Pith. sign in

REVIEW 5 major objections 6 minor 3 cited by

M2BeamLLM: Multimodal Sensing-empowered mmWave Beam Prediction with Large Language Models

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read M2BeamLLM claims that fusing camera, radar, LiDAR, and GPS through attention and a mostly frozen GPT-2 predicts a V2I mmWave beam with 68.9% Top-1 accuracy on DeepSense 6G Scenario 32, 13.9 points above the next-best baseline.

desk verdict A credible multimodal beam-prediction engine with useful ablations, but the evaluation's loose ends — no error bars, a mislabeled few-shot setting, and an untested GPS memorization risk — keep the headline margin from being proven. read the letter →

arxiv 2506.14532 v1 pith:VTGEABWJ submitted 2025-06-17 cs.CL

Can Zheng , Jiguang He , Chung G. Kang , Guofa Cai , Zitong Yu , Merouane Debbah This is my paper · ORCID
classification cs.CL
keywords beampredictionmmWavecommunicationslargelanguagemodelsmultimodalfusionDeepSense6Gvehicle-to-infrastructuremanagementGPT-2
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that beam selection in a vehicle-to-infrastructure mmWave link can be predicted ahead of time by treating multimodal sensor streams as a language-model input rather than by training a purpose-built network from scratch. It claims that fusing camera, radar, LiDAR, and GPS features through contrastive alignment and attention, then running the fused sequence through a mostly frozen GPT-2 fine-tuned on only its last layers, reaches 68.9% Top-1 accuracy on DeepSense 6G Scenario 32, beating the next-best baseline by 13.9 percentage points and remaining the top method in few-shot settings. If true, the payoff is practical: base stations with standard sensors could anticipate the right beam without repeated beam sweeps, saving the time slots that beam training consumes.

What carries the argument

The load-bearing object is a pipeline that compresses four heterogeneous sensor streams into 64-dimensional embeddings, aligns them on the unit sphere by maximizing pairwise cosine similarity, fuses them with multi-head self-attention over the modality dimension, and then treats the fused 8-step history as a token sequence for a mostly frozen GPT-2 whose final two layers are unfrozen for supervised fine-tuning. This is what carries the argument: every claimed gain is attributed to the combination of shared embedding, attention-based fusion, and the pretrained decoder's temporal reasoning, while the contrastive loss regularizes the fusion against modality-specific noise.

What would settle it

Re-train M2BeamLLM on a split that places each physical route and weather condition wholly in one set, and also train it with modality streams time-shifted against the beam labels; the fusion-and-LLM claim is confirmed only if the 13.9-point lead survives the stricter split and collapses under the misalignment check.

Watch

Extended reading notes

Core claim

The central claim is that beam index prediction in a V2I mmWave system can be reformulated as a sequence modeling task over fused sensor embeddings. Each modality—camera, radar, LiDAR, and GPS—is encoded into a shared 64-dimensional feature space, aligned by a contrastive cosine-similarity objective, and fused by multi-head self-attention over the modality dimension. The fused historical sequence is then fed to a distilled GPT-2 backbone, with most layers frozen and the last two layers fine-tuned, and the output is projected to per-beam scores. On DeepSense 6G Scenario 32, the paper reports a Top-1 accuracy of 68.9%, outperforming a BERT-based variant (55.0%) and several time-series baselines, while also reporting that accuracy improves as modalities are added: vision-plus-radar reaches 51.1%, adding LiDAR adds 10.7 percentage points, and adding GPS adds another 7.1 percentage points.

Load-bearing premise

The reported gains rest on the assumption that the dataset's 70/30 split keeps training and test drive events independent and that the sensor streams and beam labels are synchronized and correctly labeled; if similar routes or weather appear in both splits, or labels are noisy, the accuracy gap could reflect the data rather than the model.

