajectory. On the nuScenes benchmark, VLAD's average collision rate is 0.15%, versus 0.22% for VAD-Base and 0.18% for the stronger Senna baseline, reductions the authors report as 31.82% and 16.7%. The paper takes this as evidence that VLM-provided semantic guidance, not additional sensors or rules, can make an end-to-end planner safer while making its decisions interpretable.

What carries the argument

The load-bearing object is the meta-action: a discrete high-level command (e.g., GO_STRAIGHT, TURN_LEFT) that the fine-tuned VLM (CLIP ViT-L/14 vision encoder plus Vicuna-7b language model) emits from surround camera imagery, together with a textual rationale. VAD's planning head consumes that command as the conditioning token $c$ alongside ego, agent, and map queries, and decodes the future trajectory. The idea is that semantic scene understanding—especially attention to vulnerable road users—enters the planner through the command channel rather than through new sensor inputs or hand-written cost functions.

What would settle it

Retrain VAD under the identical two-stage protocol using (a) the original nuScenes command labels and (b) randomized commands, then compare average collision rates on the validation split; if either control matches the 0.15% collision rate, the VLM's meta-actions are not the cause of the improvement.

Watch

Extended reading notes

Core claim

VLAD's central claim is that a frozen end-to-end planner (VAD) becomes measurably safer when its high-level command input is replaced by meta-actions produced by a fine-tuned VLM. The VLM is trained on 365,666 question-answer pairs generated by a larger teacher model, covering perception, prediction, and planning; after that, VAD is retrained end-to-end with the VLM frozen. The result is an average collision rate of 0.15% on nuScenes validation, down from 0.22% for VAD-Base and 0.18% for Senna, at the cost of a slightly higher L2 displacement error. The paper argues this safety-for-geometry trade-off is the signature of the VLM prioritizing collision avoidance, and it also demonstrates that the same fine-tuned VLM can emit concise driving explanations in about 0.8 seconds, which it counts as real-time enough for human oversight.

Load-bearing premise

The paper assumes that the drop in collision rate is caused by the VLM's high-level commands, but it never compares against retraining VAD with the original or random commands under the same protocol, and it reports no error bars or significance tests.

Editorial extensions

If this is right

  • If the reported collision reduction holds, VLM guidance can improve safety metrics of open-loop end-to-end planners without extra sensors.
  • The architecture produces natural-language justifications at about 0.8 seconds per explanation, suggesting that interpretability can be provided at a rate useful for human oversight.
  • The large gap between full fine-tuning and LoRA (BLEU 64.60 versus 19.83; planning accuracy 90.15% versus 44.36%) indicates that parameter-efficient tuning with one epoch is insufficient for driving-domain reasoning.
  • The safety-versus-displacement trade-off implies that collision rate and L2 error should be read together, since a planner can be safer while deviating from ground-truth paths.

Reading between the lines

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

  • A direct test the paper does not run is to retrain VAD on the original human command labels or on random commands under the same two-stage protocol; without that control, the collision reduction cannot be attributed specifically to the VLM's meta-actions rather than to retraining or to the command distribution.
  • The roughly 0.8-second explanation latency suggests the VLM could be used as a continuous back-seat driver that monitors the planner, but only if the meta-action loop itself runs faster than the planner's update rate.
  • The safety gain might be amplified or altered in closed-loop settings, where the vehicle can act on conservative commands earlier rather than being scored against a fixed ground-truth trajectory.
  • Since the QA data is generated by a larger teacher model (LLaVA-34B), the student's ceiling may be set by teacher errors, so distilling from a stronger or ensemble teacher could be a direct extension.
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 / 6 minor

Summary. The paper proposes VLAD, a hybrid autonomous-driving framework that integrates a fine-tuned vision-language model (Vicuna-7B with a CLIP ViT-L/14 vision encoder) with the VAD end-to-end planner. The VLM is trained on a custom question-answer dataset generated by LLaVA-1.6-34B and outputs high-level meta-actions plus natural-language explanations; VAD converts these meta-actions into trajectories. Training is performed in two stages: first the VLM is fine-tuned while VAD is frozen, then VAD is trained end-to-end while the VLM is frozen. Experiments on nuScenes report trajectory L2 error, collision rates, explanation-quality metrics (BLEU, METEOR, ROUGE-L, CIDEr, GPT-Score), meta-action accuracy, and inference times. The headline claims are a 31.82% average collision-rate reduction over VAD-Base, a 16.7% reduction over Senna, and 0.8-second short-form explanations.

