Pith. sign in

REVIEW 4 major objections 4 minor 45 references

Geometric Cross-Modal Token Selection for Latency-Constrained Multimodal Token Communication

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

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

Pith's one-line read The paper claims that tokens lying in the overlap of several anchor query regions in a shared query-key space are the ones worth transmitting, and reports up to 31.6% VQA and 29.2% AVQA accuracy gains over pairwise baselines at fixed latenc

desk verdict A well-engineered but empirically under-verified extension of the authors' own IBS idea; worth refereeing after fixing the abstract numbers, adding error bars, and running a k=1 control. read the letter →

arxiv 2608.03141 v1 pith:W7SDFLQA submitted 2026-08-04 eess.SP

Joohyuk Park , Junyong Shin , Yongjeong Oh , Jihong Park , Yo-Seb Jeon This is my paper
classification eess.SP
keywords cross-modaltokenselectioncommunicationcross-attentionsharedquery-keyspaceangulardistancesemanticgrainregionslatencyconstrainttoken-wiseerasurechannels
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 is trying to establish a selection rule for multimodal token communication: when only a fraction of image, text, and audio tokens can be transmitted within a latency budget, the tokens worth sending are the ones that multiple tokens of the other modality agree on. It constructs a shared query-key space by reusing the query and key projections of a cross-attention layer, defines an angular 'grain region' around each anchor query, and selects non-anchor tokens whose keys fall inside at least two such regions—the intersection-based selection (IBS) rule. The central claim is that this multi-query consensus captures shared cross-modal relevance better than pairwise attention scores, and the reported simulations support it with up to 31.6% VQA and 29.2% AVQA relative accuracy gains over existing selection baselines under the same latency, plus a robust variant for channels that erase tokens. A reader should care because it points to a cheap, geometry-based way to make remote multimodal inference adapt to bandwidth and channel conditions without retraining the model.

What carries the argument

The load-bearing object is the shared query-key space built from head-averaged cross-attention projections: the anchor modality's projected tokens act as queries, every other modality's projected tokens act as keys, and the metric is the angular distance $\Theta_{i,j} = \arccos(S_{i,j})$ between the $i$-th anchor query and the $j$-th non-anchor key. Around each anchor query sits a semantic grain region of angular radius $\theta_i$, modeled after germ-grain constructions. The IBS rule selects a non-anchor key only if it lies inside at least $k$ of these grain regions, and the optimization surrogate for task utility is the sum of cosine similarities over selected anchor--non-anchor pairs, subj

What would settle it

Measure each non-anchor token's marginal task utility by deleting it from the selected set and observing the VQA/AVQA accuracy change, then compare that rank order with the token's cosine similarity to its $k$ nearest anchor queries. If the rank correlation is near zero—or if randomly shuffling the key rows of one non-anchor modality leaves the reported IBS gains unchanged—the geometric alignment hypothesis is not carrying the result.

Watch

Extended reading notes

Core claim

The authors claim that shared cross-modal relevance is revealed better by multiplicity than by pairwise strength. They reuse the query and key projections from a sequential cross-attention layer to embed cross-modal tokens of all modalities in one space. The smallest modality defines anchor queries; around each anchor a semantic grain region is cut out by a per-query angular threshold. A non-anchor token is selected only when its key lies in the intersection of at least $k$ grain regions (IBS), and the anchor tokens selected are those whose regions contain such keys. The angular thresholds are optimized, under a bit-budget latency constraint, to maximize the summed cosine similarity of selec

Load-bearing premise

The load-bearing premise is that cosine similarity between a projected anchor query and a projected non-anchor key in the shared space reflects how much that key helps the downstream task; if that link is weak, the optimized objective is maximizing the wrong quantity.

Editorial extensions

If this is right

  • Under a tight latency budget, IBS is claimed to retain cross-modal content that pairwise top-$k$ attention drops, which is where the largest VQA/AVQA gains appear.
  • Optimizing each anchor's angular threshold to the local key distribution is claimed to matter for handling the uneven geometry of the shared query-key space.
  • On token-wise erasure channels, R-IBS is claimed to keep the same geometric selection effective by replacing each cosine similarity with its expectation over erasure states.
  • The greedy variant is claimed to approach the BCD-optimized accuracy in polynomial time, making the selection rule practical for large token counts.

