pith. sign in

arxiv: 2605.29705 · v1 · pith:R6LJZMN2new · submitted 2026-05-28 · 💻 cs.AI

BitTP: The Lightweight Trajectory Prediction Model with BitLLM for Edge-Devices

Mincheol Kang , Hyunjin Lim , Bomin Kang , Daehee Park This is my paper

Pith reviewed 2026-06-29 07:54 UTC · model grok-4.3

classification 💻 cs.AI
keywords trajectory predictionLLM quantizationedge devicesbitlinear architectureautonomous systemsweight-only quantizationspatio-temporal reasoningmodel compression
0
0 comments X

The pith

BitTP shows that quantizing only the weights of an LLM trajectory predictor to 1.58 bits improves accuracy over the full-precision baseline while cutting memory and latency.

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

The paper develops BitTP to adapt LLM-based trajectory predictors for deployment on edge devices by converting them to a bitlinear architecture. It identifies weight-only quantization to 1.58 bits as optimal, with the result that this version reduces average displacement error by 14.29 percent and final displacement error by 20.97 percent compared with the original BF16 model. Activations are kept in full precision because quantizing them causes severe degradation in the model's ability to reason about multi-agent interactions. The approach simultaneously lowers memory footprint and inference time relative to other quantization schemes. The authors conclude that the quantization step functions as a regularizer that aids generalization in spatio-temporal tasks.

Core claim

BitTP converts an LLM-based trajectory predictor into a lightweight bitlinear architecture. Weight-only quantization to 1.58 bits (BitTP-Weight) not only preserves but improves prediction quality over the full-precision BF16 baseline, reducing ADE by 14.29 percent and FDE by 20.97 percent on average. Memory usage and inference latency decrease compared with other quantization methods. Activations must remain in full precision; quantizing them produces severe degradation and instability in spatio-temporal reasoning. The quantization thereby acts as an effective regularizer that enables practical deployment of sophisticated LLM-based reasoning on edge devices.

What carries the argument

Bitlinear architecture with weight-only 1.58-bit quantization while keeping activations in full precision.

If this is right

  • LLM-based trajectory predictors become deployable on resource-constrained edge hardware such as on-board computers in autonomous robots.
  • Quantization to low-bit weights alone can improve rather than degrade accuracy in multi-agent interaction modeling.
  • The method reduces both memory consumption and inference latency relative to alternative quantization approaches.
  • Language-based trajectory representations can be used in real-time autonomous systems without full-precision compute.

Where Pith is reading between the lines

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

  • The same weight-only scheme might transfer to other sequential prediction tasks that rely on contextual reasoning.
  • Hybrid precision could be explored for additional compression while monitoring stability in interaction modeling.
  • Real-world testing on varied edge hardware would clarify whether the observed latency gains hold under varying loads.
  • Improved accuracy from quantization raises questions about how the regularizer effect scales to larger LLMs.

Load-bearing premise

Quantizing activations produces severe degradation and instability in spatio-temporal reasoning, so they must stay in full precision.

What would settle it

An experiment in which both weights and activations are quantized to 1.58 bits and the model still shows stable or improved ADE/FDE without instability would falsify the requirement for full-precision activations.

Figures

Figures reproduced from arXiv: 2605.29705 by Bomin Kang, Daehee Park, Hyunjin Lim, Mincheol Kang.

Figure 1
Figure 1. Figure 1: Overview of the BitTP framework. (a) The base￾line trajectory prediction model [2] utilizing a full-precision (bf16,fp16,fp32) sequence-to-sequence architecture. (b) Our proposed BitTP, which replaces standard nn.Linear layers with BitLinear modules to selectively quantize model weights to 1.58-bit [-1,0,1]. and to plan safe, socially compliant actions. However, real￾world environments are highly interacti… view at source ↗
Figure 2
Figure 2. Figure 2: BitTP Model Architecture and Selective Quantization Strategies. (Left) The overall BitTP framework, based on a T5-small architecture. All standard nn.Linear layers within the Feed-Forward Networks (FFN) and Attention blocks are dynamically replaced by our custom BitLinear modules. (Right) We systematically investigate three selective quantization strategies: (a) BitLinear-Both (Eq. 3), which quantizes both… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative visualization of stochastic trajectory predictions on the ETH/UCY benchmark. ated with temperature 0.7. Quantization Settings. BitTP introduces lightweight BitLinear modules for weight-only, activation-only, and combined quantization, as described in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Training loss convergence curves in ETH (Y-axis in log-scale). (a) Comparison of quantization targets (lr = 1×10−4 ). (b) Sensitivity analysis of different learning rates for each BitTP strategy (batch size=256). termittent instability, likely due to accumulated quantiza￾tion noise propagating across layers [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
read the original abstract