Editorial extensions

If this is right

  • If the 68.9% Top-1 result holds, a V2I base station could skip most beam-training sweeps and still pick the correct beam roughly two times out of three, reserving sweeps for uncertain cases.
  • The reported modality-ablation ladder (vision+radar 51.1%, plus LiDAR 61.8%, plus GPS 68.9%) implies that each added sensor buys a meaningful accuracy gain, so sensor-rich roadside units would be worth deploying.
  • The few-shot results (H=3, T=10) imply the method can adapt to a new intersection with very little labeled data because the pretrained backbone supplies a temporal prior.
  • The authors' finding that unfreezing 6–8 layers is a cost-effective setting, with Top-3 accuracy around 90%, means resource-constrained deployments can trade training cost against accuracy without changing the architecture.
  • Since the paper reports comparable or faster inference than lighter baselines despite a larger parameter count, real-time deployment is claimed to be feasible within the 100 ms synchronization interval of the dataset.

Reading between the lines

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

  • A control the paper leaves implicit is a multimodal transformer trained from scratch without a pretrained language backbone; running it would show what fraction of the 13.9-point gap comes from GPT-2's pretraining rather than from the fusion design.
  • Because the paper itself notes that weather changes received power despite nearly identical GPS locations, a natural extension is to test whether modality-diversity gains persist across weather and lighting conditions rather than being driven by a single dominant sensor.
  • The conclusion suggests pretraining a wireless-specific foundation model on unlabeled drive data before fine-tuning; adding a self-supervised stage over the sensor encoders would test whether the label-efficiency gains can be pushed even further.
  • A testable extension is to evaluate the same architecture on other DeepSense 6G scenarios with different obstructions and route patterns to see whether the reported accuracy gap generalizes beyond Scenario 32.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes M2BeamLLM, a multimodal beam prediction framework for V2I mmWave systems that fuses camera, radar, LiDAR, and GPS data via per-modality encoders, contrastive alignment, and multi-head attention, followed by a GPT-2 backbone with partial fine-tuning. On DeepSense 6G Scenario 32, the authors report Top-1 accuracy of 68.9% for standard prediction with H=8, T=5, outperforming the BERT-based variant and traditional RNN/Transformer/linear baselines, and claim consistent gains as modalities are added, plus strong performance in a 'few-shot' H=3, T=10 setting. The paper includes modality and layer-unfreezing ablations and a complexity analysis.

Significance. If the central result survives scrutiny, the paper provides a useful empirical demonstration that pretrained language models can serve as backbones for multimodal physical-layer prediction, and that fusing diverse sensing modalities improves beam prediction on a real-world dataset. The comparison against several generic baselines is a strength, and the modality ablation (though incomplete) suggests an intuitive monotonic trend. However, the significance is currently limited by unresolved methodological issues: potential train/test overlap in the event-level split, absence of a GPS-only baseline, an apparently vacuous alignment loss, and an internal inconsistency between the headline configuration and the best configuration reported in the layer-unfreezing ablation. These issues must be addressed before the reported 13.9 percentage-point improvement can be taken as reliable evidence for the framework's claims.