Reading between the lines

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

  • Inference: If the alignment hypothesis generalizes, the same projection-reuse recipe could serve as a selection signal for any cross-attention Transformer without adding learned modules at deployment time.
  • Inference: A clean next experiment would compare IBS with $k=2$ and $k=1$ under identical cosine geometry and optimizer, isolating the overlap contribution from the choice of similarity score.
  • Inference: If cosine alignment tracks task utility, the angular-threshold optimization could run in closed loop, adjusting radii from feedback about which transmitted tokens the decoder actually consumed—an online variant the paper does not develop.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a geometry-based joint cross-modal token selection framework for latency-constrained multimodal token communication. The authors reuse cross-attention query/key projections to construct a shared query-key space, define semantic grain regions around anchor queries using angular distances, and introduce intersection-based token selection (IBS), which prioritizes non-anchor keys covered by at least k anchor grain regions. They formulate a latency-constrained surrogate optimization problem (P2) with query-specific angular thresholds, solve it via block coordinate descent and a low-complexity greedy algorithm, and extend the approach to token-wise erasure channels (R-IBS). Experiments on VQA and AVQA report task-accuracy gains over several baselines.

Significance. If the two working hypotheses were convincingly supported, the framework would be a useful contribution to semantic/token communication: it gives a tractable geometric surrogate for cross-modal token relevance, an explicit latency-constrained formulation, a polynomial-time greedy alternative to an NP-hard MINLP, and a channel-aware erasure extension. The paper also ships a concrete algorithmic machinery (BCD, MILP/ILP subproblems, greedy marginal-gain selection) that is clearly stated. However, the empirical evidence for the central causal claim is currently incomplete, and the headline results are not reported consistently. The significance is therefore conditional on the missing ablations and statistical validation being supplied.

major comments (4)
  1. [Sec. III-B, Sec. IV-B, Eq. (28a)] Hypothesis 2 is the load-bearing claim: intersection-based selection should improve accuracy because multi-query consensus is more reliable than single-query alignment. The only supporting evidence is the IBS-vs-OBS comparison, but OBS differs from IBS simultaneously in the relevance metric (symmetrized softmax attention scores in Eq. (49) vs. head-averaged cosine similarity in Eq. (22)), in the selection rule (aggregate pairwise relevance maximization vs. k-way intersection), and in the optimization mechanism (fixed scoring vs. query-specific angular thresholds). No ablation fixes the metric and threshold optimization and varies only the overlap parameter k. In particular, no k=1 single-anchor variant of the same geometric framework is reported. Without such a control, the claimed 27.6%/31.6% gains cannot be attributed to multi-query consensus; they may stem from the cosine metric or fr
  2. [Sec. III-B, Sec. IV-B, Eq. (28a)] Hypothesis 1 is stated as a working hypothesis but is not directly validated. The surrogate in (28a) replaces the true downstream utility U(.) with a sum of cosine similarities of selected anchor–non-anchor pairs. The paper says empirical support is provided in Sec. VI, but Sec. VI only compares the final IBS/R-IBS schemes against baselines; it does not test whether the cosine-similarity objective correlates with task accuracy. If cosine similarity is only weakly related to VQA/AVQA accuracy, the entire threshold optimization and selection rule lose their justification. Please provide direct evidence, e.g., a scatter/correlation plot between the surrogate (28a) and held-out accuracy for a range of selections, or an ablation in which the geometric objective is replaced by an oracle accuracy objective.
  3. [Abstract vs. Contributions vs. Sec. VI] The headline numbers are inconsistent. The abstract reports 'up to 31.6% and 29.2% task accuracy gains on VQA and AVQA, respectively.' The Contributions bullet reports latency-constrained gains of 27.6% (VQA) and 31.6% (AVQA), and erasure gains of 29.2% (VQA) and 27.1% (AVQA). Sec. VI states the 27.6% VQA gain at T_target=4.4 ms and the 29.2% R-IBS gain at erasure probabilities (0.2, 0.6). The abstract appears to swap the VQA and AVQA numbers and to omit the erasure-vs-latency distinction. Please correct the abstract and ensure every reported gain is tied to the correct task and channel condition.
  4. [Sec. VI, Figs. 3-4] The experimental evidence consists of single curves without error bars, number of seeds, or statistical tests. The reported improvements (27.6%, 29.2%, etc.) are point estimates, and typical fine-tuning variance for Transformer-based models is large enough that these claims need replication. In addition, the fine-tuning protocol does not state whether token selection is applied during training, at inference only, or both. If selection is inference-only, there is a train/test mismatch; if applied during training, the details (e.g., re-selection at each step, threshold initialization, synchronization with the BCD/greedy solvers) are missing. Please report mean±std over at least three runs and explicitly clarify when the selection operates.
minor comments (4)
  1. [Title page] The author line contains a spacing artifact: 'Y ongjeong Oh' should be 'Yongjeong Oh'.
  2. [Sec. IV-D, Eq. (35)] When the marginal cost in (33) is zero, the cost-efficiency ratio becomes infinite regardless of the epsilon term; the tie-breaking rule for such cases should be specified explicitly.
  3. [Sec. VI] The overlap threshold is fixed at k=2 in all experiments. Since this is a free parameter of the method and central to the multi-query consensus hypothesis, a sensitivity analysis over k (including k=1, 2, 3) is needed to support the choice and to separate modeling from optimization effects.
  4. [References] Reference [30] is the authors' own conference paper listed as 'to be published.' Please clarify the precise overlap and explicitly state which parts of the current manuscript are extended beyond [30].