Significance. If the collision-reduction claim is causal, the contribution is significant: it would demonstrate that a fine-tuned VLM providing high-level commands can make an end-to-end planner measurably safer without additional sensors, while also unifying planning supervision and interpretability in one architecture. The paper has concrete strengths: it builds on a public benchmark (nuScenes), reports multiple fine-tuning configurations, and describes a clear two-stage training protocol. It also candidly states in the conclusion that closed-loop evaluation is left for future work. However, the central quantitative claims are not yet established, because the experimental protocol lacks the controls and uncertainty quantification needed to attribute the observed collision differences to the VLM, and the explanation-quality evaluation is circular with respect to the training data. Given the load-bearing nature of the headline numbers, the contribution is promising but requires substantial additional experimental support.

major comments (4)
  1. [Section III-D2 and Table I] The collision-reduction claim is not causally identified. In the second training stage, VAD is retrained end-to-end while the VLM is frozen, yet no control condition is reported in which VAD is retrained under the same protocol with the original nuScenes commands, with commands shuffled across scenes, or with random commands drawn from the same marginal distribution. Without such a control, the observed gap between VLAD‡ (0.15%) and VAD-Base (0.22%) or Senna (0.18%) could be produced by the retraining procedure itself, by a different command distribution, or by optimization noise rather than by the VLM's semantic interpretation of the scene. This is the load-bearing experimental gap for the headline safety claim.
  2. [Table I and Section IV-B1] The reported collision differences lack statistical grounding. No seeds, confidence intervals, or significance tests are provided, and the absolute per-horizon differences are small: at 3s, VLAD‡ is 0.33% versus 0.41% for VAD-Base and 0.33% for Senna, which on the nuScenes validation split likely corresponds to a handful of collision events. The LoRA variant VLAD* actually has a worse average collision rate (0.29%) than VAD-Base (0.22%), showing that the outcome is highly sensitive to the fine-tuning method and training protocol. The paper should report multiple seeds with confidence intervals and a statistical test (e.g., a bootstrap over scenarios) before claiming a 31.82% improvement.
  3. [Section III-C and Tables II and III] The planning-accuracy and explanation-quality evaluations are circular. The validation set used in Tables II and III is drawn from the same teacher-generated QA dataset used for fine-tuning, so high BLEU, METEOR, ROUGE-L, CIDEr, and meta-action accuracy largely measure how faithfully the student model replicates LLaVA-1.6-34B's outputs rather than objective correctness. To support the interpretability contribution, the authors should evaluate against independent human annotations or an external benchmark such as DriveBench or DriveLM, or at minimum demonstrate that the reference set was not used during training.
  4. [Section IV-A2 and Section V] The evaluation protocol for collision rates is underspecified, which makes the numbers difficult to interpret or reproduce. The paper does not state whether the planner runs in an open-loop receding-horizon fashion, how the ego trajectory is executed relative to recorded actor trajectories, how collisions are computed (e.g., bounding-box overlap at which timesteps), or which validation scenarios are used. Section V defers closed-loop evaluation to future work; therefore the safety claim should be qualified as open-loop planning performance, not operational safety.
minor comments (6)
  1. [Section IV-A3] The text 'Models were trained for one epoch (an additional full-parameter fine-tuning was conducted for 10 epochs)' is ambiguous; please clarify which entries in Table I correspond to each training budget and whether the 10-epoch run uses the same data or an extended schedule.
  2. [Tables II, III, IV] Tables II and III report point estimates only; standard deviations and the number of evaluated explanations should be given, and the hardware used for the inference-time measurements in Table IV should be specified.
  3. [Figure 3] The caption states that the VLM 'keeps its recommendation to proceed straight' despite an original left-turn command; please clarify where the original command comes from and whether the scenario is representative or selected as a qualitative example.
  4. [Abstract and Introduction] The abstract and introduction state a 31.82% reduction 'compared to baseline methodologies,' but Table I shows the comparison is specifically to VAD-Base; this overstatement should be corrected to name the actual baseline.
  5. [References] Several references lack complete bibliographic information, e.g., [2] Senna and [6] LanguageMPC have no arXiv identifiers, and [10] DriveBench and [24] ADAPT have incomplete entries; please complete these for reproducibility.
  6. [Section III-C and GitHub statement] The manuscript mentions an online repository with dataset samples but does not release the full QA dataset, trained weights, or evaluation scripts; releasing these artifacts would materially strengthen the reproducibility of the central claims.