major comments (5)
  1. [IV.E.2 and IV.C] The layer-unfreezing ablation reports Top-1 accuracy of 72.4% with 4 unfrozen layers and 85.7% with 12 unfrozen layers (Section IV.E.2), whereas the headline result in Section IV.C is 68.9% with the default of 2 unfrozen layers (Table I). Since the ablation appears to use the same standard prediction task, the main comparison does not reflect the model's best achievable performance, and the claimed margin over the next-best baseline may be understated or, conversely, the chosen configuration may be arbitrary. The authors should specify which configuration is used in the comparison and why, and should report the best configuration as the primary result if they recommend full fine-tuning.
  2. [IV.A.1, IV.B, IV.E.1] The event-level 70/30 split does not guarantee that test events are spatially disjoint from training events; the paper's own visualization (Fig. 6) shows nearly identical UE locations across samples with different weather (samples 540 and 163). Since beam index is strongly coupled to UE position (as the paper notes in Fig. 9), a GPS encoder could memorize a location-to-beam mapping, making the contributions of radar/LiDAR/image fusion and the LLM backbone incidental. The modality ablation in Fig. 14 reports only three single-modality configurations and omits a GPS-only baseline, so the memorization hypothesis is not tested. A GPS-only experiment and/or a route- or location-based split is needed to validate the central claim.
  3. [III.F, Eq. (21)] The multimodal alignment loss in Eq. (21) is not well-defined: S_{ω1,ω2}[t] is a scalar, so σSoftmax(S/α) = 1 and L2 ≡ 0; moreover, multiplying by the one-hot beam label p_m[t] is dimensionally inconsistent. This means the alignment component, which is presented as a key part of the framework, is either vacuous or incorrectly transcribed. Please provide a correct contrastive loss (e.g., an InfoNCE-style loss over modality pairs) and verify that it is non-zero and that it indeed affects training.
  4. [IV.D and Section I] The 'few-shot' setting (H=3, T=10) is not a few-shot regime in the usual sense of reduced labeled training data; it is a shorter-history/longer-horizon prediction task. The introduction and contribution 3 claim that SFT enables adaptation 'with minimal training data,' but no experiment varies the amount of training data. The terminology and claims should be revised to match what is actually evaluated.
  5. [IV.C, IV.E] All results are reported as point estimates from a single training run; no random seeds, confidence intervals, or statistical tests are provided. Given the dataset size (2,138 events after event-level split) and the sensitivity of fine-tuning, the reported margins (e.g., 68.9% vs 55.0%) may not be stable. Please report mean and standard deviation over at least 3-5 seeds for the main results and ablations.
minor comments (6)
  1. [Table II] The header 'interference cost' should read 'inference cost'.
  2. [Eq. (25)] The summation index 'n−1' appears to be a typo; it should be 'n=1' to 'NTest'.
  3. [Section III, Notations] The symbol M is overloaded: it denotes both the codebook size (Eq. 2) and the feature embedding dimension (Section III.A). Please use distinct symbols to avoid ambiguity.
  4. [Section IV.A.1] The paper does not report the number of beams M (codebook size), which is needed to interpret the Top-1 accuracy values in context.
  5. [Section IV.E.1] The text says 'three single-modality' configurations but only explicitly names vision-only and radar-only in the comparison; please list all three explicitly, and also note that GPS-only is absent.
  6. [Abstract and Conclusion] The abstract and conclusion claim 'robustness,' but the experimental design does not include controlled robustness tests such as sensor dropout or weather variation; this claim should be softened or supported.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: the 68.9% accuracy claim is an empirical comparison on the external DeepSense 6G benchmark; the only self-citation (BeamLLM [10]) is contextual and not load-bearing.

full rationale

The paper's central claim is an empirical result — 68.9% Top-1 accuracy on DeepSense 6G Scenario 32 against GRU/LSTM/NLinear/Informer/BERT baselines — and no analytic result is derived from an assumption that already contains the conclusion. The ground-truth beam index enters through supervised losses L1 and L2 (Eqs. 20–21), but that is ordinary training signal evaluated on held-out data, not a definitional identification of prediction with fit. The only self-citation, BeamLLM [10], is used to motivate the work ('its reliance on a single sensing modality ... limits its adaptability') and is not invoked to justify the reported gains or the architecture. The paper does contain experimental-design weaknesses that could inflate the support for the claim: the event-level 70/30 split rule is unspecified, GPS min-max normalization in Eq. (10) uses dataset-wide statistics, and the modality ablation omits a GPS-only configuration, so the 7.1% gain attributed to adding GPS is not controlled against a GPS-memorization hypothesis. These are data-leakage/control concerns, not circularity: they do not make any equation or fitted parameter identical to the claimed prediction by construction. Accordingly, no circular step can be quoted and exhibited per the hard rules.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central empirical result rests on dataset assumptions (label reliability, split independence) and on the transferability of pretrained LLM representations to fused sensor embeddings. There are no invented physical entities. The main free choices are the contrastive temperature, the alignment loss weight, the fine-tuning depth, and the absence of a validation split.