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the selection framework is evaluated against external task accuracy rather than fitted to it, and the sole self-citation is disclosed prior work.

full rationale

The paper's derivation chain is not circular in the sense prohibited by the review rules. The central selection rule is built on two explicitly stated working hypotheses (Hypothesis 1 in Sec. III-B and Hypothesis 2 in Sec. IV-A), and the subsequent optimization problem (P2) maximizes a cosine-similarity surrogate under a latency constraint. The query-specific angular thresholds are decision variables optimized for this surrogate, not parameters fitted to the reported VQA/AVQA accuracy values. The simulations evaluate the resulting selection against held-out task accuracy on standard benchmarks (VQA v2, MUSIC-AVQA v2), which is external evidence rather than a renaming of the objective. The only self-citation is [30], the authors' own conference version, and it is used to disclose prior work rather than to justify the core claim. The absence of a k=1 ablation is a legitimate experimental-design concern: the OBS vs. IBS comparison is confounded by differences in relevance metric, selection rule, and threshold optimization, so the claimed empirical support for Hypothesis 2 is not fully isolated. However, this is a missing control or correctness issue, not circularity, because the claimed gains are measured against task accuracy and are not derived from the selection rule by construction. No equation reduces to its own input, and no fitted parameter is renamed as a prediction. Therefore no significant circularity is identified.

Assumptions & free parameters 3 free parameters · 6 assumptions · 2 invented entities

The central method is a heuristic geometric selection rule resting on two stated hypotheses about cosine similarity and multi-query consensus. It introduces no new physical entities, but it does introduce algorithmic constructions (grain regions, shared query-key space) whose value is only empirically claimed. The free parameters are the overlap threshold k and the per-query angular thresholds; the latter are optimization variables rather than accuracy-fitted constants, which keeps circularity burden low.

free parameters (3)
  • overlap threshold k = 2
    Prescribed in Sec. VI for all IBS runs; no ablation over k=3,4,... is reported, so the headline gains are conditional on this choice.
  • per-query angular thresholds {theta_i} = Optimized per query by BCD or greedy under the latency constraint
    Decision variables of (P2). They set grain region sizes and thus the selected set. They are fit to a cosine-similarity surrogate, not to task accuracy.
  • anchor modality selection rule = modality with fewest tokens
    A design choice that fixes which tokens are 'anchors'; different choices change grain geometry and selection outcomes.
assumptions (6)
  • domain assumption Geometric alignment hypothesis (Hypothesis 1): higher cosine similarity between anchor query and non-anchor key implies greater downstream task utility.
    Stated in Sec. III-B and used to build the surrogate objective in Eq. (28a); simulations are the only evidence.
  • domain assumption Multi-query consensus hypothesis (Hypothesis 2): a non-anchor key aligned with multiple anchors carries more reliable cross-modal relevance.
    Stated in Sec. IV-A and enforced by constraint (28b) with k>=2; validated only indirectly through accuracy comparisons.
  • ad hoc to paper The sum of cosine similarities in (28a) tracks the true utility U(.) in (P1).
    Used to turn (P1) into tractable (P2); no formal bound connects the surrogate to task accuracy.
  • domain assumption Head-averaged query and key projections, Eqs. (20)-(21), form a faithful shared metric space for token comparison.
    Assumes learned projections remain valid for token selection outside their cross-attention scoring role.
  • domain assumption Token-wise erasures are independent Bernoulli variables with probabilities from the outage model in (38).
    Used in the R-IBS expected cosine derivation; ignores correlations across tokens and subchannels.
  • ad hoc to paper Erased tokens are replaced by the bias row, Eqs. (44)-(45).
    A modeling choice for the receiver's handling of erased tokens; not derived from a specific real system.
invented entities (2)
  • semantic grain region G_i(theta_i)
    purpose: Set of non-anchor keys within angular threshold theta_i of anchor i; the intersection of k such regions defines IBS-selected tokens.
    A geometric construction introduced by the paper; its value is measured only through the paper's own simulations and no external falsifiable handle.
  • shared query-key space
    purpose: Common representation built by head-averaged query/key projections in Eqs. (20)-(21); used for all angular distances.
    Assumes the learned projections are a valid metric for token selection outside their attention role; no independent validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Geometric Cross-Modal Token Selection for Latency-Constrained Multimodal Token Communication." pith.science (2026). https://pith.science/paper/W7SDFLQA