Circularity Check

2 steps flagged · score 6.0 of 10

Explanation and meta-action metrics are measured against the same teacher-generated QA targets used for fine-tuning; the nuScenes collision-rate claim remains external and non-circular.

  1. self definitional [Sec. IV-B.2 (Table II); Sec. III-C (QA data generation)]
    "We compare VLAD explanations with the ground-truth ones generated offline in our QA Dataset, and different metrics are employed."

    The explanation ground truth is the offline QA Dataset produced by the LLaVA-v1.6-34b teacher, and the same pipeline generated the 365,666 QA pairs used to fine-tune the VLM (Sec. III-C). BLEU, METEOR, ROUGE-L, CIDEr and GPT-Score are therefore computed against the teacher outputs the model was optimized to imitate. The separate 76,930-pair validation set avoids exact overfitting, but the target distribution is the same teacher, so high scores demonstrate imitation fidelity rather than correctness against an independent source; the 'accurate explanations' claim is defined in terms of the training labels.

  2. self definitional [Sec. IV-A.2 and Sec. IV-B.2 (Table III); Sec. III-C (Planning QA generation)]
    "Finally, we compute the accuracy of the model to provide correct meta-actions based on the specific driving scenarios."

    The 'correct' meta-actions are the teacher-generated planning answers from the QA Dataset used in first-stage fine-tuning: Section III-C says the planning QA pairs are constructed from ground-truth observations and future trajectories, ending with the model generating a meta-action. Table III's 90.76% accuracy therefore measures how well the fine-tuned VLM reproduces the teacher's labels, not how often the meta-action is objectively correct in the world. The held-out split reduces memorization but does not break the definitional link between the accuracy metric and the training target.

full rationale

The headline collision reduction (Table I) is not circular: nuScenes collision events are external to the VLM's QA training data and are computed from planned trajectories, so the 31.82% number has independent content. However, the paper's two other quantitative claims -- explanation quality and meta-action planning accuracy -- are evaluated against the same teacher-generated QA dataset that defined the fine-tuning objective. In both cases the 'ground truth' is LLaVA-v1.6-34b's output, so the high BLEU/CIDEr/GPT-Score and 90.76% accuracy values are largely measures of imitation fidelity. The paper contains no load-bearing self-citation chain: VAD, Senna, and the other baselines are external prior work, and no uniqueness theorem is imported from the authors' own papers. The missing control for stage-2 VAD retraining and the absence of error bars are real validity threats to the causal attribution of the collision reduction, but they are experimental-control issues rather than circularity. Because the main safety metric is independent, the circularity is partial, not total; hence the score is 6 rather than higher.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim depends on training choices (full-parameter fine-tuning, QA dataset composition) and on the assumption that open-loop nuScenes metrics and teacher-generated labels are valid evidence. No new physical or conceptual entity is introduced; the meta-action is a discrete command label inherited from navigation systems. The main burden is empirical: the paper does not establish that the VLM, rather than VAD retraining or noise, causes the measured improvement.

