Pith. sign in

REVIEW 4 major objections 4 minor 4 cited by

A compact trajectory predictor distilled from a high-capacity teacher, then refined with traffic safety rewards, runs in real time on edge hardware without large accuracy loss.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 22:50 UTC pith:WHTEK4GB

load-bearing objection Solid systems packaging of known modules for edge multi-agent prediction; headline Jetson numbers matter if the missing body backs them, but PPO-as-bias-correction remains the softest claim. the 4 major comments →

arxiv 2604.10169 v3 pith:WHTEK4GB submitted 2026-04-11 cs.AI cs.LG

MAVEN-T: Reinforced Heterogeneous Distillation for Real-Time Multi-Agent Trajectory Prediction

Wenchang Duan , Zhenguo Gao , Jinguo Xian , Yi Shi This is my paper
classification cs.AI cs.LG
keywords trajectory predictionknowledge distillationreinforcement learningmulti-agent systemsautonomous drivingedge deploymentmixture of expertsproximal policy optimization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Autonomous cars must forecast how nearby vehicles will move so they can plan safely, but the most accurate predictors are often too heavy for on-board chips. This paper argues that the usual fix—static knowledge distillation from a large teacher to a small student—leaves safety-relevant teacher mistakes uncorrected. MAVEN-T first trains a strong teacher that models directed local interactions with a surround-aware graph, efficient temporal filtering, shifted-window spatial attention, and a sparse Mixture-of-Experts head for different maneuvers. A compact GRU–Squeeze-and-Excitation student with a low-rank adapted policy head is then aligned to the teacher at feature, attention, and semantic levels. After that, the student is refined with Proximal Policy Optimization using rewards for collision avoidance, comfort, and progress, stabilized by a complexity-aware curriculum and Elastic Weight Consolidation. On standard highway and urban motion datasets the student reaches 6.2 imes fewer parameters, 3.7 imes faster inference, and 14.6 ms latency on a Jetson AGX Orin while remaining competitively accurate and improving closed-loop safety. The claim is that task-level reinforcement after heterogeneous distillation is a practical way to make multi-agent trajectory prediction both deployable and safety-aware.

Core claim

Static imitation of a high-capacity multi-agent trajectory teacher is insufficient for real-time deployment because it does not correct safety-relevant teacher bias; MAVEN-T shows that heterogeneous distillation (feature, attention, semantic) of a surround-aware graph-plus-MoE teacher into a compact GRU–SE student, followed by PPO refinement with collision, comfort, and progress rewards under curriculum and EWC stabilization, yields a student with 6.2 imes parameter compression, 3.7 imes speed-up, and 14.6 ms edge latency while preserving competitive accuracy and improving closed-loop safety.

What carries the argument

Reinforced heterogeneous distillation: multi-level (feature/attention/semantic) transfer from a surround-aware graph + sparse MoE teacher into a compact GRU–SE student with LoRA policy head, then PPO task-level refinement with safety/comfort/progress rewards, complexity-aware curriculum, and Elastic Weight Consolidation.

Load-bearing premise

Hand-designed rewards for collision avoidance, comfort, and progress, applied after distillation, will systematically fix the teacher’s safety mistakes and still generalize across traffic datasets and real onboard planning without collapsing multimodal futures.

What would settle it

On a held-out closed-loop traffic simulator or real-vehicle log that includes dense interactions absent from the training reward design, measure whether the PPO-refined student still reduces collision rate and prediction error relative to pure static distillation; a rise in collisions or collapse of mode diversity would falsify the corrective claim.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes MAVEN-T, a two-stage framework for real-time multi-agent trajectory prediction: a high-capacity teacher (surround-aware graph encoder, linear-time temporal filtering with shifted-window spatial attention, sparse MoE maneuver decoder) is distilled into a compact GRU–Squeeze-and-Excitation student with a LoRA policy head via feature-, attention-, and semantic-level losses; the student is then refined with PPO using collision, comfort, and progress rewards, stabilized by a complexity-aware curriculum and Elastic Weight Consolidation. The abstract and introduction claim 6.2× parameter compression, 3.7× inference acceleration, and 14.6 ms latency on Jetson AGX Orin with competitive accuracy on NGSIM, HighD, MoCAD, Argoverse 2, and Waymo, plus improved closed-loop safety relative to static imitation. The central conceptual claim is that task-level PPO acts as a corrective refinement for safety-relevant teacher bias that conventional distillation leaves unaddressed (Fig. 1).

