Pith. sign in

REVIEW 5 major objections 6 minor 44 references

M3LLM: Model Context Protocol-aided Mixture of Vision Experts For Multimodal LLMs in Networks

T0 review · 5 major / 6 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 A multimodal LLM can improve accuracy by routing each query to a distributed vision expert chosen for both task fit and wireless channel quality, rather than using one fixed visual encoder.

desk verdict A genuinely new idea for wireless-aware expert routing, but the paper's central inference step is undefined and its headline claim doesn't match its own table. read the letter →

arxiv 2508.01805 v1 pith:7FOFVXYJ submitted 2025-08-03 cs.NI

Yongjie Zeng , Hongyang Du This is my paper · ORCID
classification cs.NI
keywords multimodallargelanguagemodelsmixtureofvisionexpertsmodelcontextprotocolwirelessedgeinferenceexpertroutingsoftactor-critichierarchicalBayesianstate-spaceretrieval-augmentedgeneration
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 sets out to show that a multimodal large language model does not have to carry every visual capability in one monolithic encoder. It proposes M3LLM, a distributed framework in which a central MLLM backbone calls specialized vision experts hosted on edge devices over wireless links, using the Model Context Protocol (MCP) to encode task context and expert capabilities in a common format. Routing happens in two stages: retrieval-augmented filtering prunes the expert pool by semantic relevance, then a dual-stream Soft Actor-Critic agent chooses among survivors by balancing task-expert compatibility against channel quality, with a hierarchical Bayesian module supplying stability priors. The paper reports that M3LLM outperforms baselines such as MoVA on both accuracy and network metrics in a simulated wireless environment.

What carries the argument

The load-bearing mechanism is the M3LLM routing pipeline: MCP-aided RAG (retrieval-augmented generation over MCP-exposed capability descriptors) for coarse expert filtering; CE-SAC, a dual-stream Soft Actor-Critic with decoupled expert and channel critics, for fine-grained joint routing; and ASEM, a hierarchical Bayesian state-space module with variational inference that extracts short-term and long-term latents from channel observations. These components feed a final weighted aggregation in which the response is $\hat{R} = \sum_i w^*[i] R_i$, where each $R_i$ is the transmitted output of expert $i$ and $w^*$ is the normalized product of expert and channel weights.

What would settle it

Run the seven listed vision experts over a real wireless testbed on a foggy traffic-light query and a medical-image query, then compare the fused output $\hat{R} = \sum_i w^*[i] R_i$ against each single expert's output at the same SNR; if the weighted aggregate does not reproduce the best single expert's correct answer, the aggregation step is the failure point.

Watch

Extended reading notes

Core claim

The central claim is that expert selection for MLLMs should be a joint, network-aware optimization rather than a purely semantic one. M3LLM makes this concrete by having the MCP-aided RAG stage produce a binary mask of semantically plausible experts, and then having CE-SAC, a dual-stream Soft Actor-Critic with separate critics for semantic reward and channel reward, produce per-expert weights that combine the two objectives multiplicatively: $w_{\text{final}} = w_{\text{expert}} \odot w_{\text{channel}} / (\|w_{\text{expert}} \odot w_{\text{channel}}\|_1 + \epsilon)$. ASEM's hierarchical Bayesian latents stabilize the state fed to the policy against bursty channel noise. The paper claims this yields up to 51% higher task accuracy under realistic wireless conditions than MoVA, and shows a 15.9% higher LLM Quality Score on the MME benchmark.

Load-bearing premise

The load-bearing premise is that outputs produced by different vision experts on different devices can be weighted and summed without losing the task-relevant visual information.

Editorial extensions

If this is right

  • If M3LLM works as claimed, an MLLM's visual capabilities no longer have to be fixed at training time; new specialists can join the network and be used as they appear.
  • Routing that explicitly weighs task-expert fit against channel conditions should reduce communication cost and latency, since poor links are avoided before large outputs are transmitted.
  • The decoupled critic design provides a template for combining semantic and communication objectives in other networked-AI decision problems without reward interference.
  • Existing centralized MoE-style MLLMs could be converted into distributed systems by exposing their vision experts through MCP-compatible endpoints and keeping the same backbone.