free parameters (4)
  • Full-parameter fine-tuning (vs LoRA) and epoch count = full-parameter, 10 epochs for best VLAD; LoRA r=128, alpha=256 for 1 epoch underperforms
    The claimed collision and accuracy improvements appear only after full-parameter fine-tuning; the choice is empirical and not predicted by any theory (Section IV-B, Tables I-III).
  • QA dataset design (teacher model, prompt template, pair counts) = LLaVA-v1.6-34b teacher; 365,666 training / 76,930 validation pairs
    The VLM's behavior is shaped by teacher-generated targets and custom prompts; the authors selected these without a formal justification (Section III-C).
  • VAD second-stage training budget = not specified
    The paper does not report how many epochs or steps VAD is retrained on nuScenes in the second stage, making the comparison to the frozen original VAD ambiguous (Section III-D2).
  • Meta-action set A = discrete commands such as GO_STRAIGHT, TURN_LEFT
    The granularity of the command set is a design choice that determines how much control the VLM has over VAD's planning head (Section III-B).
assumptions (3)
  • domain assumption Open-loop collision rate on nuScenes is a meaningful proxy for real-world driving safety.
    The paper's headline safety claim is based on this metric; no closed-loop evaluation is performed.
  • ad hoc to paper Teacher-generated QA labels are correct enough to serve as both training targets and evaluation references.
    Section III-C: the QA dataset is produced by LLaVA-v1.6-34b with GT positions; Section IV-B2 uses the same generated references as ground truth for BLEU/METEOR/ROUGE/CIDEr/GPT-Score.
  • domain assumption The student VLM, seeing only camera images, can recover the information the teacher had from ground-truth annotations.
    The teacher was prompted with precise agent positions and future trajectories; at inference VLAD is camera-only, and the paper does not quantify this information gap.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VLAD: A VLM-Augmented Autonomous Driving Framework with Hierarchical Planning and Interpretable Decision Process." pith.science (2026). https://pith.science/paper/LV22MLTP

@misc{pith2026250701284,
  author       = {Pith},
  title        = {Pith review of: VLAD: A VLM-Augmented Autonomous Driving Framework with Hierarchical Planning and Interpretable Decision Process},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LV22MLTP}},
  note         = {Machine review of arXiv:2507.01284}
}
read the original abstract

Recent advancements in open-source Visual Language Models (VLMs) such as LLaVA, Qwen-VL, and Llama have catalyzed extensive research on their integration with diverse systems. The internet-scale general knowledge encapsulated within these models presents significant opportunities for enhancing autonomous driving perception, prediction, and planning capabilities. In this paper we propose VLAD, a vision-language autonomous driving model, which integrates a fine-tuned VLM with VAD, a state-of-the-art end-to-end system. We implement a specialized fine-tuning approach using custom question-answer datasets designed specifically to improve the spatial reasoning capabilities of the model. The enhanced VLM generates high-level navigational commands that VAD subsequently processes to guide vehicle operation. Additionally, our system produces interpretable natural language explanations of driving decisions, thereby increasing transparency and trustworthiness of the traditionally black-box end-to-end architecture. Comprehensive evaluation on the real-world nuScenes dataset demonstrates that our integrated system reduces average collision rates by 31.82% compared to baseline methodologies, establishing a new benchmark for VLM-augmented autonomous driving systems.

Figures

Figures reproduced from arXiv: 2507.01284 by the authors.

Figure 1
Figure 1. VLAD Architecture. VLAD consists of two modules: a VLM integrated with an end-to-end system, VAD. The VLM encodes multi-view images and outputs a high-level command along with explanations of the selected behavior, based on the driving scenario. VAD processes the high-level command along with camera images and produces the final trajectory. a) Ego–Agent Interaction: Q ′ ego = TransDec(q = Qego, k = Qa, v = Qa, qpos … view at source ↗
Figure 2
Figure 2. Comparison of different LLaVA models. The image shows a real-world sample data from nuScenes, where two models, namely LLaVA-v1.6-34B and LLaVA-1.6-Mistral-7B, are prompted to generate a scene description as driving agents. LLaVA-v1.6-34B is able to generate an accurate description, analyzing well the driving setting. Conversely, LLaVA-1.6-Mistral-7B, is not able to correctly detect important details like the car in… view at source ↗
Figure 3
Figure 3. VLAD Safety-First Approach Demonstration. Despite an original high-level command to turn left, the VLM detects pedestrians crossing the intersection and keeps its recommendation to ”proceed straight”, postponing the execution of the turn until the road is clear, as also justified in its explanation. This adaptation enables the end-to-end system to focus on the vulnerable road users, reducing vehicle speed and preven… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 40 canonical work pages

  1. [1]