Significance. If the reported compression/latency numbers and closed-loop gains hold under proper multimodality and transfer checks, the work is practically relevant for on-board autonomous driving predictors, where high-capacity graph/attention/generative models often miss real-time budgets. The framing that heterogeneous multi-level distillation plus PPO refinement can correct teacher bias—rather than only imitate it—is a useful systems contribution relative to static KD and pure lightweight predictors. Multi-dataset evaluation spanning highway and urban interactive benchmarks, plus explicit edge-device latency, would strengthen deployability claims if fully substantiated. The free-parameter surface (distillation weights, reward weights, LoRA rank, MoE sparsity, curriculum, EWC) is large, so significance depends on whether ablations isolate PPO’s corrective role from reward hacking or mode collapse.

major comments (4)
  1. The manuscript text available for review is truncated: after the introduction it jumps to the reference list (pp. 13–14) and author bios, omitting methods, equations, loss definitions, training protocol, tables, ablations, and closed-loop evaluation details. Load-bearing claims in the abstract (6.2× compression, 3.7× speedup, 14.6 ms Orin latency, competitive multi-dataset accuracy, closed-loop safety gains from PPO) cannot be verified without those sections. A complete methods/results body is required before the central claims can be assessed.
  2. Abstract/Introduction and Fig. 1 frame PPO (collision/comfort/progress) as a corrective refinement of safety-relevant teacher bias after multi-level distillation. Trajectory prediction is multimodal; PPO optimizes expected return of a policy head. Without explicit multimodality preservation (e.g., mode-conditioned rewards, diversity regularizers, constrained updates, or reported minADE/minFDE vs. mode-coverage/NLL before and after PPO), reward refinement can collapse futures toward a single high-return trajectory that looks good in closed-loop training scenes while harming calibrated multimodality needed for planning. This is the least secured causal link in the strongest claim and needs dedicated ablations and metrics.
  3. The weakest assumption is that hand-designed PPO rewards plus EWC/curriculum systematically correct teacher bias and transfer across NGSIM/HighD/MoCAD/Argoverse 2/Waymo rather than overfit proxy rewards or closed-loop training scenes. The paper must isolate (i) student after static multi-level KD only, (ii) KD+PPO, and (iii) PPO without full heterogeneous distillation, with open-loop accuracy, mode diversity, and closed-loop collision/comfort/progress on held-out interactive urban sets. EWC and curriculum stabilize stage-wise training but do not by themselves establish bias correction over reward hacking.
  4. Free parameters listed in the pipeline (feature/attention/semantic distillation weights, PPO reward weights, LoRA rank/placement, MoE expert count and sparsity, complexity curriculum schedule, EWC strength) are numerous. Sensitivity or ablations on these knobs—especially reward weights and whether PPO is applied to the full multimodal head or a single policy mode—are load-bearing for reproducibility and for the claim that gains are not fragile hyperparameter fits.
minor comments (4)
  1. Title and running header inconsistently render the name as “MA VEN-T” / “MAVEN-T”; standardize spelling throughout.
  2. Abstract and introduction largely restate the same pipeline; once the full methods section is restored, tighten the intro to avoid duplication and point to specific equations/tables for compression and closed-loop results.
  3. Fig. 1 caption contrasts generic policy distillation with MAVEN-T; ensure the figure itself (not only the caption) makes the three-level distillation vs. action-policy transfer distinction visually clear for readers unfamiliar with KD+RL hybrids.
  4. Related-work citations in the truncated reference block mix trajectory predictors, Mamba/KD variants, and PPO; when the full related-work section is present, explicitly position against prior KD-for-trajectory and RL-for-prediction works (e.g., RAIN, proximal policy distillation) so novelty of “reinforced heterogeneous distillation” is crisp.