@misc{pith2026260803141,
  author       = {Pith},
  title        = {Pith review of: Geometric Cross-Modal Token Selection for Latency-Constrained Multimodal Token Communication},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W7SDFLQA}},
  note         = {Machine review of arXiv:2608.03141}
}
read the original abstract

This paper proposes a geometry-based joint cross-modal token selection framework for latency-constrained multimodal token communications. To capture cross-modal token dependencies, we leverage the cross-attention mechanism to project modality-specific tokens into a shared query-key space, where the modality with the fewest tokens serves as the anchor modality and the others as non-anchor modalities. Inspired by germ-grain models, we define an angular-distance metric and construct semantic grain regions around anchor queries. Based on this geometric representation, we identify cross-modal evidence shared across multiple anchor queries in this space and develop an intersection-based token selection (IBS) strategy that prioritizes non-anchor tokens whose key are covered by multiple grain regions. We further develop an erasure-aware extension, termed robust-IBS (R-IBS), for token-wise erasure channels using an expected angular-distance formulation. In both IBS and R-IBS, the grain regions are optimized for individual queries under a latency constraint, using block coordinate descent and a low-complexity greedy algorithm. Simulations corroborate the effectiveness of IBS and R-IBS under latency-constrained and token-wise erasure channels, achieving up to 31.6% and 29.2% task accuracy gains on visual question answering (VQA) and audio-visual question answering (AVQA) tasks, respectively, over existing token selection baselines.

Figures

Figures reproduced from arXiv: 2608.03141 by the authors.

Figure 1
Figure 1. An illustration of the proposed geometry-based join [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. An illustration of the proposed sequential cross-at [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Comparison of VQA and AVQA accuracy across various to [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Comparison of VQA and AVQA accuracy across various to ˆ [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Visualization of the cross-modal token selection re ˆ [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 37 canonical work pages

  1. [30]

    Geometr y-driven joint token selection for multimodal token communications,

    J. Park, J. Shin, Y . Oh, J. Park, and Y .-S. Jeon, “Geometr y-driven joint token selection for multimodal token communications,” in Proc. IEEE Global Commun. Conf. (GLOBECOM) , 2026, to be published

    work page 2026

  2. [1]

    Semantic communications : Overview, open issues, and future research directions,

    X. Luo, H.-H. Chen, and Q. Guo, “Semantic communications : Overview, open issues, and future research directions,” IEEE Wireless Commun. , vol. 29, no. 1, pp. 210–219, Feb. 2022

    2022

  3. [2]

    A su rvey on multimodal large language models,

    S. Yin, C. Fu, S. Zhao, K. Li, X. Sun, T. Xu, and E. Chen, “A su rvey on multimodal large language models,” Natl. Sci. Rev. , vol. 11, no. 12, pp. 1–20, Dec. 2024

    work page 2024

  4. [3]

    Visual instruction tun ing,

    H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tun ing,” in Adv. Neural Inf. Process. Syst. (NeurIPS), New Orleans, LA, USA, Dec. 2023, pp. 34 892–34 916

    work page 2023

  5. [4]

    Matryoshka query transformer for large vision-language m odels,

    W. Hu, Z.-Y . Dou, L. H. Li, A. Kamath, N. Peng, and K.-W. Cha ng, “Matryoshka query transformer for large vision-language m odels,” in Adv. Neural Inf. Process. Syst. (NeurIPS) , V ancouver, BC, Canada, Dec. 2024

    work page 2024

  6. [5]