    V AD: Vectorized Scene Represen- tation for Efficient Autonomous Driving,

    Bo Jiang, Shaoyu Chen, Qing Xu, “V AD: Vectorized Scene Represen- tation for Efficient Autonomous Driving,” ICCV , 2023

    work page 2023

  2. [2]

    Bo Jiang, Shaoyu Chen, Bencheng Liao, “Senna: Bridging Large Vision- Language Models and End-to-End Autonomous Driving,“ arXiv, 2024

    work page 2024

  3. [3]

    Haotian Liu, Chunyuan Li, Qingyang Wu, “Visual Instruction Tuning,“ NeurIPS, 2023

    work page 2023

  4. [4]

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, “Judging LLM-as-a- Judge with MT-Bench and Chatbot Arena,“ NeurIPS, 2023

    work page 2023

  5. [5]

    nuScenes: A multimodal dataset for autonomous driving,

    Holger Caesar, Varun Bankiti, Alex H. Lang, “nuScenes: A multimodal dataset for autonomous driving,” CVPR, 2020

    work page 2020

  6. [6]

    LanguageMPC: Large Language Models as Decision Makers for Autonomous Driving,

    Hao Sha, Yao Mu, Yuxuan Jiang, “LanguageMPC: Large Language Models as Decision Makers for Autonomous Driving,” arXiv, 2025

    work page 2025

  7. [7]

    DriveGPT4: Interpretable End-to-end Autonomous Driving via Large Language Model,

    Zhenhua Xu, Yujia Zhang, Enze Xie, “DriveGPT4: Interpretable End-to-end Autonomous Driving via Large Language Model,” IEEE ROBOTICS AND AUTOMATION LETTERS, 2024

    work page 2024

  8. [8]

    LingoQA: Visual Question Answering for Autonomous Driving,

    Ana-Maria Marcu, Long Chen, Jan H ¨unermann, “LingoQA: Visual Question Answering for Autonomous Driving,” ECCV , 2024

    work page 2024

Show all 40 references
  1. [9]

    Chonghao Sima, Katrin Renz, Kashyap Chitta, “DriveLM: Driving with Graph Visual Question Answering,“ ECCV , 2024

    2024

  2. [10]

    Shaoyuan Xie, Lingdong Kong, Yuhao Dong, “Are VLMs Ready for Autonomous Driving? An Empirical Study from the Reliability, Data, and Metric Perspectives,“ arXiv, 2025

    2025

  3. [11]

    Alexey Dosovitskiy, German Ros, Felipe Codevilla, “CARLA: An Open Urban Driving Simulator,“ CoRL, 2017

    2017

  4. [12]

    Tsun-Hsuan Wang, Alaa Maalouf, Wei Xiao, “Drive Anywhere: Gener- alizable End-to-end Autonomous Driving with Multi-modal Foundation Models,“ arXiv, 2024

    2024

  5. [13]

    Hao Shao, Yuxuan Hu, Letian Wang, “LMDrive: Closed-Loop End-to- End Driving with Large Language Models,“ arXiv, 2024

    2024

  6. [14]

    Katrin Renz, Long Chen, Ana-Maria Marcu, “CarLLaV A: Vision lan- guage models for camera-only closed-loop driving,“ arXiv, 2024

    2024

  7. [15]

    Wenhai Wang, Jiangwei Xie, ChuanYang Hu, “DriveMLM: Aligning Multi-Modal Large Language Models with Behavioral Planning States for Autonomous Driving,“ arXiv, 2023

    2023

  8. [16]

    Xiaoyu Tian, Junru Gu, Bailin Li, “DriveVLM: The Convergence of Autonomous Driving and Large Vision-Language Models,“ CoRL, 2024

    2024

  9. [17]

    Zilin Huang, Zihao Sheng, Yansong Qu, “VLM-RL: A Unified Vision Language Models and Reinforcement Learning Framework for Safe Autonomous Driving,“ arXiv, 2024

    2024

  10. [18]

    Katrin Renz, Long Chen, Elahe Arani, “SimLingo: Vision-Only Closed- Loop Autonomous Driving with Language-Action Alignment,“ CVPR, 2025

    2025

  11. [19]