Circularity Check

0 steps flagged

Empirical multi-agent trajectory distillation paper; claims are measured on external public benchmarks, not forced by definition or self-citation chains.

full rationale

MAVEN-T proposes a teacher–student pipeline (surround-aware graph + temporal filtering/shifted-window attention + sparse MoE teacher; GRU–SE + LoRA student; feature/attention/semantic distillation; then PPO with collision/comfort/progress rewards, curriculum, and EWC) and reports compression, latency, accuracy, and closed-loop safety on NGSIM, HighD, MoCAD, Argoverse 2, and Waymo. These are empirical systems results evaluated against held-out public data and standard baselines, not algebraic identities or fitted constants renamed as predictions. Standard components (PPO, GAE, EWC, LoRA, SE, Swin-style windows, MoE) are cited from independent prior work. Author-overlapping self-citation (e.g., adaptive context-length MARL) appears only as related work and is not used as a uniqueness theorem or load-bearing premise that forces the main accuracy/compression claims. No self-definitional loop, no fitted-input-as-prediction reduction, and no ansatz smuggled in as a first-principles derivation. Residual risks (reward design, possible mode collapse under PPO) are correctness/generalization concerns, not circularity. Score 0 is the honest finding.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 3 invented entities

Load-bearing content is architectural and training-design choices rather than physical axioms. The central claim rests on public datasets, standard optimizers/architectures, and many free design knobs (loss weights, reward terms, ranks, curriculum). No formal derivation; invented pieces are named modules whose value is purely empirical.

free parameters (6)
  • distillation loss weights (feature / attention / semantic)
    Relative weights among multi-level KD terms are design choices that control what the student imitates; not fixed by theory.
  • PPO reward weights for collision, comfort, progress
    Hand-designed multi-objective rewards directly shape the claimed safety correction; different weights can reverse conclusions.
  • LoRA rank and adaptation placement on policy head
    Rank and which layers are adapted determine student capacity and stability; chosen by authors.
  • MoE expert count / sparsity routing in teacher
    Controls teacher multimodality capacity and thus the distillation target quality.
  • complexity-aware curriculum schedule
    Stage ordering and difficulty ramps are free training design that can affect final accuracy/safety tradeoffs.
  • EWC regularization strength
    Controls retention vs plasticity across stage-wise training; fitted/tuned rather than derived.
axioms (4)
  • domain assumption A high-capacity teacher’s intermediate features, attentions, and semantic outputs are useful supervision for a compact trajectory student.
    Standard KD assumption invoked throughout the distillation stage (abstract/intro).
  • ad hoc to paper PPO with collision/comfort/progress rewards improves downstream safety of open-loop trajectory predictions without destroying multimodal validity.
    Core corrective claim of MAVEN-T vs static imitation (Fig. 1 and abstract).
  • domain assumption Public offline trajectory datasets plus the authors’ closed-loop evaluation sufficiently represent onboard multi-agent interaction for deployment claims.
    Implicit in multi-dataset and closed-loop safety evaluation claims.
  • standard math Standard deep learning components (GAT-style graphs, shifted-window attention, MoE, GRU, SE, LoRA, PPO, EWC) behave as in prior literature.
    Background methods cited in references; used compositionally.
invented entities (3)
  • MAVEN-T reinforced heterogeneous distillation pipeline no independent evidence
    purpose: Name the full teacher–multi-level-KD–PPO–curriculum–EWC system as a deployable trajectory framework.
    The paper’s primary proposed object; value is empirical, not independently evidenced outside this work.
  • surround-aware graph encoder (teacher) no independent evidence
    purpose: Model directed local multi-agent interactions for the high-capacity teacher.
    Presented as a teacher module; no external falsifiable handle beyond this paper’s ablations (not visible in supplied text).
  • complexity-aware curriculum for stage-wise distillation/RL no independent evidence
    purpose: Stabilize progressive training from imitation to PPO refinement.
    Training construct introduced to support the pipeline; success is internal to reported runs.

pith-pipeline@v1.1.0-grok45 · 11120 in / 3344 out tokens · 38979 ms · 2026-07-12T22:50:20.476956+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of MAVEN-T: Reinforced Heterogeneous Distillation for Real-Time Multi-Agent Trajectory Prediction." pith.science (2026). https://pith.science/paper/WHTEK4GB

