Pith. sign in

REVIEW 4 major objections 6 minor 64 references

The paper tests five LLM modules for driving motion generation and finds that, with domain-specific adaptations, they push the Waymo realism score to 0.778.

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 · deepseek-v4-flash

2026-08-05 11:25 UTC pith:KF6GTESG

load-bearing objection Useful systematic ablation of LLM modules for driving motion generation, but the headline claim is only as solid as the Waymo realism metric, which the authors themselves show is partly biased against safer policies. the 4 major comments →

arxiv 2509.02754 v1 pith:KF6GTESG submitted 2025-09-02 cs.AI

Do LLM Modules Generalize? A Study on Motion Generation for Autonomous Driving

Mingyi Wang , Jingke Wang , Tengju Ye , Junbo Chen , Kaicheng Yu This is my paper
classification cs.AI
keywords large language modelsmotion generationautonomous drivingWaymo Sim Agentstokenizer designpositional embeddingGRPOtest-time computing
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.

The paper asks whether the modules that powered large language models—tokenizers, positional embeddings, pretraining, post-training, and test-time computation—still work when transplanted into autonomous-driving motion generation. It argues that most do, but only after adapting them to traffic: motion is continuous, agents share a spatial scene, and safety can conflict with human-likeness. Each module is ablated alone on the Waymo Sim Agents benchmark, and the winners are combined: Verlet-Agent tokenization, Global-DRoPE positional encoding, a 5.3M-parameter autoregressive model, GRPO post-training, and search-plus-cluster selection. That pipeline reaches a 0.778 realism score on the test leaderboard, competitive with published methods. The paper also records where transfer fails—vanilla 1D positional encodings and unconstrained reinforcement learning—and flags a benchmark bias in which the likelihood-based realism metric can penalize safer policies in roughly 6% of scenarios.

Core claim

The paper's central claim is that LLM modules transfer to autonomous-driving motion generation non-uniformly, and the right adaptations unlock their value. With all other components fixed, an agent-centric Verlet tokenizer with 169 acceleration tokens beats both a scene-centric Verlet tokenizer and a data-driven 2,048-token vocabulary, because the same physical action always maps to the same token. Keeping scene tokens in the global self-driving-car coordinate frame while injecting relative pose through rotary attention (Global-DRoPE) beats vanilla sinusoidal positional encoding, no positional encoding, and local-frame DRoPE, because local frames make lane tokens semantically indistinguishab

What carries the argument

The load-bearing machinery is a controlled module-by-module ablation inside one fixed architecture: a scene encoder plus a GPT-style autoregressive motion-token decoder. Three adaptations carry most of the argument. First, the Verlet-Agent tokenizer encodes each trajectory in the moving agent's own coordinate frame, then discretizes the X and Y dimensions into 13 levels each, yielding a 169-token vocabulary with a consistent token-to-physical-action mapping. Second, Global-DRoPE keeps all scene and map tokens in the global coordinate system so lane elements retain distinct semantics, while using rotary relative-position embeddings during attention. Third, the GRPO post-training objective nor

Load-bearing premise

The load-bearing premise is that the Waymo Sim Agents realism score is the right yardstick for whether a module 'generalizes'; the paper itself shows this likelihood-based metric can reward collisions and penalize safer policies when the ground-truth human trajectory hits a road boundary, in roughly 6% of scenarios.

What would settle it

Re-score the same ablations after excluding the roughly 6% of Sim Agents scenarios where the ground-truth trajectory collides with a road boundary (for example scenario dc9b2c377e20b2be) and recomputing the off-road and collision likelihood components; if the final combination's 0.778 score, or GRPO's advantage over REINFORCE and A2C, reverses, the conclusion that these modules transfer generally fails.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Data-driven trajectory tokenizers with large vocabularies transfer poorly; small-vocabulary model-based Verlet tokenizers achieve higher next-token accuracy and better closed-loop behavior.
  • Vanilla 1D positional embeddings from NLP hurt driving tasks; relative-pose rotary encodings help, and encoding scene tokens in the global coordinate frame is the strongest of the four variants tested.
  • Autoregressive motion generation exhibits scaling-law-like gains when training data is sufficient, but without enough diverse data larger models overfit, so model size should grow alongside data.
  • GRPO is the only post-training method tested that improves realism while reducing collisions and off-road rate; REINFORCE and A2C buy safety at a substantial realism cost.
  • Test-time search over many parallel rollouts is the largest single contributor to the final safety gains, cutting the off-road rate from 0.153 to 0.052 in the local evaluation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: if the 2025 Sim Agents scoring corrects the likelihood bias created by false-positive ground-truth collisions, the relative ranking of GRPO and search-based methods could shift, since both are currently penalized for being safer than a flawed ground truth.
  • Beyond the paper: the recipe of agent-centric discrete tokenization plus global-coordinate relative position encodings is a promising design principle for other multi-agent sequence generation problems, such as pedestrian prediction, robot fleet coordination, or token-based traffic world models.
  • Beyond the paper: a falsifiable extension is to re-run the same five module ablations under a cost-weighted safety metric instead of likelihood realism; the claim that these modules generalize would be much stronger if the ranking survives that replacement.