    MobileVLM V2: Faster and stronger baseline for vision language model,

    X. Chu et al. , “MobileVLM V2: Faster and stronger baseline for vision language model,” 2024, arXiv:2402.03766

    arXiv 2024

  7. [6]

    Honeybee: Locality-e nhanced projector for multimodal LLM,

    J. Cha, W. Kang, J. Mun, and B. Roh, “Honeybee: Locality-e nhanced projector for multimodal LLM,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) , Seattle, W A, USA, Jun. 2024, pp. 13 817– 13 827

    work page 2024

  8. [7]

    BLIP-2: Bootstrapp ing language- image pre-training with frozen image encoders and large lan guage models,

    J. Li, D. Li, S. Savarese, and S. Hoi, “BLIP-2: Bootstrapp ing language- image pre-training with frozen image encoders and large lan guage models,” in Proc. Int. Conf. Mach. Learn. (ICML) , Honolulu, HI, USA, Jul. 2023, pp. 19 730–19 742

    work page 2023

Show all 45 references
  1. [8]

    Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond ,

    J. Bai et al. , “Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond ,” 2023, arXiv:2308.12966

    2023 arXiv

  2. [9]

    TG-LLaV A: Text guided LLaV A via learnable latent embeddings,

    D. Y an et al. , “TG-LLaV A: Text guided LLaV A via learnable latent embeddings,” in Proc. AAAI Conf. Artif. Intell. , Philadelphia, PA, USA, Feb. 2025

    2025

  3. [10]

    Token communications: A large model-driven Framewor k for cross-modal context-aware semantic communications,

    L. Qiao, M. B. Mashhadi, Z. Gao, R. Tafazolli, M. Bennis, and D. Niy- ato, “Token communications: A large model-driven Framewor k for cross-modal context-aware semantic communications,” IEEE Wireless Commun., vol. 32, no. 5, pp. 80–88, Oct. 2025

    2025

  4. [11]

    Task-oriented mul timodal token transmission in resource-constrained multiuser net works,

    J. Zhang, W. Ni, P . Wang, and D. Wang, “Task-oriented mul timodal token transmission in resource-constrained multiuser net works,” IEEE Wireless Commun. Lett. , vol. 15, pp. 570–574, Nov. 2025

    2025

  5. [12]

    ToDMA: Large model-driven token-domain multiple access for seman tic commu- nications,

    L. Qiao, M. B. Mashhadi, Z. Gao, R. Schober, and D. G¨ und¨uz, “ToDMA: Large model-driven token-domain multiple access for seman tic commu- nications,” 2025, arXiv:2505.10946

    2025 arXiv

  6. [13]

    Adaptive semantic token communication for transformer-b ased edge inference,

    A. Devoto, J. Pomponi, M. Merluzzi, P . D. Lorenzo, and S. Scardapane, “Adaptive semantic token communication for transformer-b ased edge inference,” IEEE Trans. Mach. Learn. Commun. Netw. , vol. 4, pp. 422– 437, Jan. 2026

    2026

  7. [14]

    Conte xt-aware iterative token detection and masked transmission for wire less token communication,

    J. Shin, J. Park, J. Park, J. Choi, and Y .-S. Jeon, “Conte xt-aware iterative token detection and masked transmission for wire less token communication,” in Proc. AAAI Conf. Artif. Intell. , Singapore, Jan. 2026

    2026

  8. [15]

    Text-guided token communication for wireless image transmission,

    B. Liu et al. , “Text-guided token communication for wireless image transmission,” in Proc. IEEE Int. Conf. Commun. China (ICCC) , Shang- hai, China, Aug. 2025

    2025

  9. [16]

    Video TokenCom: Textual intent-guided multi-rate video t oken com- munications with UEP-based adaptive source–channel codin g,

    J. Men, M. B. Mashhadi, N. Wang, Y . Ma, M. Nilsson, and R. T afazolli, “Video TokenCom: Textual intent-guided multi-rate video t oken com- munications with UEP-based adaptive source–channel codin g,” 2026, arXiv:2603.02470

    2026

  10. [17]

    Tok en communication in the era of large models: An information bot tleneck- based approach,