@misc{pith2026260410169,
  author       = {Pith},
  title        = {Pith review of: MAVEN-T: Reinforced Heterogeneous Distillation for Real-Time Multi-Agent Trajectory Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WHTEK4GB}},
  note         = {Machine review of arXiv:2604.10169}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Trajectory prediction is a key component of autonomous driving systems because future motions directly affect collision checking, behavior planning, and control. The task remains challenging under dense interactions, heterogeneous behaviors, multimodal futures, and limited on-board computation. Existing graph, attention, and generative predictors improve interaction reasoning or uncertainty modeling, but their high-capacity designs are often costly for real-time deployment. Lightweight predictors and conventional distillation reduce inference cost, yet usually rely on static imitation and do not explicitly correct safety-relevant teacher bias. This paper proposes \textbf{MAVEN-T}, a reinforced heterogeneous distillation framework for real-time multi-agent trajectory prediction. A high-capacity teacher models directed local interactions with a surround-aware graph encoder, combines efficient temporal filtering with shifted-window spatial attention, and decodes maneuver-specific futures through a sparse Mixture-of-Experts head. A compact GRU--Squeeze-and-Excitation student with a Low-Rank Adapted policy head is trained by feature-, attention-, and semantic-level distillation. To align prediction with downstream behavior, the student is further refined by Proximal Policy Optimization rewards for collision avoidance, comfort, and progress, while a complexity-aware curriculum and Elastic Weight Consolidation stabilize stage-wise training. Experiments on NGSIM, HighD, MoCAD, Argoverse~2, and the Waymo Open Motion Dataset evaluate accuracy, efficiency, generalization, robustness, and closed-loop safety. The student achieves 6.2$\times$ parameter compression, 3.7$\times$ inference acceleration, and 14.6,ms latency on an NVIDIA Jetson AGX Orin while maintaining competitive accuracy.

Figures

Figures reproduced from arXiv: 2604.10169 by Jinguo Xian, Wenchang Duan, Yi Shi, Zhenguo Gao.

Figure 1
Figure 1. Figure 1: Architecture overview of the proposed teacher–student framework for autonomous-driving policy learning. The teacher network (upper) incorporates [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Hybrid attention mechanism in the teacher model. The architec [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Lane keeping scenario [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Left lane change scenario [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Right lane change scenario. curriculum learning accelerates convergence by 37%, reducing training time from 45 to 28 epochs. F. Cross-Dataset Generalization G. Computational Complexity Analysis 1) Theoretical Complexity: Teacher Complexity: O(N 2 d + T d2 + M d3 ) (34) Student Complexity: O(T dh + dh2 ) (35) Speedup Ratio: N2d + T d2 + M d3 T dh + dh2 ≈ 3.7× (36) where N is the number of agents, T is seque… view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Chain-Aware Encoding for Microservice Trace Anomaly Detection

    cs.SE 2026-07 conditional novelty 6.0

    Encoding trace events as (endpoint, root-to-span chain) pairs lets a dual-task LSTM detect structural path anomalies online, raising F1 from 89.0% to 94.3% and path recall to 99.1% on TrainTicket.

  2. OC-Distill: Ontology-aware Contrastive Learning with Cross-Modal Distillation for ICU Risk Prediction

    cs.LG 2026-04 conditional novelty 6.0

    An ICU risk model pretrained with ICD-ontology-aware contrastive learning and fine-tuned by distilling clinical notes achieves the best vitals-only results on MIMIC tasks.

  3. OC-Distill: Ontology-aware Contrastive Learning with Cross-Modal Distillation for ICU Risk Prediction

    cs.LG 2026-04 unverdicted novelty 5.0

    OC-Distill combines ontology-aware contrastive pretraining with cross-modal distillation to improve ICU risk prediction performance and label efficiency while using only vital signs at inference.

  4. OC-Distill: Ontology-aware Contrastive Learning with Cross-Modal Distillation for ICU Risk Prediction

    cs.LG 2026-04 unverdicted novelty 4.0

    Ontology-aware contrastive pretraining plus note-to-vitals distillation improves MIMIC ICU risk and length-of-stay prediction using only vital signs at inference.