Reading between the lines

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

  • The paper leaves the representation of each expert output $R_i$ unspecified; a natural test is to define $R_i$ as a probability distribution over a shared answer space and check whether weighted log-likelihood fusion preserves accuracy.
  • Because MCP is an open protocol, the same interface could plausibly coordinate non-vision AI services across the edge, making the routing mechanism itself task-agnostic.
  • A cleaner evaluation of the headline claim would fix a single accuracy metric on MME/ScienceQA and report the gain over MoVA on that metric, since the abstract claims up to 51% while Table I shows a 15.9% quality-score gap.
  • The two reported failure modes (semantically ambiguous queries and extreme channel volatility) suggest direct stress tests: near-duplicate task tags and rapid SNR bursts should degrade routing stability if the failure analysis is accurate.
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 M3LLM, a distributed framework for multimodal LLM inference in wireless edge networks. The framework uses the Model Context Protocol (MCP) to expose vision experts on edge devices, a RAG-based coarse filtering stage to select semantically relevant experts, and a dual-stream Soft Actor-Critic (CE-SAC) agent to jointly optimize expert selection and channel-aware routing. An Adaptive Stability Enhancement Module (ASEM) based on hierarchical Bayesian variational inference is introduced to produce stability-aware state representations. Experiments are reported on MME and ScienceQA against Random, MoVA, EdgeViT, and MoE-LLaVA baselines, with claims of improved task accuracy, reduced communication cost, and enhanced routing adaptability.

Significance. If the proposed framework worked as described, it would address a timely and important problem: dynamically composing specialized vision experts distributed across wireless edge devices for multimodal LLM inference. The architectural ideas—MCP-based capability advertisement, decoupled critics for semantic and communication rewards, and stability-aware state representations—are plausible and could be useful to the networking and MLLM communities. However, the manuscript as written does not establish the central claims. The core aggregation operation is unspecified, the headline accuracy improvement is contradicted by the reported table, the communication-cost benefit is claimed but never measured, and the network evaluation metrics largely re-use the optimized reward terms. These issues prevent the results from being reproduced or interpreted as evidence for the proposed system.

major comments (5)
  1. [Section I and Table I] The abstract and Section I claim that M3LLM 'achieves up to 51% higher multimodal task accuracy under realistic wireless conditions compared to state-of-the-art MLLMs such as MoVA.' Table I reports an LLM Quality Score of 0.730 for M3LLM versus 0.630 for MoVA, which is a 15.9% relative improvement, and the table contains no separate task-accuracy metric. The 51% figure is therefore unsupported by the data presented and should be corrected or removed.
  2. [Algorithm 1, lines 15-21] The response aggregation step is structurally underspecified. Line 18 defines R_i as the result of e_i.invoke(M_i, channel_i), and line 21 computes \hat R = \sum_i w^*[i] R_i. The expert pool in Section IV-A includes DINOv2 (feature embeddings), Co-DETR (bounding boxes), SAM (segmentation masks), Pix2Struct/Deplot (structured text), and BiomedCLIP (multimodal embeddings). These outputs live in incompatible spaces, and no projection, fusion layer, or algebraic structure is provided that would make a weighted sum meaningful. Without this definition, the reported LLM Quality Score cannot be traced to a concrete inference procedure, and the accuracy claims are not reproducible.
  3. [Abstract and Section IV] The abstract claims that M3LLM 'reduces communication cost,' and the conclusion repeats the claim, but no communication-cost metric, measurement, or table appears anywhere in Section IV. The evaluation covers task-semantic and wireless network-aware metrics, but communication overhead is never quantified. This is a load-bearing claimed benefit and needs either explicit measurement or removal from the claims.
  4. [Equation (21), Section III-B.1] ASEM is presented as a central contribution, but its model is not actually specified. Equation (21) is a generic ELBO with no statement of the priors p(z_1, z_2), the likelihood p(q | z_1, z_2), the encoder and decoder architectures, or the training procedure. The ablation in Table II attributes a 14.9% Channel Stability drop to removing ASEM, but without a precise model definition it is impossible to assess whether the implementation matches the claimed hierarchical Bayesian design or to reproduce the result.
  5. [Table I and Section II-C.2] The wireless network-aware evaluation metrics appear to be the same quantities that are optimized by the channel reward. Equation (13)-(17) define R_channel as a weighted sum of normalized average SNR, a stability term, a load-distribution entropy, and spectral efficiency. Table I's 'Channel Quality,' 'SNR Quality,' and 'Channel Stability' scores are not defined independently in the evaluation section, and they read as direct reports of the reward components in (14) and (15). The reported network gains therefore partly re-report the optimized objective, not an independent evaluation. Similarly, the primary 'LLM Quality Score' resembles the R4 reward component in Eq. (12). Independent, clearly defined evaluation metrics are needed before the comparison can be interpreted.
minor comments (6)
  1. [Algorithm 1] The loop structure 'while t = 1 do ... t <- t + 1' executes exactly once; this should be written as a single pass or as a genuine loop with a termination condition.
  2. [Section IV-D] In the MoVA Baseline paragraph, 'network-unawre design' should be 'network-unaware design.'
  3. [Section V-C] The text cites 'MoE-RL [15]' but reference [15] is MoE-LLaVA; the intended MoE-RL work is not present in the bibliography.
  4. [Table I] The metrics 'Task-Expert Alignment' and 'Expert Diversity' are reported numerically but their computation is never defined, so these results cannot be reproduced or interpreted.
  5. [Conclusion] The conclusion states that M3LLM achieves 'near-oracle task semantic accuracy,' but no oracle comparison is reported anywhere in the evaluation.
  6. [References] Reference [43] is incomplete; the entry for Mnih et al. ends mid-sentence without publication venue or year.