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 / 6 minor

Summary. The paper presents a systematic, module-by-module study of whether five LLM design components—tokenizer, positional embedding, pretraining/scaling, post-training, and test-time computation—transfer to multi-agent motion generation for autonomous driving. Using a GPT-style autoregressive model trained on the Waymo Open Motion Dataset, the authors propose Verlet-Agent tokenization and Global-DRoPE, compare several post-training methods and test-time selection strategies, and finally combine the best variants into a model that scores 0.778 realism on the Waymo Sim Agents test leaderboard. The central claim is that, with domain-specific adaptation, LLM modules can significantly improve autonomous-driving motion generation.

Significance. If the empirical claims are supported, this would be a useful systematic reference for the community: it identifies which LLM modules transfer, proposes two concrete adaptations (Verlet-Agent and Global-DRoPE), and ships code and a project page. The paper is also commendably transparent, especially in Section 4.1, where it acknowledges that the benchmark's likelihood-based metrics can penalize safer policies. However, the central quantitative claims are currently stronger than the evidence. The headline realism score is a weighted likelihood metric that the authors themselves show is biased against the safety improvements produced by their post-training and test-time modules, and all ablation comparisons are single point estimates without uncertainty. These issues need to be addressed before the claimed significance is established.

major comments (4)
  1. [§4.1, Tables 7, 4–6] The realism meta-metric is the yardstick for the headline conclusions, but the paper itself shows the metric systematically penalizes safer outputs. In Table 7, scenario dc9b2c377e20b2be, Pre-train (6.7% offroad) scores offroad likelihood 0.771 and realism 0.700, while GRPO and Search (0.5% offroad) score 0.176 and ≈0.55; the authors estimate ~6% of validation scenarios are affected. Since GRPO's realism gain over baseline is only 0.728 vs 0.724 (Table 4) and even the search gain in Table 5 is affected by the same weighting, the observed ranking may reflect matching erroneous ground-truth collisions instead of better driving. The abstract's claim of 'significantly improve performance for autonomous driving motion generation' is not established without re-computing realism excluding/flagging false-positive boundary-collision scenarios or showing that module rankings are stable under such
  2. [Tables 1–5] All ablation tables report single point estimates on one locally sampled validation set, with no seeds, confidence intervals, or significance tests. Several central comparisons are numerically small: GRPO vs baseline realism is 0.728 vs 0.724; SFT is 0.726 vs 0.724 (Table 4); tokenizer ADE differences are roughly 0.07–0.2 (Table 1); positional-embedding ADE differences are around 0.02–0.05 (Table 2). Without repeated runs, these differences are within plausible run-to-run noise, so 'significantly improve' and 'best overall performance' are unsupported. Please report at least three seeds with mean±std, or paired significance tests, for the module-level decisions that drive the final configuration.
  3. [§3.3, Table 3, Fig. 9] The scaling-law claim is asserted qualitatively: no power-law function is fitted, no exponents or goodness-of-fit are reported, and the 'Aug data' axis is an eightfold augmented version of the same data, not an independent increase in data diversity. Table 3 also shows saturation and overfitting for the Large model, and Fig. 9(b) shows a slight deviation from 100% to 800%. The statement that 'performance follows a power-law–like scaling during pretraining' should either be backed by an explicit fit on independent data or downgraded to a monotone-improvement observation within the studied range.
  4. [§3.5, Table 5] The test-time computing comparison confounds the selection strategy with the number of rollouts. Baseline uses 32 rollouts (runtime 0.69 s), while the proposed variants use 'MoreRollouts' (1024 in Table 8; runtimes 5.2–11.3 s). No condition with 1024 rollouts but without clustering/search is reported, so the improvements in realism (0.732–0.759 vs 0.724) cannot be attributed to clustering/search rather than to the larger sample size. Please include the MoreRollouts-only control or re-frame the contribution accordingly.
minor comments (6)
  1. [Table 3] The caption says 'Scaling Law on GPT Layers' but the rows are model-parameter scales; rename to 'Model Parameters' or clarify what is varied.
  2. [Introduction, contribution list] There is a duplicated word: 'to to motion generation in autonomous driving.'
  3. [Tables 4 and 5] Both tables list the same 'Baseline' realism value 0.724. Please state explicitly that this is the same local-validation baseline and give the number of scenarios/rollouts used, so cross-table comparisons are meaningful.
  4. [§4.1, Table 7] Please specify how the ~6% estimate was obtained (manual review subset, inter-rater agreement, etc.) and, if possible, report the realism score for all methods on the validation set after excluding those scenarios.
  5. [Appendix §5.2.6, Eq. (9)] The notation 'Â_{i,t} = er_i = ...' appears to contain a typo; define r_i and the normalization step explicitly.
  6. [Algorithms 2 and 3] The variable K is used for different purposes: in Algorithm 2 it controls the number of output rollouts, while in Algorithm 3 it is the number of agents. Rename to avoid ambiguity.