    H. Wei, W. Ni, W. Wang, W. Xu, D. Niyato, and P . Zhang, “Tok en communication in the era of large models: An information bot tleneck- based approach,” IEEE Wireless Commun. Lett. , vol. 15, pp. 186–190, Oct. 2025

    2025

  11. [18]

    TokenCom: Vision-language model for multimodal token communications,

    F. Jiang et al., “TokenCom: Vision-language model for multimodal token communications,” 2026, arXiv:2603.00482

    2026

  12. [19]

    Joint seman tic-channel coding and modulation for token communications,

    J. Ying, Z. Qin, Y . Feng, L. Wang, and X. Tao, “Joint seman tic-channel coding and modulation for token communications,” IEEE Trans. Wireless Commun., vol. 25, pp. 8179–8193, Dec. 2025

    2025

  13. [20]

    Learning transferable visual models from natural language supervision,

    A. Radford et al. , “Learning transferable visual models from natural language supervision,” in Proc. Int. Conf. Mach. Learn. (ICML) , Jul. 2021, pp. 8748–8763

    2021

  14. [21]

    A tria ngle enables multimodal alignment beyond cosine similarity,

    G. Cicchetti, E. Grassucci, and D. Comminiello, “A tria ngle enables multimodal alignment beyond cosine similarity,” in Adv. Neural Inf. Process. Syst. (NeurIPS) , San Diego, CA, USA, Dec. 2025

    2025

  15. [22]

    CA DeepSC: Cross-modal alignment for multi-modal semantic communications,

    W. Wang, M. Liu, and M. Chen, “CA DeepSC: Cross-modal alignment for multi-modal semantic communications,” in Proc. IEEE Global Commun. Conf. (GLOBECOM) , Kuala Lumpur, Malaysia, 2023, pp. 5871–5876

    2023

  16. [23]

    Unifying modalities through sem antic embed- dings,

    E. Kutay and A. Y ener, “Unifying modalities through sem antic embed- dings,” in Proc. IEEE Global Commun. Conf. (GLOBECOM) , Taipei, Taiwan, 2025, pp. 4675–4680

    2025

  17. [24]

    Large model e mpowered multi-modal semantic communication with selective tokens for training,

    J. Peng, H. Xing, Z. Xiao, L. Xu, and X. Lei, “Large model e mpowered multi-modal semantic communication with selective tokens for training,” IEEE Signal Process. Lett. , vol. 32, pp. 2967–2971, Jul. 2025

    2025

  18. [25]

    Not all patches are what you need: Expediting vision transformers v ia token reorganizations,

    Y . Liang, C. Ge, Z. Tong, Y . Song, J. Wang, and P . Xie, “Not all patches are what you need: Expediting vision transformers v ia token reorganizations,” in Proc. Int. Conf. Learn. Represent. (ICLR) , Apr. 2022

    2022

  19. [26]

    S. N. Chiu, D. Stoyan, W. S. Kendall, and J. Mecke, Stochastic Geometry and Its Applications . Chichester, U.K.: John Wiley & Sons, Ltd., 2013

    2013

  20. [27]

    Block coordinate descent on smoot h manifolds: Convergence theory and twenty-one examples,

    L. Peng and R. Vidal, “Block coordinate descent on smoot h manifolds: Convergence theory and twenty-one examples,” 2 023, arXiv:2305.14744

    arXiv

  21. [28]

    Making the V in VQA matter: Elevating the role of image understandin g in visual question answering,

    Y . Goyal, T. Khot, D. Summers-Stay, D. Batra, and D. Pari kh, “Making the V in VQA matter: Elevating the role of image understandin g in visual question answering,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) , Honolulu, HI, USA, Jul. 2017, pp. 6904–6913

    2017

  22. [29]

    Tackling data bias in MUSI C-A VQA: Crafting a balanced dataset for unbiased question-answeri ng,

    X. Liu, Z. Dong, and P . Zhang, “Tackling data bias in MUSI C-A VQA: Crafting a balanced dataset for unbiased question-answeri ng,” in Proc. IEEE/CVF Winter Conf. Appl. Comput. Vis. (WACV) , 2024, pp. 4478– 4487

    2024

  23. [31]

    A unified multi- task semantic communication system for multimodal data,

    G. Zhang, Q. Hu, Z. Qin, Y . Cai, G. Y u, and X. Tao, “A unified multi- task semantic communication system for multimodal data,” IEEE Trans. Commun., vol. 72, no. 7, pp. 4101–4116, Jul. 2024