Circularity Check

2 steps flagged · score 6.0 of 10

Network-performance numbers in Table I re-report the Rchannel reward by construction, and the semantic score overlaps the LLM-eval reward, so the headline gains are not independently grounded.

  1. self definitional [Section II-C2 (Eqs. 13-17) and Section IV-A 'Wireless Network-aware Performance']
    "Rchannel = w1 ¯Q + w2S + w3D + w4E, (13) where ¯Q = 1/|A| Σ_{i∈A} (SNRi(t)-SNRmin)/(SNRmax-SNRmin), (14) S = 1 - σ_SNR(A)/(¯Q+ϵ), (15) D = -Σ_{i∈A} p_i log p_i, (16) E = Σ_{i∈A} log2(1+SNRi(t))/(|A|·log2(1+SNRmax)), (17). ... The Channel Quality score provides a holistic assessment integrating SNR, stability, load distribution, and spectral efficiency."

    CE-SAC is trained to maximize Rchannel, whose components are exactly average normalized SNR, SNR-variation stability, load-distribution entropy, and spectral efficiency. The evaluation metrics 'Channel Quality', 'SNR Quality', and 'Channel Stability' are the same functions of the same channel-state quantities. Table I's network columns therefore report the optimized reward itself rather than an independent outcome such as transmitted bytes, latency, or end-to-end task accuracy under channel errors. The claimed 'reduces communication cost' is supported only by the entropy term D inside the reward, so it holds by construction.

  2. self definitional [Algorithm 1 (lines 22-23) and Section III-B3 'Reward Decomposition and Meta-Analysis']
    "Compute quality scores Qsemantic, Qchannel; Qtotal ← αQsemantic + βQchannel. ... Although the expert and channel critics are trained independently on their respective rewards, we compute a combined score Rfinal = αRLLM + βRchannel as system overall performance."

    The evaluation pipeline computes Qsemantic and Qchannel and combines them with the same α, β weights used to form the training reward Rfinal. The semantic reward RLLM contains an LLM-evaluation term R4 = LLMeval(...), and the primary reported semantic metric is an 'LLM Quality Score' judging output accuracy, relevance, and coherence; no standard MME/ScienceQA accuracy is reported. Thus the 'improves task accuracy' and 'reduces communication cost' claims are evaluated by the same objective the policy was trained to maximize, making the headline results a re-statement of the optimized reward rather than an independent test.

full rationale

The core circularity is evaluative: the wireless metrics in Table I are constructed from the same quantities (Eqs. 14-17) as the Rchannel reward, and the semantic quality score is not separated from the LLM-eval reward used in RLLM. Algorithm 1's Qtotal / Rfinal identity makes the overlap explicit. I did not find load-bearing self-citation, imported uniqueness theorems, or ansatz-smuggling through citations; the citations to prior wireless and MLLM work are background. The 51% versus 15.9% discrepancy is an internal inconsistency, not circularity. The underspecified aggregation of heterogeneous expert outputs (DINOv2 embeddings, SAM masks, Deplot tables) in Algorithm 1 is a serious correctness risk but is not a circularity pattern. Overall, the network-performance gains reduce by construction to the training reward, while the semantic gains partially overlap the reward, giving a partial-circularity score of 6.