Circularity Check

0 steps flagged

No significant circularity: empirical ablation study with an external benchmark; the single self-citation (LTP scene encoder) is not load-bearing.

full rationale

This paper is an empirical study rather than a derivation. Each LLM module (tokenizer, positional embedding, pre-training scale, post-training method, test-time computation) is ablated independently while other components are held fixed, and results are reported on the Waymo Sim Agents benchmark. The final combined configuration is assembled from the ablation winners and then submitted to the separate test leaderboard, so the headline realism score of 0.778 is not the same quantity as any fitted parameter or training objective. No equation in the paper reduces to its own inputs: Verlet-Agent tokenization is compared against external baselines, Global-DRoPE is an architectural variant of externally cited DRoPE, GRPO is taken from external prior work, and the search-based selection is an inference-time heuristic. The only self-citation is the scene encoder 'inspired by the scene encoding module in LTP [15]' (Appendix 5.2.2), which is a baseline component shared across all comparisons and is not load-bearing for the paper's transferability conclusions; it also involves no uniqueness claim. The Section 4.1 discussion of the Sim Agents realism metric penalizing safer policies is a candid external-validity limitation, not a circularity in the derivation, because the paper explicitly reports the issue and the metric is still an external benchmark rather than a quantity defined by the paper's own model. Therefore, no specific circular step can be exhibited, and the appropriate score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

This is an empirical benchmark study, so the ledger contains evaluation and modeling assumptions rather than mathematical axioms or physical postulates. The free parameters are training and inference hyperparameters that are hand-set or chosen by validation. No new physical entities are introduced.

free parameters (5)
  • Tokenizer hyperparameters = 128 bins per coordinate, range [-18m, 18m], 169 token vocabulary
    These hand-chosen values define the motion vocabulary and directly affect the tokenizer comparison in Table 1. The range is justified by reference to MotionLM, not measured on this paper's data.
  • Model sizes = 0.8M, 3.7M, 5.3M, 11.6M parameters
    The scaling-law analysis in Table 3 uses four manually selected sizes. No power-law exponents are fitted or reported, so the scaling claim is descriptive rather than predictive.
  • GRPO coefficients = lambda_KL = 0.8, lambda_H = 0.01, gamma = 0.5
    These post-training weights are chosen by hand in Appendix 5.2.6 and 5.2.8 and are central to the claim that GRPO balances safety and realism.
  • Test-time rollouts and cluster count = 1024 rollouts, temperature 1.0, K-Medoids centers
    The test-time computing results in Table 5 depend on a 1024-rollout budget and the choice of K-Medoids clustering, with the exact number of clusters left underspecified in Algorithm 2.
  • Data augmentation factor = 8x via K = 8 anchor times
    The pretraining scaling study uses a synthetic 800% dataset created by eight anchors. The authors note this augmented data is redundant and deviates from ideal scaling, so it functions as a tuned setting rather than new data.
axioms (5)
  • domain assumption Waymo Sim Agents realism metrics are a valid proxy for motion generation quality.
    The benchmark score is the main evidence for every module comparison and the final leaderboard claim (Section 3 and Appendix 5.2.1).
  • domain assumption One-at-a-time ablations isolate the contribution of each LLM module.
    The paper fixes a baseline and changes only the module under study in Tables 1, 2, 4, and 5, implicitly assuming no strong interaction effects between modules.
  • domain assumption The locally sampled validation set is representative of the official leaderboard.
    All ablation tables except Section 3.6 report results on a locally sampled validation set, with no error bars or comparison to official leaderboard scores for those configurations.
  • domain assumption The 8x anchor-time augmentation produces training signal comparable to more diverse data.
    The scaling study in Section 5.2.5 treats augmented tokens as an 800% dataset, while acknowledging redundancy and weaker scaling behavior at that setting.
  • domain assumption The scene encoder and motion generator design choices do not confound module comparisons.
    The baseline removes M2M self-attention for speed and uses a scene-centric coordinate frame, so the findings may be specific to this architecture (Section 5.2.2).