Trajectory prediction is a fundamental task for autonomous systems, requiring complex reasoning about multi-agent interactions and intents. Large language models (LLMs) have recently been adopted for this task, as they provide strong contextual reasoning and interpretable, language-based trajectory representations. However, these LLM-based predictors are extremely memory- and compute-intensive, making them difficult to deploy on resource-constrained edge devices such as on-board computers in autonomous robots. To bridge this gap, we propose BitTP, which converts an LLM-based trajectory predictor into a lightweight bitlinear architecture. We demonstrate that weight-only quantization to 1.58-bit (BitTP-Weight) is optimal. Crucially, activations must remain in full precision, as quantizing them leads to severe degradation and instability in spatio-temporal reasoning. Empirically, BitTP-Weight not only preserves but improves prediction quality over the full-precision (BF16) LLM baseline, reducing ADE by 14.29% and FDE by 20.97% on average, while simultaneously reducing memory usage and inference latency relative to other quantization methods. These results demonstrate that carefully designed quantization acts as an effective regularizer, enabling the practical deployment of sophisticated LLM-based reasoning on edge devices. Code is available at: https://github.com/MintCat98/BitTP.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper proposes BitTP, a method to convert LLM-based trajectory predictors into a lightweight bitlinear architecture. It claims that weight-only quantization to 1.58 bits (BitTP-Weight) is optimal when activations remain in full precision, as quantizing activations causes severe degradation. Empirically, BitTP-Weight improves average ADE by 14.29% and FDE by 20.97% over the BF16 LLM baseline while reducing memory usage and inference latency, with quantization acting as an effective regularizer for edge-device deployment in autonomous systems.

Significance. If the central empirical claims hold after addressing baseline ambiguities, the work would be significant for enabling deployment of complex LLM-based spatio-temporal reasoning on resource-constrained edge devices, demonstrating that targeted quantization can improve rather than degrade prediction quality in trajectory forecasting tasks.

major comments (1)
  1. [Abstract] Abstract: The central claim attributes the ADE/FDE reductions (14.29%/20.97%) and regularization effect specifically to 1.58-bit weight quantization. However, the baseline is the unmodified full-precision LLM; the method first converts the model to a bitlinear architecture before applying quantization. Without an explicit ablation measuring the bitlinear architecture at BF16 precision, it is impossible to isolate whether the observed gains arise from quantization or from the architectural substitution itself.
minor comments (2)
  1. [Abstract] The abstract states that 'activations must remain in full precision' as crucial, but provides no quantitative comparison or ablation showing the degradation from activation quantization.
  2. No details are provided on the specific datasets, baselines, or error bars supporting the reported ADE/FDE improvements.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for highlighting the baseline ambiguity in isolating the contribution of 1.58-bit quantization versus the bitlinear architectural conversion. We address this point directly below and commit to revisions that strengthen the empirical claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim attributes the ADE/FDE reductions (14.29%/20.97%) and regularization effect specifically to 1.58-bit weight quantization. However, the baseline is the unmodified full-precision LLM; the method first converts the model to a bitlinear architecture before applying quantization. Without an explicit ablation measuring the bitlinear architecture at BF16 precision, it is impossible to isolate whether the observed gains arise from quantization or from the architectural substitution itself.

    Authors: We agree that the current baseline comparison does not fully disentangle the bitlinear conversion from the subsequent weight-only quantization. The bitlinear architecture was introduced specifically to enable efficient 1.58-bit operations, but an explicit BF16 bitlinear ablation is indeed required to confirm that the observed ADE/FDE improvements and regularization effect stem from quantization rather than the architectural change alone. In the revised manuscript we will add this ablation (bitlinear model at BF16 versus quantized BitTP-Weight versus original LLM), report the corresponding ADE/FDE and latency numbers, and update the abstract and experimental sections accordingly to clarify the source of the gains. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical quantization study with no derivations or self-referential reductions