free parameters (5)
  • alpha (contrastive temperature) = 0.07
    Temperature in the multimodal alignment loss (Eq. 21); chosen by hand without a validation sweep, affects softness of similarity distribution.
  • lambda (alignment loss weight) = 1
    Weight in composite loss L = L1 + lambda*L2 (Eq. 22); chosen ad hoc, no sensitivity analysis.
  • number of unfrozen LLM layers = 2 (default)
    SFT depth; ablation shows 12 unfrozen layers gives 85.7% Top-1, so the default is not performance-optimal and selection appears test-based.
  • feature embedding dimension M = 64 per modality (Fig. 3); codebook size unspecified
    Architecture dimension chosen by hand; must match codebook size for output projection.
  • train/test split ratio = 70/30
    No validation set is described; hyperparameters seem selected on the test split, inflating reported accuracy.
assumptions (4)
  • domain assumption DeepSense 6G Scenario 32 ground-truth beam labels and sensor synchronization are reliable.
    Section IV.A.1 uses the dataset as ground truth without verification of label noise or sync errors.
  • domain assumption Event-level 70/30 split isolates train and test without temporal or environmental leakage.
    Section IV.A.1; no analysis of whether weather, route, or time overlaps between splits.
  • ad hoc to paper Pretrained GPT-2 representations transfer to non-linguistic fused sensor embeddings.
    Section III.E; the paper assumes 'knowledge activation' explains gains, but no random-initialization baseline is provided.
  • standard math The optimal beam index is determined by maximum channel gain (Eq. 2), and the dataset label reflects this.
    Section II.A; standard assumption, not independently verified for the dataset.

how reviews work

0 comments
Cite this review

Pith. "Pith review of M2BeamLLM: Multimodal Sensing-empowered mmWave Beam Prediction with Large Language Models." pith.science (2026). https://pith.science/paper/VTGEABWJ

@misc{pith2026250614532,
  author       = {Pith},
  title        = {Pith review of: M2BeamLLM: Multimodal Sensing-empowered mmWave Beam Prediction with Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VTGEABWJ}},
  note         = {Machine review of arXiv:2506.14532}
}
read the original abstract

This paper introduces a novel neural network framework called M2BeamLLM for beam prediction in millimeter-wave (mmWave) massive multi-input multi-output (mMIMO) communication systems. M2BeamLLM integrates multi-modal sensor data, including images, radar, LiDAR, and GPS, leveraging the powerful reasoning capabilities of large language models (LLMs) such as GPT-2 for beam prediction. By combining sensing data encoding, multimodal alignment and fusion, and supervised fine-tuning (SFT), M2BeamLLM achieves significantly higher beam prediction accuracy and robustness, demonstrably outperforming traditional deep learning (DL) models in both standard and few-shot scenarios. Furthermore, its prediction performance consistently improves with increased diversity in sensing modalities. Our study provides an efficient and intelligent beam prediction solution for vehicle-to-infrastructure (V2I) mmWave communication systems.

Figures

Figures reproduced from arXiv: 2506.14532 by the authors.