pith-pipeline@v1.4.0-alltime-deepseek-medium · 20892 in / 10307 out tokens · 115602 ms · 2026-08-05T11:25:05.618896+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Do LLM Modules Generalize? A Study on Motion Generation for Autonomous Driving." pith.science (2026). https://pith.science/paper/KF6GTESG

@misc{pith2026250902754,
  author       = {Pith},
  title        = {Pith review of: Do LLM Modules Generalize? A Study on Motion Generation for Autonomous Driving},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KF6GTESG}},
  note         = {Machine review of arXiv:2509.02754}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Recent breakthroughs in large language models (LLMs) have not only advanced natural language processing but also inspired their application in domains with structurally similar problems--most notably, autonomous driving motion generation. Both domains involve autoregressive sequence modeling, token-based representations, and context-aware decision making, making the transfer of LLM components a natural and increasingly common practice. However, despite promising early attempts, a systematic understanding of which LLM modules are truly transferable remains lacking. In this paper, we present a comprehensive evaluation of five key LLM modules--tokenizer design, positional embedding, pre-training paradigms, post-training strategies, and test-time computation--within the context of motion generation for autonomous driving. Through extensive experiments on the Waymo Sim Agents benchmark, we demonstrate that, when appropriately adapted, these modules can significantly improve performance for autonomous driving motion generation. In addition, we identify which techniques can be effectively transferred, analyze the potential reasons for the failure of others, and discuss the specific adaptations needed for autonomous driving scenarios. We evaluate our method on the Sim Agents task and achieve competitive results.

Figures

Figures reproduced from arXiv: 2509.02754 by Jingke Wang, Junbo Chen, Kaicheng Yu, Mingyi Wang, Tengju Ye.

Figure 1
Figure 1. Figure 1: The autonomous driving motion generation task bears a striking resemblance to large [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Schematic diagram of different PEs. When using a local coordinate system, the lane [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of different post-training methods, showing the distribution of agents’ end [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: A limitation of using likelihood-based evaluation. Due to observation noise, the ground [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The limitation of our method. Acknowledgement This work is partially supported by the National Natural Science Foundation of China (NSFC) under grant No. 62403389 and the Provincial Natural Science Foundation of Zhejiang under grant No. QKWL25F0301. This work was carried out during the first author’s internship at Udeer.AI, in collaboration with Westlake University. The authors gratefully acknowledge the v… view at source ↗
Figure 6
Figure 6. Figure 6: The architecture of the scene encoder. Motion Generator Similar to MotionLM, we adopt an autoregressive generation scheme for multi￾agent motion generation. However, MotionLM encodes the environment separately for each agent in their respective local coordinate systems. This design, while intuitive, suffers from the curse of dimensionality and thus primarily focuses on motion generation for only two agents… view at source ↗
Figure 7
Figure 7. Figure 7: The architecture of the motion generator. [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Influence of initial velocity on the travel distance induced by the same motion token [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Scaling Law of Data Volume and Model Parameters [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: In our reinforcement learning framework, each transition from [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visualization of lane segment features under DRoPE using inputs from local (top row) [PITH_FULL_IMAGE:figures/full_fig_p025_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Under the DRoPE setting, we compare motion generation results using local (top row) [PITH_FULL_IMAGE:figures/full_fig_p025_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: The results of three parallel rollouts after pre-training. [PITH_FULL_IMAGE:figures/full_fig_p026_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: A qualitative comparison between GRPO and the pre-trained model. In (a), Vehicle 2 [PITH_FULL_IMAGE:figures/full_fig_p026_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative comparison of test-time computing methods: Cluster vs. Search vs. Clus [PITH_FULL_IMAGE:figures/full_fig_p027_15.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