    2024

  24. [32]

    How much training is need ed in multiple-antenna wireless links?,

    B. Hassibi and B. M. Hochwald, “How much training is need ed in multiple-antenna wireless links?,” IEEE Trans. Inf. Theory , vol. 49, no. 4, pp. 951–963, Apr. 2003

    2003

  25. [33]

    ViLBERT: Pretrai ning task- agnostic visiolinguistic representations for vision-and -language tasks,

    J. Lu, D. Batra, D. Parikh, and S. Lee, “ViLBERT: Pretrai ning task- agnostic visiolinguistic representations for vision-and -language tasks,” in Adv. Neural Inf. Process. Syst. (NeurIPS) , V ancouver, BC, Canada, Dec. 2019, pp. 13–23

    2019

  26. [34]

    An image is worth 16 ×16 words: Transformers for image recognition at scale,

    A. Dosovitskiy et al. , “An image is worth 16 ×16 words: Transformers for image recognition at scale,” in Proc. IEEE Int. Conf. Learn. Repre- sent. (ICLR) , May 2021, pp. 1–21

    2021

  27. [35]

    Hubs i n Space: Popular Nearest Neighbors in High-Dimensional Data,

    M. Radovanovic, A. Nanopoulos, and M. Ivanovic, “Hubs i n Space: Popular Nearest Neighbors in High-Dimensional Data,” J. Mach. Learn. Res., vol. 11, no. 86, pp. 2487–2531, Sep. 2010

    2010

  28. [36]

    A branch-and-cut algorithm for the resolution of large-scale symmetric traveling salesman pr oblems,

    M. Padberg and G. Rinaldi, “A branch-and-cut algorithm for the resolution of large-scale symmetric traveling salesman pr oblems,” SIAM Rev., vol. 33, no. 1, pp. 60–100, Mar. 1991

    1991

  29. [37]

    G. L. Nemhauser and L. A. Wolsey, Integer and Combinatorial Opti- mization. New Y ork, NY , USA: John Wiley & Sons, 1988

    1988

  30. [38]

    Synchronous multi-modal semantic communication system with packet-level coding,

    Y . Tian, J. Ying, Z. Qin, Y . Jin, and X. Tao, “Synchronous multi-modal semantic communication system with packet-level coding,” IEEE Trans. Wireless Commun., vol. 24, no. 5, pp. 3684–3697, May 2025

    2025

  31. [39]

    Low-complexity se mantic packet aggregation for token communication via lookahead search,

    S. Lee, J. Park, J. Choi, and H. Park, “Low-complexity se mantic packet aggregation for token communication via lookahead search, ” 2025, arXiv:2506.19451

    2025 arXiv

  32. [40]

    Goldsmith, Wireless Communications

    A. Goldsmith, Wireless Communications. Cambridge, U.K.: Cambridge Univ. Press, 2005

    2005

  33. [41]

    BERT: Pre-training of deep bidirectional transformers for language understan ding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of deep bidirectional transformers for language understan ding,” in Proc. Conf. North Amer . Chapter Assoc. Comput. Linguistics: Hum. Lang. Technol., Minneapolis, MN, USA, Jun. 2019, pp. 4171–4186

    2019

  34. [42]

    wav2vec 2. 0: A framework for self-supervised learning of speech represen tations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2. 0: A framework for self-supervised learning of speech represen tations,” in Adv. Neural Inf. Process. Syst. (NeurIPS) , Dec. 2020, pp. 12 449–12 460

    2020

  35. [43]

    Decoupled weight decay re gularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay re gularization,” in Proc. Int. Conf. Learn. Represent. (ICLR) , New Orleans, LA, USA, May 2019

    2019

  36. [44]

    An image is worth 1/2 tokens after layer 2: Plug-and-play inference accelera tion for large vision-language models,

    L. Chen, H. Zhao, T. Liu, B. Chang, and Z. Jiang, “An image is worth 1/2 tokens after layer 2: Plug-and-play inference accelera tion for large vision-language models,” in Proc. Eur . Conf. Comput. Vis. (ECCV) , Seattle, W A, USA, Nov. 2024, pp. 19–35

    2024

  37. [45]

    Multidimensional scaling: I. Theory and method,

    W. S. Torgerson, “Multidimensional scaling: I. Theory and method,” Psychometrika, vol. 17, no. 4, pp. 401–419, Dec. 1952

    1952

Pith tools

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