    Xiangru Mu, Tong Qin, Songan Zhang, “Pix2Planning: End-to-End Planning by Vision-language Model for Autonomous Driving on Carla Simulator,“ IEEE Intelligent Vehicles Symposium (IV), 2024

    2024

  12. [20]

    Jyh-Jing Hwang, Runsheng Xu, Hubert Lin, “EMMA: End-to-End Multimodal Model for Autonomous Driving,“ arXiv, 2024

    2024

  13. [21]

    Shuo Xing, Chengyuan Qian, Yuping Wang, “OpenEMMA: Open- Source Multimodal Model for End-to-End Autonomous Driving,“ LLVM-AD, 2025

    2025

  14. [22]

    Yi Xu, Yuxin Hu, Zaiwei Zhang, “VLM-AD: End-to-End Autonomous Driving through Vision-Language Model Supervision,“ arXiv, 2024

    2024

  15. [23]

    Rui Zhao, Qirui Yuan, Jinyu Li, “DriveLLaV A: Human-Level Behavior Decisions via Vision Language Model,“ Sensors (Basel), 2024

    2024

  16. [24]

    Bu Jin, Xinyu Liu, Yupeng Zheng, “ADAPT: Action-aware Driving Caption Transformer,“ ICRA, 2023

    2023

  17. [25]

    Ding, Kairui, Chen, “Hint-AD: Holistically Aligned Interpretability in End-to-End Autonomous Driving,“ CoRL, 2024

    2024

  18. [26]

    Yihan Hu, Jiazhi Yang, Li Chen, “Planning-Oriented Autonomous Driving,“ CVPR, 2023

    2023

  19. [27]

    Pei Liu, Haipeng Liu, Haichao Liu, “VLM-E2E: Enhancing End-to-End Autonomous Driving with Multimodal Driver Attention Fusion,“ arXiv, 2025

    2025

  20. [28]

    Alec Radford, Jong Wook Kim, Chris Hallacy, “Learning Transferable Visual Models From Natural Language Supervision,“ International Con- ference on Machine Learning, 2021

    2021

  21. [29]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, “Attention Is All You Need,“ NeurIPS, 2017

    2017

  22. [30]

    Papineni, S

    K. Papineni, S. Roukos, T. Ward, “Bleu: a method for automatic evaluation of machine translation,“ ACL, 2002

    2002

  23. [31]

    Banerjee and A

    S. Banerjee and A. Lavie, “Meteor: An automatic metric for mt evaluation with improved correlation with human judgments,“ ACL, 2005

    2005

  24. [32]

    Chin-Yew Lin, “ROUGE: A Package for Automatic Evaluation of Summaries,“ Text summarization branches, 2004

    2004

  25. [33]

    Vedantam, C

    R. Vedantam, C. Lawrence Zitnick, and D. Parikh, “Cider: Consensus- based image description evaluation,“ CVPR, 2015

    2015

  26. [34]

    Jinlan Fu, See-Kiong Ng, Zhengbao Jiang, “GPTScore: Evaluate as You Desire,“ arXiv, 2023

    2023

  27. [35]

    Hu, Yelong Shen, Phillip Wallis, “LoRA: Low-Rank Adap- tation of Large Language Models,“ ICLR, 2022

    Edward J. Hu, Yelong Shen, Phillip Wallis, “LoRA: Low-Rank Adap- tation of Large Language Models,“ ICLR, 2022

    2022

  28. [36]

    S. Hu, L. Chen, P. Wu, “St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning,“ ECCV , 2022

    2022

  29. [37]

    Codevilla, E

    F. Codevilla, E. Santana, A. M. Lopez, “Exploring the Limitations of Behavior Cloning for Autonomous Driving,“ ICCV , 2019

    2019

  30. [38]

    D. Chen, B. Zhou, V . Koltun, “Learning by Cheating,“ CoRL, 2020

    2020

  31. [39]

    Peiyun Hu, Aaron Huang, John Dolan, “Safe Local Motion Planning with Self-Supervised Freespace Forecasting,“ CVPR, 2021

    2021

  32. [40]

    Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, “ZeRO: Memory Optimizations Toward Training Trillion Parameter Models,“ SC20, 2020

    2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.