Figure 1
Figure 1. Illustration of the V2I system model: The BS is equipped with a camera, radar, and LiDAR, while a GPS-RTK system provides the UE’s localization information to the BS. II. SYSTEM MODEL AND PROBLEM FORMULATION A. System Model As shown in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The model framework of M2BeamLLM. In the sensing data encoding part, it encodes and aligns multimodal sensing data from camera, radar, LiDAR, and GPS. It then fuses these diverse inputs using a multi-head attention mechanism. Subsequently, the fused data undergoes input projection before being fed into an LLM backbone, which includes both frozen and unfrozen pre-trained components. Finally, an output projection and … view at source ↗
Figure 3
Figure 3. Multimodal data encoding module for the M [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Optimal beam index frequency distribution. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Comparison of power profiles for the first [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 8
Figure 8. Figure 8: Visualization of LiDAR point cloud data of sample [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: BS and UE GPS location map for the first [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 7
Figure 7. Figure 7: Radar data visualization of sample 86. by their distinct positions in the angle-distance spectrum and their strong concentration along the 0 velocity line in the RV map. Furthermore, the radar system is also capable of identifying distinct moving targets at specific di…
Figure 10
Figure 10. Figure 10: Average Top-K accuracy performance of the proposed method comparing to several baselines in the standard predic￾tion task. in GPT-2) for future beam index prediction, compared to the bidirectional encoder architecture used in BERT. Traditional RNN-based models show re…
Figure 12
Figure 12. Figure 12: Average Top-K accuracy performance of the proposed method comparing to several baselines in the few-shot predic￾tion task. models. However, the Top-2 and Top-3 accuracy metrics remain largely unaffected. E. Comparative Analysis and Component Ablation Study In this sec…
Figure 15
Figure 15. Figure 15: Average Top-K accuracy performance of different combinations of sensing data [PITH_FULL_IMAGE:figures/full_fig_p010_15.png]
Figure 16
Figure 16. Figure 16: Average DBA-score performance of different combi [PITH_FULL_IMAGE:figures/full_fig_p010_16.png]
Figure 19
Figure 19. Figure 19: Comparison of average Top-K accuracy for different numbers of frozen layers [PITH_FULL_IMAGE:figures/full_fig_p011_19.png]
Figure 20
Figure 20. Figure 20: Comparison of average DBA-score accuracy for [PITH_FULL_IMAGE:figures/full_fig_p011_20.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

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

  1. Topological sum rule for geometric phases of quantum gates

    quant-ph 2026-03 unverdicted novelty 6.0 of 10

    Geometric phases of a two-qubit gate over a complete basis sum to a multiple of the Hamiltonian winding number, so topology is necessary for entanglement generation.

  2. M3F-UAV: A Missing-Modality Multimodal Foundation Model for Low-Altitude Wireless Sensing

    eess.SP 2026-07 conditional novelty 5.0 of 10

    A missing-modality multimodal foundation model fuses RGB, depth, LiDAR, and CSI to support UAV localization, beam prediction, and CSI prediction with graceful degradation when a sensor is missing.

  3. Data-Free Knowledge Distillation for LiDAR-Aided Beam Tracking in MmWave Systems

    eess.SP 2025-09 conditional novelty 4.0 of 10

    A data-free knowledge distillation method trains a compact student for LiDAR-aided mmWave beam tracking entirely on synthetic data generated from a teacher's statistics, nearly matching teacher accuracy.

Reference graph

Works this paper leans on

29 extracted references · 22 canonical work pages · cited by 3 Pith papers

  1. [1]

    Mi llimeter wave communications for future mobile networks,

    M. Xiao, S. Mumtaz, Y . Huang, L. Dai, Y . Li, M. Matthaiou, G . K. Karagiannidis, E. Bjornson, K. Y ang, C.-L. I, and et al., “Mi llimeter wave communications for future mobile networks,” IEEE J. Sel. Areas Commun., vol. 35, no. 9, p. 1909–1935, Sep 2017

    work page 1909

  2. [2]

    A s urvey of beam management for mmwave and thz communications toward s 6g,

    Q. Xue, C. Ji, S. Ma, J. Guo, Y . Xu, Q. Chen, and W. Zhang, “A s urvey of beam management for mmwave and thz communications toward s 6g,” IEEE Commun. Surveys Tuts. , vol. 26, no. 3, pp. 1520–1559, 2024

    work page 2024

  3. [3]

    Deep learning for mmwave beam and blockage prediction using sub-6 ghz channels,

    M. Alrabeiah and A. Alkhateeb, “Deep learning for mmwave beam and blockage prediction using sub-6 ghz channels,” IEEE Trans. Commun. , vol. 68, no. 9, pp. 5504–5518, 2020

    work page 2020

  4. [4]