full rationale

The paper presents an empirical evaluation of weight-only 1.58-bit quantization on an LLM-based trajectory predictor, reporting ADE/FDE improvements over a BF16 baseline without any mathematical derivations, equations, fitted parameters presented as predictions, or load-bearing self-citations. The central claims rest on experimental measurements rather than a derivation chain that reduces to its own inputs by construction. The noted experimental design choice (baseline is unmodified LLM rather than bitlinear BF16) is a potential validity concern but does not constitute circularity under the specified patterns, as no step equates a result to an input via definition or self-citation. The work is self-contained as a report on observed effects.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no architectural equations, training details, or parameter counts are provided, so the ledger cannot be populated beyond noting the implicit assumption that the bitlinear conversion preserves the original model's reasoning capability.

pith-pipeline@v0.9.1-grok · 5769 in / 1056 out tokens · 21695 ms · 2026-06-29T07:54:00.475666+00:00 · methodology

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

56 extracted references · 9 canonical work pages · 6 internal anchors

  1. [1]

    So- cial lstm: Human trajectory prediction in crowded spaces

    Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. So- cial lstm: Human trajectory prediction in crowded spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961–971, 2016. 1, 2

    2016

  2. [2]

    Can language beat numerical regression? language-based multimodal tra- jectory prediction

    Inhwan Bae, Junoh Lee, and Hae-Gon Jeon. Can language beat numerical regression? language-based multimodal tra- jectory prediction. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 753–766, 2024. 1, 2, 3, 5, 6, 7

    2024

  3. [3]

    Lan- guage models are few-shot learners.Advances in neural in- formation processing systems, 33:1877–1901, 2020

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Sub- biah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakan- tan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Lan- guage models are few-shot learners.Advances in neural in- formation processing systems, 33:1877–1901, 2020. 1, 3

    1901

  4. [4]

    nuscenes: A multi- modal dataset for autonomous driving

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom. nuscenes: A multi- modal dataset for autonomous driving. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621–11631, 2020. 2

    2020

  5. [5]

    NuPlan: A closed-loop ML-based planning benchmark for autonomous vehicles

    Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based plan- ning benchmark for autonomous vehicles.arXiv preprint arXiv:2106.11810, 2021. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  6. [6]

    Intentnet: Learning to predict intention from raw sensor data

    Sergio Casas, Wenjie Luo, and Raquel Urtasun. Intentnet: Learning to predict intention from raw sensor data. InCon- ference on robot learning, pages 947–956. PMLR, 2018. 2

    2018

  7. [7]

    Multipath: Multiple probabilistic anchor tra- jectory hypotheses for behavior prediction.arXiv preprint arXiv:1910.05449, 2019

    Yuning Chai, Benjamin Sapp, Mayank Bansal, and Dragomir Anguelov. Multipath: Multiple probabilistic anchor tra- jectory hypotheses for behavior prediction.arXiv preprint arXiv:1910.05449, 2019. 2

    work page arXiv 1910

  8. [8]

    Argoverse: 3d tracking and forecasting with rich maps

    Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jag- jeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8748–8757, 2019. 2

    2019

  9. [9]

    The case for 4-bit pre- cision: k-bit inference scaling laws

    Tim Dettmers and Luke Zettlemoyer. The case for 4-bit pre- cision: k-bit inference scaling laws. InInternational Confer- ence on Machine Learning, pages 7750–7774. PMLR, 2023. 6

    2023

  10. [10]

    LLM.int8(): 8-bit matrix multiplication for transformers at scale

    Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. LLM.int8(): 8-bit matrix multiplication for transformers at scale. InAdvances in Neural Information Processing Systems, 2022. 2, 3

    2022

  11. [11]

    8-bit optimizers via block-wise quantization

    Tim Dettmers, Mike Lewis, Sam Shleifer, and Luke Zettle- moyer. 8-bit optimizers via block-wise quantization. In 9th International Conference on Learning Representations, ICLR, 2022

    2022

  12. [12]

    Qlora: Efficient finetuning of quantized llms,

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms,

  13. [13]

    Bert: Pre-training of deep bidirectional trans- formers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional trans- formers for language understanding. InProceedings of the 2019 conference of the North American chapter of the asso- ciation for computational linguistics: human language tech- nologies, volume 1 (long and short papers), pages 4171– 4186, 2019. 1, 3