Reference graph

Works this paper leans on

43 extracted references · 1 linked inside Pith · cited by 2 Pith papers

  1. [1]

    C2F-TP: A coarse-to- fine denoising framework for uncertainty-aware trajectory prediction,

    Y . Wen, H. Tang, S. Wang, Y . Pan, and W. Chen, “C2F-TP: A coarse-to- fine denoising framework for uncertainty-aware trajectory prediction,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, 2025

    2025

  2. [2]

    S2TNet: Spatio-temporal trans- former network for multi-agent trajectory prediction,

    S. Chen, T. Zhao, P. Wang, and M. Liu, “S2TNet: Spatio-temporal trans- former network for multi-agent trajectory prediction,” inProceedings of the Asian Conference on Machine Learning, 2021, pp. 454–469

    2021

  3. [3]

    Multimodal motion prediction with stacked transformers,

    Y . Liu, J. Zhang, L. Fang, Q. Jiang, and B. Zhou, “Multimodal motion prediction with stacked transformers,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 7577–7586

    2021

  4. [4]

    Enhancing mapless trajectory prediction through knowledge distillation,

    Y . Wang, P. Zhang, L. Bai, F. Li, and J. Xue, “Enhancing mapless trajectory prediction through knowledge distillation,”IEEE Transactions on Intelligent Transportation Systems, vol. 26, no. 10, pp. 16 665–16 675, 2025

    2025

  5. [5]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” inProceedings of the Conference on Language Modeling, 2024

    2024

  6. [6]

    Trajectory mamba: Efficient attention- mamba forecasting model based on selective state space,

    Y . Huang, M. Lin, and C. Wei, “Trajectory mamba: Efficient attention- mamba forecasting model based on selective state space,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2025

    2025

  7. [7]

    KD-Mamba: Selective state space models with knowledge distillation for trajectory prediction,

    H. Park, J. Choi, and M. Lee, “KD-Mamba: Selective state space models with knowledge distillation for trajectory prediction,”Expert Systems with Applications, vol. 248, 2025

    2025

  8. [8]

    Proximal policy distillation,

    G. Spigler, “Proximal policy distillation,”Transactions on Machine Learning Research, 2025. [Online]. Available: https://openreview.net/ forum?id=WfVXe88oMh

    2025

  9. [9]

    Attention based vehicle trajectory prediction,

    K. Messaoud, I. Yahiaoui, A. Verroust-Blondet, and F. Nashashibi, “Attention based vehicle trajectory prediction,”IEEE Transactions on Intelligent Vehicles, vol. 6, no. 1, pp. 175–185, 2021

    2021

  10. [10]

    Social GAN: Socially acceptable trajectories with generative adversarial net- works,

    A. Gupta, J. Johnson, L. Fei-Fei, S. Savarese, and A. Alahi, “Social GAN: Socially acceptable trajectories with generative adversarial net- works,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 2255–2264

    2018

  11. [11]

    Trajectron++: Dynamically-feasible trajectory forecasting with heterogeneous data,

    T. Salzmann, B. Ivanovic, P. Chakravarty, and M. Pavone, “Trajectron++: Dynamically-feasible trajectory forecasting with heterogeneous data,” in Proceedings of the European Conference on Computer Vision, 2020, pp. 683–700

    2020

  12. [12]

    I2T: From intention decoupling to vehicular trajectory prediction based on prioriformer networks,

    Y . Zhou, Z. Wang, N. Ning, Z. Jin, N. Lu, and X. S. Shen, “I2T: From intention decoupling to vehicular trajectory prediction based on prioriformer networks,”IEEE Transactions on Intelligent Transportation Systems, vol. 25, no. 8, pp. 9411–9426, 2024

    2024

  13. [13]

    Intention-aware denoising diffu- sion model for trajectory prediction,

    C. Liu, S. He, H. Liu, and J. Chen, “Intention-aware denoising diffu- sion model for trajectory prediction,”IEEE Transactions on Intelligent Transportation Systems, vol. 26, no. 5, pp. 5915–5930, 2025

    2025

  14. [14]

    RAIN: Reinforced hybrid attention inference network for motion forecasting,

    J. Li, F. Yang, M. Tomizuka, and C. Choi, “RAIN: Reinforced hybrid attention inference network for motion forecasting,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 16 096–16 106

    2021

  15. [15]

    GameFormer: Game-theoretic modeling and learning of transformer-based interactive prediction and planning for autonomous driving,

    Z. Huang, X. Mo, and C. Lv, “GameFormer: Game-theoretic modeling and learning of transformer-based interactive prediction and planning for autonomous driving,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 3903–3913

    2023

  16. [16]