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

The central system rests on standard reinforcement learning and Bayesian tools; the free parameters are largely reward weights and simulator scenario settings chosen by hand. The axioms include domain assumptions about MCP semantic encoding, weighted output aggregation, and LLM evaluation. No new physical entities are introduced; CE-SAC and ASEM are algorithmic modules built from published methods.

free parameters (4)
  • Channel reward weights w1,w2,w3,w4 = (0.4, 0.3, 0.2, 0.1)
    Hand-set weights in Eq. 13 that define the channel reward; the Channel Quality evaluation metric is built from the same weighted components, so these choices directly shape the reported network gains. No sensitivity analysis is provided.
  • LLM reward weights alpha1-alpha4 and thresholds = (0.4, 0.3, 0.2, 0.1) with theta_act=0.2, theta_sup=0.1
    Hand-set in Eq. 8-10; these determine what counts as correct expert activation and influence the semantic reward. No tuning analysis is reported.
  • Combined reward weight alpha=beta=0.5 = 0.5
    Used in R_final = alpha R_LLM + beta R_channel; hand-set without sensitivity testing.
  • Wireless simulator scenario parameters = mu_SNR=25 dB, sigma_SNR=5 dB, mu_d=275 m, sigma_d=75 m, mu_shadow=8 dB, sigma_shadow=1.3 dB, n=3.5, rho=0.9
    These parameters define the simulated network environment (Section IV-A). Results may be specific to this scenario, and no real wireless traces or deployments are used.
assumptions (5)
  • domain assumption Rayleigh fading with h_i(t) ~ CN(0,1) and Gauss-Markov shadowing are representative of real edge wireless channels
    Eqs. 4-7 define the simulator. If real channels deviate from these statistical models, the channel-aware routing gains may not transfer.
  • domain assumption MCP can encode input context into learnable representations that capture task semantics and align with expert capability descriptors
    Introduced in Section II-A and used by coarse routing in Section III-A; the authors themselves note in Section IV-D that ambiguous queries degrade this representation.
  • domain assumption Weighted linear aggregation of independently invoked expert outputs preserves task-relevant information
    Algorithm 1 (lines 15-21) forms the final response as sum_i w*[i] R_i; the representation of R_i is never specified, and the fusion is not validated.
  • domain assumption LLM-based judgments (R4 reward and the LLM Quality Score metric) are faithful proxies for task quality
    Eq. 12 uses an external LLM to assess output consistency, and the primary evaluation metric is an LLM-customized quality score; no human ground truth or standard MME accuracy is reported.
  • ad hoc to paper ASEM latent variables z1 and z2 capture channel volatility and expert reliability with a valid variational approximation
    Section III-B1 describes only a generic ELBO (Eq. 21); the generative model, priors, and inference network are unspecified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of M3LLM: Model Context Protocol-aided Mixture of Vision Experts For Multimodal LLMs in Networks." pith.science (2026). https://pith.science/paper/7FOFVXYJ

@misc{pith2026250801805,
  author       = {Pith},
  title        = {Pith review of: M3LLM: Model Context Protocol-aided Mixture of Vision Experts For Multimodal LLMs in Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7FOFVXYJ}},
  note         = {Machine review of arXiv:2508.01805}
}
read the original abstract

Current Multimodal Large Language Models (MLLMs) rely on centralized architectures and often suffer from poor alignment between the input task and their fixed visual encoding modules, which limits performance on diverse and dynamic visual tasks. With the increasing deployment of resource-efficient models on edge devices in wireless networks, a new opportunity emerges to dynamically use distributed vision experts for improved MLLM inference quality. To enable this, we propose M3LLM, where the Model Context Protocol (MCP) coordinates a mixture of vision experts to achieve distributed MLLMs. Specifically, MCP is an open protocol that structures the input task context into interpretable representations, enabling wireless network-aware coordination between the central model backbone and edge-hosted vision experts. Based on the MCP representation, M3LLM formulates vision expert routing as a joint optimization problem that balances task-expert semantic compatibility and channel performance. To solve the resulting gradient conflicts, we develop a dual-stream Soft Actor-Critic (SAC) algorithm with decoupled reward signals and introduce an Adaptive Stability Enhancement Module (ASEM) based on hierarchical Bayesian modeling to ensure effective routing. Experiments show that M3LLM improves task accuracy, reduces communication cost, and enhances expert routing adaptability under dynamic wireless network conditions.

Figures

Figures reproduced from arXiv: 2508.01805 by the authors.