    2019

  14. [14]

    Large scale interactive motion forecasting for autonomous driving: The waymo open mo- tion dataset

    Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open mo- tion dataset. InProceedings of the IEEE/CVF international conference on computer vision, pages 9710–9719, 2021. 2

    2021

  15. [15]

    The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks

    Jonathan Frankle and Michael Carbin. The lottery ticket hy- pothesis: Finding sparse, trainable neural networks.arXiv preprint arXiv:1803.03635, 2018. 2, 3

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  16. [16]

    Gptq: Accurate post-training quantization for gener- ative pre-trained transformers, 2023

    Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Al- istarh. Gptq: Accurate post-training quantization for gener- ative pre-trained transformers, 2023. 2, 3

    2023

  17. [17]

    Vectornet: Encoding hd maps and agent dynamics from vectorized rep- resentation

    Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, and Cordelia Schmid. Vectornet: Encoding hd maps and agent dynamics from vectorized rep- resentation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11525– 11533, 2020. 1, 2

    2020

  18. [18]

    llama.cpp: Llm inference in c/c++.https: //github.com/ggml- org/llama.cpp

    ggml org. llama.cpp: Llm inference in c/c++.https: //github.com/ggml- org/llama.cpp. Accessed 2026-04-08. 6

    2026

  19. [19]

    Improving multi-agent motion predic- tion with heuristic goals and motion refinement

    Carlos G ´omez-Hu´elamo, Marcos V Conde, Rafael Barea, and Luis M Bergasa. Improving multi-agent motion predic- tion with heuristic goals and motion refinement. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5323–5332, 2023. 2, 3

    2023

  20. [20]

    Stochastic trajectory pre- diction via motion indeterminacy diffusion

    Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yong- ming Rao, Jie Zhou, and Jiwen Lu. Stochastic trajectory pre- diction via motion indeterminacy diffusion. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17113–17122, 2022. 5, 6

    2022

  21. [21]

    Social gan: Socially acceptable tra- jectories with generative adversarial networks

    Agrim Gupta, Justin Johnson, Li Fei-Fei, Silvio Savarese, and Alexandre Alahi. Social gan: Socially acceptable tra- jectories with generative adversarial networks. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 2255–2264, 2018. 1, 2

    2018

  22. [22]

    Distilling the Knowledge in a Neural Network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distill- ing the knowledge in a neural network.arXiv preprint arXiv:1503.02531, 2015. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  23. [23]

    One thousand and one hours: Self-driving motion prediction dataset

    John Houston, Guido Zuidhof, Luca Bergamini, Yawei Ye, Long Chen, Ashesh Jain, Sammy Omari, Vladimir Iglovikov, and Peter Ondruska. One thousand and one hours: Self-driving motion prediction dataset. InConference on Robot Learning, pages 409–418. PMLR, 2021. 2

    2021

  24. [24]

    Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022. 2, 3

    2022

  25. [25]

    Stgat: Modeling spatial-temporal interac- tions for human trajectory prediction

    Yingfan Huang, Huikun Bi, Zhaoxin Li, Tianlu Mao, and Zhaoqi Wang. Stgat: Modeling spatial-temporal interac- tions for human trajectory prediction. InProceedings of the IEEE/CVF international conference on computer vision, pages 6272–6281, 2019. 1, 2

    2019

  26. [26]

    The trajectron: Proba- bilistic multi-agent trajectory modeling with dynamic spa- tiotemporal graphs

    Boris Ivanovic and Marco Pavone. The trajectron: Proba- bilistic multi-agent trajectory modeling with dynamic spa- tiotemporal graphs. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 2375–2384,

  27. [27]

    Social- bigat: Multimodal trajectory forecasting using bicycle-gan and graph attention networks.Advances in neural informa- tion processing systems, 32, 2019

    Vineet Kosaraju, Amir Sadeghian, Roberto Mart ´ın-Mart´ın, Ian Reid, Hamid Rezatofighi, and Silvio Savarese. Social- bigat: Multimodal trajectory forecasting using bicycle-gan and graph attention networks.Advances in neural informa- tion processing systems, 32, 2019. 2

    2019

  28. [28]

    ALBERT: A Lite BERT for Self-supervised Learning of Language Representations

    Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942, 2019. 1, 3

    work page internal anchor Pith review Pith/arXiv arXiv 1909

  29. [29]