    Minds on the move: Decoding trajectory prediction in autono- mous driving with cognitive insights,

    H. Liao, C. Wang, K. Zhu, Y . Ren, B. Gao, S. E. Li, C. Xu, and Z. Li, “Minds on the move: Decoding trajectory prediction in autono- mous driving with cognitive insights,”IEEE Transactions on Intelligent Transportation Systems, vol. 26, no. 5, pp. 6101–6115, 2025

    2025

  17. [17]

    An interactive prediction and planning method for lane change trajectories,

    H. Liu, M. Yang, H. Wang, Y .-H. Chen, and W. Xu, “An interactive prediction and planning method for lane change trajectories,”IEEE Transactions on Intelligent Transportation Systems, vol. 26, no. 2, pp. 1842–1857, 2025

    2025

  18. [18]

    How attentive are graph attention networks?

    S. Brody, U. Alon, and E. Yahav, “How attentive are graph attention networks?” inProceedings of the International Conference on Learning Representations, 2022

    2022

  19. [19]

    Root mean square layer normalization,

    B. Zhang and R. Sennrich, “Root mean square layer normalization,” Advances in Neural Information Processing Systems, vol. 32, 2019

    2019

  20. [20]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 10 012–10 022

    2021

  21. [21]

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,

    N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hinton, and J. Dean, “Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,” inProceedings of the International Confer- ence on Learning Representations, 2017

    2017

  22. [22]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018, pp. 7132–7141

    2018

  23. [23]

    LoRA: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large language models,” inProceedings of the International Conference on Learning Represen- tations, 2022

    2022

  24. [24]

    Prox- imal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Prox- imal policy optimization algorithms,”arXiv preprint arXiv:1707.06347, 2017

    Pith/arXiv arXiv 2017

  25. [25]

    High- dimensional continuous control using generalized advantage estimation,

    J. Schulman, P. Moritz, S. Levine, M. Jordan, and P. Abbeel, “High- dimensional continuous control using generalized advantage estimation,” inProceedings of the International Conference on Learning Represen- tations, 2016

    2016

  26. [26]

    Adaptive context length optimization with low-frequency truncation for multi-agent reinforcement learning,

    W. Duan, Y . Yu, J. He, and Y . Shi, “Adaptive context length optimization with low-frequency truncation for multi-agent reinforcement learning,” inAdvances in Neural Information Processing Systems (NeurIPS), 2025

    2025

  27. [27]

    Overcoming catastrophic forgetting in neural networks,

    J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska et al., “Overcoming catastrophic forgetting in neural networks,”Pro- ceedings of the National Academy of Sciences, vol. 114, no. 13, pp. 3521–3526, 2017

    2017

  28. [28]

    Graph-based spatial-temporal convolutional network for vehicle trajectory prediction in autonomous driving,

    Z. Sheng, Y . Xu, S. Xue, and D. Li, “Graph-based spatial-temporal convolutional network for vehicle trajectory prediction in autonomous driving,”IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 10, pp. 17 654–17 665, 2022

    2022

  29. [29]

    Next generation simulation (NGSIM) vehicle trajectories and supporting data,

    U.S. Federal Highway Administration, “Next generation simulation (NGSIM) vehicle trajectories and supporting data,” U.S. Department of Transportation Data Portal,

  30. [30]

    Available: https://data.transportation.gov/Automobiles/ Next-Generation-Simulation-NGSIM-Vehicle-Trajector/8ect-6jqj

    [Online]. Available: https://data.transportation.gov/Automobiles/ Next-Generation-Simulation-NGSIM-Vehicle-Trajector/8ect-6jqj

  31. [31]

    The highD dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of highly automated driving systems,

    R. Krajewski, J. Bock, L. Kloeker, and L. Eckstein, “The highD dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of highly automated driving systems,” inProceedings of the IEEE International Conference on Intelligent Transportation Systems, 2018, pp. 2118–2125

    2018

  32. [32]

    MoCAD: Macao connected autonomous driving dataset,

    H. Liao, C. Wang, Z. Li, B. Gao, and C. Xu, “MoCAD: Macao connected autonomous driving dataset,” 2024. [Online]. Available: https://github.com/Petrichor625/HLTP

    2024

  33. [33]

    Argoverse 2: Next generation datasets for self-driving percep- tion and forecasting,