64 extracted references · 32 canonical work pages · 2 internal anchors

  1. [1]

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y . Li, Y . Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

    Pith/arXiv arXiv 2024

  2. [2]

    Touvron, T

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi `ere, N. Goyal, E. Hambro, F. Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023

    Pith/arXiv arXiv 2023

  3. [3]

    J. Gao, C. Sun, H. Zhao, Y . Shen, D. Anguelov, C. Li, and C. Schmid. Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 11525–11533, 2020

    work page 2020

  4. [4]

    Z. Zhou, Z. Wen, J. Wang, Y .-H. Li, and Y .-K. Huang. Qcnext: A next-generation framework for joint multi-agent trajectory prediction. arXiv preprint arXiv:2306.10508, 2023

    Pith/arXiv arXiv 2023

  5. [5]

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

    work page 2022

  6. [6]

    Zhang, A

    Z. Zhang, A. Liniger, D. Dai, F. Yu, and L. Van Gool. Trafficbots: Towards world models for autonomous driving simulation and motion prediction. In2023 IEEE International Conference on Robotics and Automation (ICRA) , pages 1522–1529. IEEE, 2023

    work page 2023

  7. [7]

    W. Wu, X. Feng, Z. Gao, and Y . Kan. Smart: scalable multi-agent real-time motion generation via next-token prediction. Advances in Neural Information Processing Systems , 37:114048– 114071, 2024

    work page 2024

  8. [8]

    J. Zhao, J. Zhuang, Q. Zhou, T. Ban, Z. Xu, H. Zhou, J. Wang, G. Wang, Z. Li, and B. Li. Kigras: Kinematic-driven generative model for realistic agent simulation. IEEE Robotics and Automation Letters, 2024

    work page 2024

  9. [9]

    Cheng, Y

    J. Cheng, Y . Chen, and Q. Chen. Pluto: Pushing the limit of imitation learning-based planning for autonomous driving. arXiv preprint arXiv:2404.14327, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    Dauner, M

    D. Dauner, M. Hallgarten, A. Geiger, and K. Chitta. Parting with misconceptions about learning-based vehicle motion planning. In Conference on Robot Learning, pages 1268–1281. PMLR, 2023

    work page 2023

  11. [11]

    CarPlanner: Consistent Auto-regressive Trajectory Planning for Large-scale Reinforcement Learning in Autonomous Driving

    D. Zhang, J. Liang, K. Guo, S. Lu, Q. Wang, R. Xiong, Z. Miao, and Y . Wang. Carplan- ner: Consistent auto-regressive trajectory planning for large-scale reinforcement learning in autonomous driving. arXiv preprint arXiv:2502.19908, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  12. [12]

    A. Seff, B. Cera, D. Chen, M. Ng, A. Zhou, N. Nayakanti, K. S. Refaat, R. Al-Rfou, and B. Sapp. Motionlm: Multi-agent motion forecasting as language modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 8579–8590, 2023

    work page 2023

  13. [13]

    Philion, X

    J. Philion, X. B. Peng, and S. Fidler. Trajeglish: Traffic modeling as next-token prediction. arXiv preprint arXiv:2312.04535, 2023

    Pith/arXiv arXiv 2023

  14. [14]

    Z. Zhou, H. Haibo, X. Chen, J. Wang, N. Guan, K. Wu, Y .-H. Li, Y .-K. Huang, and C. J. Xue. Behaviorgpt: Smart agent simulation for autonomous driving with next-patch prediction. Advances in Neural Information Processing Systems , 37:79597–79617, 2024

    work page 2024

  15. [15]

    J. Wang, T. Ye, Z. Gu, and J. Chen. Ltp: Lane-based trajectory prediction for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, pages 17134–17142, 2022. 9

    work page 2022

  16. [16]