    Traj-llm: A new explo- ration for empowering trajectory prediction with pre-trained large language models, 2024

    Zhengxing Lan, Hongbo Li, Lingshan Liu, Bo Fan, Yisheng Lv, Yilong Ren, and Zhiyong Cui. Traj-llm: A new explo- ration for empowering trajectory prediction with pre-trained large language models, 2024. 1, 3

    2024

  30. [30]

    Desire: Distant future prediction in dynamic scenes with interacting agents

    Namhoon Lee, Wongun Choi, Paul Vernaza, Christopher B Choy, Philip HS Torr, and Manmohan Chandraker. Desire: Distant future prediction in dynamic scenes with interacting agents. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 336–345, 2017. 1, 2

    2017

  31. [31]

    Spatio-temporal graph dual-attention network for multi-agent prediction and tracking, 2021

    Jiachen Li, Hengbo Ma, Zhihao Zhang, Jinning Li, and Masayoshi Tomizuka. Spatio-temporal graph dual-attention network for multi-agent prediction and tracking, 2021. 2

    2021

  32. [32]

    Evolvegraph: Multi-agent trajectory prediction with dynam- ically evolving interaction graphs

    Xin Li, Tianchen Yu, Noriaki Kim, Wei Xu, and Renjie Liu. Evolvegraph: Multi-agent trajectory prediction with dynam- ically evolving interaction graphs. InAdvances in Neural Information Processing Systems (NeurIPS), pages 19742– 19752, 2020. 2, 3

    2020

  33. [33]

    Learning lane graph representa- tions for motion forecasting

    Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, and Raquel Urtasun. Learning lane graph representa- tions for motion forecasting. InEuropean Conference on Computer Vision, pages 541–556. Springer, 2020. 1, 2

    2020

  34. [34]

    Awq: Activation-aware weight quantization for llm compression and acceleration, 2024

    Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. Awq: Activation-aware weight quantization for llm compression and acceleration, 2024. 2, 3

    2024

  35. [35]

    It is not the journey but the destination: Endpoint conditioned trajectory prediction, 2020

    Karttikeya Mangalam, Harshayu Girase, Shreyas Agarwal, Kuan-Hui Lee, Ehsan Adeli, Jitendra Malik, and Adrien Gaidon. It is not the journey but the destination: Endpoint conditioned trajectory prediction, 2020. 2, 5, 6

    2020

  36. [36]

    From goals, waypoints & paths to long term hu- man trajectory forecasting

    Karttikeya Mangalam, Yang An, Harshayu Girase, and Jiten- dra Malik. From goals, waypoints & paths to long term hu- man trajectory forecasting. InProceedings of the IEEE/CVF international conference on computer vision, pages 15233– 15242, 2021. 3

    2021

  37. [37]

    Social-stgcnn: A social spatio-temporal graph convolutional neural network for human trajectory prediction

    Abduallah Mohamed, Kun Qian, Mohamed Elhoseiny, and Christian Claudel. Social-stgcnn: A social spatio-temporal graph convolutional neural network for human trajectory prediction. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14424– 14432, 2020. 1, 2

    2020

  38. [38]

    Scene transformer: A unified architecture for predicting multiple agent trajectories, 2022

    Jiquan Ngiam, Benjamin Caine, Vijay Vasudevan, Zheng- dong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, David Weiss, Ben Sapp, Zhifeng Chen, and Jonathon Shlens. Scene transformer: A unified architecture for predicting multiple agent trajectories, 2022. 1, 2, 3

    2022

  39. [39]

    Exploring the limits of transfer learning with a unified text-to-text transformer.Journal of machine learning research, 21(140):1–67, 2020

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer.Journal of machine learning research, 21(140):1–67, 2020. 3

    2020

  40. [40]

    Xnor-net: Imagenet classification using bi- nary convolutional neural networks

    Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using bi- nary convolutional neural networks. InEuropean conference on computer vision, pages 525–542. Springer, 2016. 2, 3

    2016

  41. [41]

    Learning social etiquette: Human trajectory prediction in crowded scenes

    A Robicquet, A Sadeghian, A Alahi, and S Savarese. Learning social etiquette: Human trajectory prediction in crowded scenes. InEuropean Conference on Computer Vi- sion (ECCV), page 5, 2017. 2

    2017

  42. [42]