Figure 1
Figure 1. The system model of M3LLM in wireless networks. adaptive MLLMs requires solving two fundamental challenges: • Interoperable expert connectivity: Defining unified inter￾faces for invoking heterogeneous vision experts across devices with different platforms and capabilities. • Network-aware expert coordination: Designing scheduling mechanisms that jointly consider task semantic require￾ments and dynamic network condit… view at source ↗
Figure 2
Figure 2. Expert routing scheme of M3LLM. Stage 1 performs coarse-grained expert filtering via MCP-aided RAG. Stage 2 executes network-aware fine-grained expert routing using a decoupled DRL agent, i.e., CE-SAC, which leverages a stability-aware state representation from ASEM. expert weight distribution wexpert, as a task semantic-aware design principle: RLLM = X 4 i=1 αi · Ri(wexpert, τ, O), (8) where αi are dimension weight… view at source ↗
Figure 3
Figure 3. Training dynamics comparison across 1, 000 episodes. (a) Total reward evolution showing M3LLM’s superior convergence and final performance. (b) LLM reward demonstrating consistent semantic quality improvements. (c) Channel reward highlighting M3LLM’s unique ability to optimize network quality while baselines remain static. Shaded areas represent confidence intervals. TABLE I COMPREHENSIVE PERFORMANCE COMPARISON ON T… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 28 canonical work pages

  1. [1]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” inProc. NeurIPS, 2017, pp. 5998–6008

    work page 2017

  2. [2]

    Flamingo: a visual language model for few-shot learning,

    J.-B. Alayrac, J. Donahue, P. Luc, A. Miech, I. Barr, Y . Hasson, K. Lenc, A. Mensch, K. Millican, M. Reynolds, R. Ring, E. Rutherford, S. Cabi, T. Han, Z. Glaese, A. Firman, R. Toth, B. Brunner, F. Menick, E. Leblond, E. Rae, R. Razavi, L. B. Uesato, C. Dyer, O. Vinyals, K. Kavukcuoglu, R. Pascanu, L. Sifre, and S. Osindero, “Flamingo: a visual language ...

    work page 2022

  3. [3]

    BLIP-2: Bootstrapping language- image pre-training with frozen image encoders and large language models,

    J. Li, D. Li, S. Savarese, and S. Hoi, “BLIP-2: Bootstrapping language- image pre-training with frozen image encoders and large language models,” in Proc. ICML, 2023, pp. 19730–19742

    work page 2023

  4. [4]

    Visual instruction tuning,

    Hao Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” in Proc. NeurIPS, 2023, pp. 34873–34885

    work page 2023

  5. [5]

    GPT-4V(ision) system card,

    OpenAI, “GPT-4V(ision) system card,” OpenAI Technical Report, 2023, https://openai.com/research/gpt-4v-system-card

    work page 2023

  6. [6]

    Gemini: A family of highly capable multimodal models,

    Gemini Team, “Gemini: A family of highly capable multimodal models,” arXiv preprint arXiv:2312.11805, 2023

    arXiv 2023

  7. [7]

    Multimodal Large Language Models: A Survey,

    J. Wu, W. Gan, Z. Chen, S. Wan, and P. S. Yu, “Multimodal Large Language Models: A Survey,” arXiv.org, Nov. 22, 2023. https://arxiv.org/abs/2311.13165

    arXiv 2023

  8. [8]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,” in Proc. ICML, 2021, pp. 8748–8763

    work page 2021

Show all 44 references
  1. [9]

    GPipe: Efficient training of giant neural networks using pipeline parallelism,

    Y . Huang, Y . Cheng, A. Bapna, O. Firat, M. X. Chen, D. Chen, H. Lee, J. Ngiam, Q. V . Le, Y . Wu, and Z. Chen, “GPipe: Efficient training of giant neural networks using pipeline parallelism,” in Proc. NeurIPS, 2019, pp. 103–112

    2019

  2. [10]

    Are we ready for autonomous driving? The KITTI vision benchmark suite,

    A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? The KITTI vision benchmark suite,” in Proc. IEEE CVPR, 2012, pp. 3354–3361

    2012

  3. [11]

    CheXpert: A large chest radiograph dataset with uncertainty labels and expert comparison,

    J. Irvin, P. Rajpurkar, M. Ko, Y . Yu, S. Ciurea-Ilcus, C. Chute, H. Marklund, B. Haghgoo, R. Ball, K. Shpanskaya, J. Seekins, D. A. Mong, S. S. Halabi, J. K. Sandberg, R. Jones, D. B. Larson, C. P. Langlotz, B. N. Patel, M. P. Lungren, and A. Y . Ng, “CheXpert: A large chest ...