    H. Min, X. Xiong, P. Wang, and Z. Zhang. A hierarchical LSTM–based vehicle trajectory prediction method considering interaction information.Automotive Innovation, 7:71–81, 2024. doi:10.1007/s42154-023-00261-0

    work page doi:10.1007/s42154-023-00261-0 2024

  17. [17]

    Z. Peng, J. Yan, H. Yin, Y . Wen, W. Ge, T. Watzel, and G. Rigoll. Efficient interaction-aware trajectory prediction model based on multi-head attention. Automotive Innovation, 7(2):258– 270, 2024. doi:10.1007/s42154-023-00269-6

    work page doi:10.1007/s42154-023-00269-6 2024

  18. [18]

    Lucente, R

    G. Lucente, R. Dariani, J. Schindler, and M. Ortgiese. A bayesian approach with prior mixed strategy nash equilibrium for vehicle intention prediction. Automotive Innovation, 6:425–437,

    work page

  19. [19]

    Cusumano-Towner, D

    M. Cusumano-Towner, D. Hafner, A. Hertzberg, B. Huval, A. Petrenko, E. Vinitsky, E. Wi- jmans, T. Killian, S. Bowers, O. Sener, P. Kr ¨ahenb¨uhl, and V . Koltun. Robust autonomy emerges from self-play. 2025. doi:10.48550/arXiv.2502.03349. URL https://arxiv.org/ abs/2502.03349

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.03349 2025

  20. [20]

    Y . Fan, Y . Li, and S. Wang. Risk-aware self-consistent imitation learning for trajectory plan- ning in autonomous driving. 2024. URL https://www.ecva.net/papers/eccv_2024/ papers_ECCV/papers/02087.pdf

    work page 2024

  21. [21]

    Montali, J

    N. Montali, J. Lambert, P. Mougin, A. Kuefler, N. Rhinehart, M. Li, C. Gulino, T. Emrich, Z. Yang, S. Whiteson, et al. The waymo open sim agents challenge. Advances in Neural Information Processing Systems, 36:59151–59171, 2023

    work page 2023

  22. [22]

    Bengio, R

    Y . Bengio, R. Ducharme, P. Vincent, and C. Jauvin. A neural probabilistic language model. Journal of machine learning research, 3(Feb):1137–1155, 2003

    work page 2003

  23. [23]

    Kudo and J

    T. Kudo and J. Richardson. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. arXiv preprint arXiv:1808.06226, 2018

    Pith/arXiv arXiv 2018

  24. [24]

    Y . Gao, N. I. Nikolov, Y . Hu, and R. H. Hahnloser. Character-level translation with self- attention. arXiv preprint arXiv:2004.14788, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2004

  25. [25]

    Zhang, P

    L. Zhang, P. Li, S. Liu, and S. Shen. Simpl: A simple and efficient multi-agent motion predic- tion baseline for autonomous driving. IEEE Robotics and Automation Letters , 2024

    work page 2024

  26. [26]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polo- sukhin. Attention is all you need. Advances in neural information processing systems , 30, 2017

    2017

  27. [27]

    Huang, J

    Y . Huang, J. Du, Z. Yang, Z. Zhou, L. Zhang, and H. Chen. A survey on trajectory-prediction methods for autonomous driving. IEEE Transactions on Intelligent V ehicles , 7(3):652–674, 2022

    work page 2022

  28. [28]

    Z. Zhou, J. Wang, Y .-H. Li, and Y .-K. Huang. Query-centric trajectory prediction. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 17863–17873, 2023

    work page 2023

  29. [29]

    J. Zhao, T. Ban, Z. Liu, H. Zhou, X. Wang, Q. Zhou, H. Qin, M. Yang, L. Liu, and B. Li. Drope: Directional rotary position embedding for efficient agent interaction modeling. arXiv preprint arXiv:2503.15029, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    J. Su, M. Ahmed, Y . Lu, S. Pan, W. Bo, and Y . Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024

    2024

  31. [31]

    Kaplan, S

    J. Kaplan, S. McCandlish, T. Henighan, T. B. Brown, B. Chess, R. Child, S. Gray, A. Rad- ford, J. Wu, and D. Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 10

    Pith/arXiv arXiv 2001

  32. [32]

    J. Wei, Y . Tay, R. Bommasani, C. Raffel, B. Zoph, S. Borgeaud, D. Yogatama, M. Bosma, D. Zhou, D. Metzler, et al. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682, 2022

    Pith/arXiv arXiv 2022

  33. [33]