    Computer vision aided beam tr acking in A real-world millimeter wave deployment,

    S. Jiang and A. Alkhateeb, “Computer vision aided beam tr acking in A real-world millimeter wave deployment,” in Proc. IEEE Globecom W orkshops (GC Wkshps), 2022, pp. 142–147

    2022

  5. [5]

    Radar aided 6G beam predic tion: Deep learning algorithms and real-world demonstration,

    U. Demirhan and A. Alkhateeb, “Radar aided 6G beam predic tion: Deep learning algorithms and real-world demonstration,” i n Proc. IEEE Wireless Communications and Networking Conference (WCNC) , 2022, pp. 2655–2660

    work page 2022

  6. [6]

    LiDAR aided futur e beam prediction in real-world millimeter wave V2I communicatio ns,

    S. Jiang, G. Charan, and A. Alkhateeb, “LiDAR aided futur e beam prediction in real-world millimeter wave V2I communicatio ns,” IEEE Wireless Commun. Lett. , vol. 12, no. 2, pp. 212–216, 2023

    work page 2023

  7. [7]

    P osition-aided beam prediction in the real world: How useful GPS locations a ctually are?

    J. Morais, A. Bchboodi, H. Pezeshki, and A. Alkhateeb, “P osition-aided beam prediction in the real world: How useful GPS locations a ctually are?” in Proc. IEEE International Conference on Communications (ICC), 2023, pp. 1824–1829

    work page 2023

  8. [8]

    GPT-4 technical report,

    OpenAI, “GPT-4 technical report,” Accessed Mar. 2023. [ Online]. Avail- able: https://arxiv.org/abs/2303.08774

    arXiv 2023

Show all 29 references
  1. [9]

    Deepseek-r1: Incentivizing reasoning capability in LLMs via reinforcement learning,

    D. Guo, D. Y ang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S . Ma, P . Wang, X. Bi et al. , “Deepseek-r1: Incentivizing reasoning capability in LLMs via reinforcement learning,” arXiv preprint arXiv:2501.12948 , 2025

    2025 arXiv

  2. [10]

    BeamLLM: Vi sion- empowered mmWave beam prediction with large language model s,

    C. Zheng, J. He, G. Cai, Z. Y u, and C. G. Kang, “BeamLLM: Vi sion- empowered mmWave beam prediction with large language model s,” arXiv preprint arXiv:2503.10432 , 2025

    2025 arXiv

  3. [11]

    Large l anguage models empower multimodal integrated sensing and communic ation,

    L. Cheng, H. Zhang, B. Di, D. Niyato, and L. Song, “Large l anguage models empower multimodal integrated sensing and communic ation,” IEEE Commun. Mag. , vol. 63, no. 5, pp. 190–197, 2025

    2025

  4. [12]

    Multimodal d eep learning empowered millimeter-wave beam prediction,

    B. Shi, M. Li, M.-M. Zhao, M. Lei, and L. Li, “Multimodal d eep learning empowered millimeter-wave beam prediction,” in Proc. IEEE 99th V ehicular Technology Conference (VTC2024-Spring), 2024, pp. 1– 6

    2024

  5. [13]

    Multimodal deep learning-empowered beam prediction in fu ture THz ISAC systems,

    K. Zhang, W. Y u, H. He, S. Song, J. Zhang, and K. B. Letaief , “Multimodal deep learning-empowered beam prediction in fu ture THz ISAC systems,” arXiv preprint arXiv:2505.02381 , 2025

    2025 arXiv

  6. [14]

    Se nsing- assisted high reliable communication: A transformer-base d beamforming Approach,

    Y . Cui, J. Nie, X. Cao, T. Y u, J. Zou, J. Mu, and X. Jing, “Se nsing- assisted high reliable communication: A transformer-base d beamforming Approach,” IEEE J. Sel. Topics Signal Process. , vol. 18, no. 5, pp. 782– 795, 2024

    2024

  7. [15]

    LIMA: Less is more for alignment,

    C. Zhou, P . Liu, P . Xu, S. Iyer, J. Sun, Y . Mao, X. Ma, A. Efr at, P . Y u, L. Y u, S. Zhang, G. Ghosh, M. Lewis, L. Zettlemoyer, and O. Lev y, “LIMA: Less is more for alignment,” in Proc. Neural Information Processing Systems (NIPS) , 2023