    B. Wilson, W. Qi, T. Agarwal, J. Lambert, J. Singh, S. Khandelwal, B. Pan, R. Kumar, A. Hartnett, J. K. Pontes, D. Ramanan, P. Carr, and J. Hays, “Argoverse 2: Next generation datasets for self-driving percep- tion and forecasting,” inAdvances in Neural Information Processing Systems Datasets and Benchmarks Track, 2023

    2023

  34. [34]

    Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset,

    S. Ettinger, S. Cheng, B. Caine, C. Liu, H. Zhao, S. Pradhan, Y . Chai, B. Sapp, C. R. Qi, Y . Zhou, Z. Yang, A. Chouard, P. Wang, J. Ngiam, H. Caesar, K. S. Refaat, A. Salz, and D. Anguelov, “Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset,” inProceedings of the IEEE/CVF International Conference on Compute...

    2021

  35. [35]

    Social LSTM: Human trajectory prediction in crowded spaces,

    A. Alahi, K. Goel, V . Ramanathan, A. Robicquet, L. Fei-Fei, and S. Savarese, “Social LSTM: Human trajectory prediction in crowded spaces,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 961–971

    2016

  36. [36]

    Intention- aware vehicle trajectory prediction based on spatial-temporal dynamic attention network for internet of vehicles,

    X. Chen, H. Zhang, F. Zhao, Y . Hu, C. Tan, and J. Yang, “Intention- aware vehicle trajectory prediction based on spatial-temporal dynamic attention network for internet of vehicles,”IEEE Transactions on Intel- ligent Transportation Systems, vol. 23, no. 10, pp. 19 471–19 483, 2022

    2022

  37. [37]

    WSiP: Wave superposition inspired pooling for dynamic interactions-aware trajectory prediction,

    R. Wang, S. Wang, H. Yan, and X. Wang, “WSiP: Wave superposition inspired pooling for dynamic interactions-aware trajectory prediction,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, 2023, pp. 4685–4692

    2023

  38. [38]

    GRIP++: Enhanced graph-based interaction-aware trajectory prediction for autonomous driving,

    X. Li, X. Ying, and M. C. Chuah, “GRIP++: Enhanced graph-based interaction-aware trajectory prediction for autonomous driving,” inPro- ceedings of the IEEE/CVF International Conference on Computer Vision Workshops, 2019, pp. 3515–3524

    2019

  39. [39]

    HiVT: Hierarchical vector transformer for multi-agent motion prediction,

    Z. Zhou, L. Ye, J. Wang, K. Wu, and K. Lu, “HiVT: Hierarchical vector transformer for multi-agent motion prediction,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 8823–8833

    2022

  40. [40]

    Motion transformer with global intention localization and local movement refinement,

    S. Shi, L. Jiang, D. Dai, and B. Schiele, “Motion transformer with global intention localization and local movement refinement,” inAdvances in Neural Information Processing Systems, vol. 35, 2022, pp. 6531–6543

    2022

  41. [41]

    NEST: A neuromodu- lated small-world hypergraph trajectory prediction model for autono- mous driving,

    J. Hu, Y . Wang, J. Xu, Y . Wang, and W. Jiang, “NEST: A neuromodu- lated small-world hypergraph trajectory prediction model for autono- mous driving,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, 2025. IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS 14

    2025

  42. [42]

    Less is more: Efficient brain-inspired learning for autonomous driving trajectory prediction,

    H. Liao, Y . Li, Z. Li, C. Wang, G. Li, C. Tian, Z. Bian, K. Zhu, Z. Cui, and J. Hu, “Less is more: Efficient brain-inspired learning for autonomous driving trajectory prediction,” inECAI 2024, ser. Frontiers in Artificial Intelligence and Applications. IOS Press, 2024

    2024

  43. [43]

    Wayformer: Motion forecasting via simple & efficient at- tention networks,

    N. Nayakanti, R. Al-Rfou, A. Zhou, K. Goel, K. S. Refaat, and B. Sapp, “Wayformer: Motion forecasting via simple & efficient at- tention networks,” inProceedings of the IEEE International Conference on Robotics and Automation, 2023, pp. 2592–2598. Wenchang Duan(Graduate Student Member, IEEE) received the B.S. degree in statistics from the Xi- angtan Unive...

    2023