    Hoffmann, S

    J. Hoffmann, S. Borgeaud, A. Mensch, E. Buchatskaya, T. Cai, E. Rutherford, D. d. L. Casas, L. A. Hendricks, J. Welbl, A. Clark, et al. Training compute-optimal large language models. arXiv preprint arXiv:2203.15556, 2022

    Pith/arXiv arXiv 2022

  34. [34]

    P. F. Christiano, J. Leike, T. Brown, M. Martic, S. Legg, and D. Amodei. Deep reinforcement learning from human preferences. Advances in neural information processing systems , 30, 2017

    2017

  35. [35]

    D. M. Ziegler, N. Stiennon, J. Wu, T. B. Brown, A. Radford, D. Amodei, P. Christiano, and G. Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019

    Pith/arXiv arXiv 1909

  36. [36]

    Ouyang, J

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems , 35:27730–27744, 2022

    work page 2022

  37. [37]

    Y . Lu, J. Fu, G. Tucker, X. Pan, E. Bronstein, R. Roelofs, B. Sapp, B. White, A. Faust, S. White- son, et al. Imitation is not enough: Robustifying imitation with reinforcement learning for chal- lenging driving scenarios. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7553–7560. IEEE, 2023

    work page 2023

  38. [38]

    Z. Peng, W. Luo, Y . Lu, T. Shen, C. Gulino, A. Seff, and J. Fu. Improving agent behaviors with rl fine-tuning for autonomous driving. In European Conference on Computer Vision , pages 165–181. Springer, 2024

    work page 2024

  39. [39]

    Zhang, P

    Z. Zhang, P. Karkus, M. Igl, W. Ding, Y . Chen, B. Ivanovic, and M. Pavone. Closed-loop supervised fine-tuning of tokenized traffic models. arXiv preprint arXiv:2412.05334, 2024

    Pith/arXiv arXiv 2024

  40. [40]

    L. Rowe, R. Girgis, A. Gosselin, B. Carrez, F. Golemo, F. Heide, L. Paull, and C. Pal. Ctrl-sim: Reactive and controllable driving agents with offline reinforcement learning. arXiv preprint arXiv:2403.19918, 2024

    Pith/arXiv arXiv 2024

  41. [41]

    Cusumano-Towner, D

    M. Cusumano-Towner, D. Hafner, A. Hertzberg, B. Huval, A. Petrenko, E. Vinitsky, E. Wij- mans, T. Killian, S. Bowers, O. Sener, et al. Robust autonomy emerges from self-play. arXiv preprint arXiv:2502.03349, 2025

    Pith/arXiv arXiv 2025

  42. [42]

    Zhang, L

    S. Zhang, L. Dong, X. Li, S. Zhang, X. Sun, S. Wang, J. Li, R. Hu, T. Zhang, F. Wu, et al. Instruction tuning for large language models: A survey. arXiv preprint arXiv:2308.10792 , 2023

    arXiv 2023

  43. [43]

    R. J. Williams. Simple statistical gradient-following algorithms for connectionist reinforce- ment learning. Machine learning, 8:229–256, 1992

    work page 1992

  44. [44]

    V . Mnih, A. P. Badia, M. Mirza, A. Graves, T. Lillicrap, T. Harley, D. Silver, and K. Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International conference on machine learning, pages 1928–1937. PmLR, 2016

    work page 1928

  45. [45]

    Zarrieß, H

    S. Zarrieß, H. V oigt, and S. Sch¨uz. Decoding methods in neural language generation: a survey. Information, 12(9):355, 2021

    work page 2021

  46. [46]

    Radford, J

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019

    2019

  47. [47]

    W. X. Zhao, K. Zhou, J. Li, T. Tang, X. Wang, Y . Hou, Y . Min, B. Zhang, J. Zhang, Z. Dong, Y . Du, C. Yang, Y . Chen, Z. Chen, J. Jiang, R. Ren, Y . Li, X. Tang, Z. Liu, P. Liu, J.-Y . Nie, and J.-R. Wen. A survey of large language models. Mar 2023. URL http://arxiv.org/ abs/2303.18223v16. 11

    Pith/arXiv arXiv 2023

  48. [48]

    A. Fan, M. Lewis, and Y . Dauphin. Hierarchical neural story generation. arXiv preprint arXiv:1805.04833, 2018

    Pith/arXiv arXiv 2018

  49. [49]