    2023

  8. [16]

    ImageNet: A large-scale fierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fe i, “ImageNet: A large-scale fierarchical image database,” in Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , Jun. 2009, pp. 248– 255

    2009

  9. [17]

    Deep residual learni ng for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learni ng for image recognition,” in Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , Jun. 2016, pp. 770–778

    2016

  10. [18]

    High accuracy range estimation of FMCW lev el radar based on the phase of the zero-padded FFT,

    Q. Guoqing, “High accuracy range estimation of FMCW lev el radar based on the phase of the zero-padded FFT,” in Proc. 7th International Conference on Signal Processing (ICSP) , vol. 3, 2004, pp. 2078–2081

    2004

  11. [19]

    Gradient-based learning applied to document recognition,

    Y . Lecun, L. Bottou, Y . Bengio, and P . Haffner, “Gradient-based learning applied to document recognition,” Proc. IEEE, vol. 86, no. 11, pp. 2278– 2324, 1998

    1998

  12. [20]

    Multimodal alignment and fusion: A su rvey,

    S. Li and H. Tang, “Multimodal alignment and fusion: A su rvey,” arXiv preprint arXiv:2411.17040, 2024

    2024

  13. [21]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Ag arwal, G. Sastry, A. Askell, P . Mishkin, J. Clark et al. , “Learning transferable visual models from natural language supervision,” in Proc. International Conference on Machine Learning (ICML) , Jul. 2021, pp. 8748–8763

    2021

  14. [22]

    DeepSense 6G: A large-scale r eal-world multi-modal sensing and communication dataset,

    A. Alkhateeb, G. Charan, T. Osman, A. Hredzak, J. Morais , U. Demirhan, and N. Srinivas, “DeepSense 6G: A large-scale r eal-world multi-modal sensing and communication dataset,” IEEE Commun. Mag. , vol. 61, no. 9, pp. 122–128, 2023

    2023

  15. [23]

    Language models are unsupervised multitask learn- ers,

    A. Radford et al. , “Language models are unsupervised multitask learn- ers,” OpenAI blog , vol. 1, no. 8, p. 9, Feb. 2019

    2019

  16. [24]

    On the properties of neural machine translation: encoder-decode r approaches,

    K. Cho, B. V an Merri¨ enboer, D. Bahdanau, and Y . Bengio, “On the properties of neural machine translation: encoder-decode r approaches,” arXiv preprint arXiv:1409.1259 , 2014

    2014 arXiv

  17. [25]

    Long short-term mem ory,

    S. Hochreiter and J. Schmidhuber, “Long short-term mem ory,” Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997

    1997

  18. [26]

    Are transformers e ffective for time series forecasting?

    A. Zeng, M. Chen, L. Zhang, and Q. Xu, “Are transformers e ffective for time series forecasting?” 2023

    2023

  19. [27]

    Informer: Beyond efficient transformer for long sequence t ime-series Forecasting,

    H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W . Zhang, “Informer: Beyond efficient transformer for long sequence t ime-series Forecasting,” in Proc. AAAI Conference on Artificial Intelligence , vol. 35, May. 2021, pp. 11 106–11 115. 13

    2021

  20. [28]

    Adam: A method for stochastic opt imization,

    D. P . Kingma and J. Ba, “Adam: A method for stochastic opt imization,” in Proc. International Conference on Learning Representatio ns (ICLR) , May. 2015

    2015

  21. [29]

    Multi-modal transformer and reinforcement learning-bas ed beam man- agement,

    M. Ghassemi, H. Zhang, A. Afana, A. B. Sediq, and M. Erol- Kantarci, “Multi-modal transformer and reinforcement learning-bas ed beam man- agement,” IEEE Netw. Lett. , vol. 6, no. 4, pp. 222–226, 2024

    2024

Pith tools

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