    Trajectron++: Dynamically-feasible trajec- tory forecasting with heterogeneous data

    Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectron++: Dynamically-feasible trajec- tory forecasting with heterogeneous data. InEuropean Con- ference on Computer Vision, pages 683–700. Springer, 2020. 1, 2

    2020

  43. [43]

    DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter

    Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter.arXiv preprint arXiv:1910.01108,

    work page internal anchor Pith review Pith/arXiv arXiv 1910

  44. [44]

    Drivelm: Driving with graph visual question answering

    Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beißwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. InEuropean conference on computer vision, pages 256–274. Springer, 2024. 1

    2024

  45. [45]

    LLaMA: Open and Efficient Foundation Language Models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth´ee Lacroix, Baptiste Rozi`ere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  46. [46]

    Spectra 1.1: Scaling laws and efficient inference for ternary language models.arXiv preprint arXiv:2506.23025, 2025

    Tejas Vaidhya, Ayush Kaushal, Vineet Jain, Francis Cou- ture Harpin, Prashant Shishodia, Majid Behbahani, Yuriy Nevmyvaka, and Irina Rish. Spectra 1.1: Scaling laws and efficient inference for ternary language models.arXiv preprint arXiv:2506.23025, 2025. 6

    work page arXiv 2025

  47. [47]

    Bitnet: 1-bit pre-training for large language models.Journal of Machine Learning Research, 26(125):1–29, 2025

    Hongyu Wang, Shuming Ma, Lingxiao Ma, Lei Wang, Wen- hui Wang, Li Dong, Shaohan Huang, Huaijie Wang, Jilong Xue, Ruiping Wang, et al. Bitnet: 1-bit pre-training for large language models.Journal of Machine Learning Research, 26(125):1–29, 2025. 2, 3, 4, 5, 7

    2025

  48. [48]

    So- cialvae: Human trajectory prediction using timewise latents

    Pei Xu, Jean-Bernard Hayet, and Ioannis Karamouzas. So- cialvae: Human trajectory prediction using timewise latents. InEuropean Conference on Computer Vision, pages 511–

  49. [49]

    Springer, 2022. 5, 6

    2022

  50. [50]

    Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting

    Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris M Kitani. Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting. InProceedings of the IEEE/CVF international conference on computer vision, pages 9813– 9823, 2021. 1, 2, 3

    2021

  51. [51]

    Tnt: Target-driven trajectory pre- diction

    Hang Zhao, Jiyang Gao, Tian Lan, Chen Sun, Ben Sapp, Balakrishnan Varadarajan, Yue Shen, Yi Shen, Yuning Chai, Cordelia Schmid, et al. Tnt: Target-driven trajectory pre- diction. InConference on robot learning, pages 895–904. PMLR, 2021. 1, 2

    2021

  52. [52]

    Uni4d-llm: A unified spatiotemporal-aware vlm for 4d understanding and gener- ation.arXiv preprint arXiv:2509.23828, 2025

    Hanyu Zhou and Gim Hee Lee. Uni4d-llm: A unified spatiotemporal-aware vlm for 4d understanding and gener- ation.arXiv preprint arXiv:2509.23828, 2025. 1, 3

    work page arXiv 2025

  53. [53]

    Hivt: Hierarchical vector transformer for multi-agent motion prediction

    Zikang Zhou, Luyao Ye, Jianping Wang, Kui Wu, and Ke- jie Lu. Hivt: Hierarchical vector transformer for multi-agent motion prediction. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 8823–8833, 2022. 2

    2022

  54. [54]

    A unified frame- work for multimodal, multi-part human motion synthesis,

    Zixiang Zhou, Yu Wan, and Baoyuan Wang. A unified frame- work for multimodal, multi-part human motion synthesis,

  55. [55]

    Query-centric trajectory prediction

    Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17863–17873, 2023. 1, 2 BitTP: The Lightweight Trajectory Prediction Model with BitLLM for Edge-Devices Supplementary Material S1. Implementation Details and Evaluation M...

    2023

  56. [56]

    framework, BitTP-Weight successfully achieves a throughput of 2.18 sequences/s on the ETH subset. This result confirms that running sophisticated trajectory predic- tion models entirely on resource-constrained on-board com- puters without GPU acceleration is practically possible. To measure realistic edge-device latency, our CPU evaluation uses single-sam...