    Holtzman, J

    A. Holtzman, J. Buys, L. Du, M. Forbes, and Y . Choi. The curious case of neural text degen- eration. arXiv preprint arXiv:1904.09751, 2019

    Pith/arXiv arXiv 1904

  50. [50]

    Zhang, S

    Y . Zhang, S. Sun, M. Galley, Y .-C. Chen, C. Brockett, X. Gao, J. Gao, J. Liu, and B. Dolan. Dialogpt: Large-scale generative pre-training for conversational response generation. arXiv preprint arXiv:1911.00536, 2019

    Pith/arXiv arXiv 1911

  51. [51]

    Y . Hu, S. Chai, Z. Yang, J. Qian, K. Li, W. Shao, H. Zhang, W. Xu, and Q. Liu. Solving motion planning tasks with a scalable generative model. InEuropean Conference on Computer Vision, pages 386–404. Springer, 2024

    work page 2024

  52. [52]

    L. Lin, X. Lin, K. Xu, H. Lu, L. Huang, R. Xiong, and Y . Wang. Revisit mixture models for multi-agent simulation: Experimental study within a unified framework. arXiv preprint arXiv:2501.17015, 2025

    Pith/arXiv arXiv 2025

  53. [53]

    L. Xue, A. Barua, N. Constant, R. Al-Rfou, S. Narang, M. Kale, A. Roberts, and C. Raffel. Byt5: Towards a token-free future with pre-trained byte-to-byte models. Transactions of the Association for Computational Linguistics , 10:291–306, 2022

    work page 2022

  54. [54]

    Y . Tay, V . Q. Tran, S. Ruder, J. Gupta, H. W. Chung, D. Bahri, Z. Qin, S. Baumgartner, C. Yu, and D. Metzler. Charformer: Fast character transformers via gradient-based subword tokeniza- tion. arXiv preprint arXiv:2106.12672, 2021

    Pith/arXiv arXiv 2021

  55. [55]

    Z. Dai, Z. Yang, Y . Yang, J. Carbonell, Q. V . Le, and R. Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860 , 2019

    Pith/arXiv arXiv 1901

  56. [56]

    P. He, X. Liu, J. Gao, and W. Chen. Deberta: Decoding-enhanced bert with disentangled attention. arXiv preprint arXiv:2006.03654, 2020

    Pith/arXiv arXiv 2006

  57. [57]

    Chang, H

    H. Chang, H. Zhang, L. Jiang, C. Liu, and W. T. Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11315–11325, 2022

    work page 2022

  58. [58]

    Strudel, C

    R. Strudel, C. Tallec, F. Altch ´e, Y . Du, Y . Ganin, A. Mensch, W. Grathwohl, N. Savinov, S. Dieleman, L. Sifre, et al. Self-conditioned embedding diffusion for text generation. arXiv preprint arXiv:2211.04236, 2022

    Pith/arXiv arXiv 2022

  59. [59]

    Rafailov, A

    R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn. Direct pref- erence optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728–53741, 2023

    work page 2023

  60. [60]

    T. Liu, Y . Zhao, R. Joshi, M. Khalman, M. Saleh, P. J. Liu, and J. Liu. Statistical rejection sampling improves preference optimization. arXiv preprint arXiv:2309.06657, 2023

    Pith/arXiv arXiv 2023

  61. [61]

    Q. Yu, Z. Zhang, R. Zhu, Y . Yuan, X. Zuo, Y . Yue, T. Fan, G. Liu, L. Liu, X. Liu, et al. Dapo: An open-source llm reinforcement learning system at scale. arXiv preprint arXiv:2503.14476, 2025

    Pith/arXiv arXiv 2025

  62. [62]

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V . Le, D. Zhou, et al. Chain-of- thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  63. [63]

    human-like

    C. Gulino, J. Fu, W. Luo, G. Tucker, E. Bronstein, Y . Lu, J. Harb, X. Pan, Y . Wang, X. Chen, et al. Waymax: An accelerated, data-driven simulator for large-scale autonomous driving research. Advances in Neural Information Processing Systems , 36:7730–7742, 2023. 12 Limitations 4.1 Limitations of Sim Agents Benchmark Overall, the Waymo Sim Agents leaderb...

    work page 2023

  64. [2023]

    doi:10.1007/s42154-023-00229-0

    work page doi:10.1007/s42154-023-00229-0