    2019

  4. [12]

    On the opportunities and risks of foundation models,

    R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S. von Arx, M. S. Bernstein, J. Bohg, A. Bosselut, E. Brunskill, et al., “On the opportunities and risks of foundation models,” arXiv preprint arXiv:2108.07258, 2021

    2021 arXiv

  5. [13]

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

    N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hinton, and J. Dean, “Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,” in Proc. ICLR, 2017

    2017

  6. [14]

    MoV A: Adapting mixture of vision experts to multimodal context,

    Z. Zong, B. Ma, D. Shen, G. Song, H. Shao, D. Jiang, H. Li, and Y . Liu, “MoV A: Adapting mixture of vision experts to multimodal context,” arXiv preprint arXiv:2404.13046, 2024

    2024 arXiv

  7. [15]

    MoE-LLaV A: Mixture of experts for large vision-language models,

    B. Lin, Z. Tang, Y . Ye, J. Cui, B. Zhu, P. Jin, J. Zhang, M. Ning, L. Yuan, and W. Liu, “MoE-LLaV A: Mixture of experts for large vision-language models,” arXiv preprint arXiv:2401.15947, 2024

    2024 arXiv

  8. [16]

    Uni-MoE: Scaling Unified Multimodal LLMs with Mixture of Experts,

    Y . Li et al., “Uni-MoE: Scaling Unified Multimodal LLMs with Mixture of Experts,” arXiv (Cornell University), May 2024, doi: https://doi.org/10.48550/arxiv.2405.11273

    2024 doi arXiv

  9. [17]

    Z. Lin, C. Liu, R. Zhang, P. Gao, L. Qiu, H. Xiao, H. Qiu, C. Lin, W. Shao, K. Chen, J. Han, S. Huang, Y . Zhang, X. He, H. Li, and Y . Qiao, ”SPHINX: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models,” arXiv preprint arXiv:2311.07...

    2023 arXiv

  10. [18]

    Quantization and training of neural networks for efficient integer-arithmetic-only inference,

    B. Jacob, S. Kligys, B. Chen, M. Zhu, M. Tang, A. Howard, H. Adam, and D. Kalenichenko, “Quantization and training of neural networks for efficient integer-arithmetic-only inference,” in Proc. CVPR, 2018, pp. 2704–2713

    2018

  11. [19]

    Semantic communi- cations for future Internet: Fundamentals, applications, and challenges,

    H. Du, Y . Zhang, X. Li, Z. Ding, and H. V . Poor, “Semantic communi- cations for future Internet: Fundamentals, applications, and challenges,” IEEE Communications Surveys & Tutorials , vol. 25, no. 1, pp. 213–250, 2023

    2023

  12. [20]

    Model Context Protocol: An open standard for connecting AI assistants to the world,

    Anthropic, “Model Context Protocol: An open standard for connecting AI assistants to the world,” Anthropic Technical Report, 2024, [Online]. Available: https://www.anthropic.com/news/model-context-protocol. Ac- cessed: 2025-07-23

    2024

  13. [21]

    Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,

    T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine, “Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,” in Proc. ICML, 2018, pp. 1861–1870

    2018

  14. [22]

    Variational inference: A review for statisticians,

    D. M. Blei, A. Kucukelbir, and J. D. McAuliffe, “Variational inference: A review for statisticians,” Journal of the American Statistical Association , vol. 112, no. 518, pp. 859–877, 2017

    2017

  15. [23]

    T. S. Rappaport, Wireless Communications: Principles and Practice , 2nd ed. Upper Saddle River, NJ, USA: Prentice Hall, 2002

    2002

  16. [24]

    Goldsmith, Wireless Communications

    A. Goldsmith, Wireless Communications. Cambridge University Press, 2005

    2005

  17. [25]

    Correlation model for shadow fading in mobile radio systems,

    M. Gudmundson, “Correlation model for shadow fading in mobile radio systems,” Electronics Letters, vol. 27, no. 23, p. 2145, 1991, doi: https://doi.org/10.1049/el:19911328

    1991 doi

  18. [26]

    G. L. St ¨uber, Principles of Mobile Communication , 4th ed. Springer, 2017

    2017

  19. [27]

    Billion-scale similarity search with GPUs,

    J. Johnson, M. Douze, and H. Jegou, “Billion-scale similarity search with GPUs,” IEEE Transactions on Big Data, vol. 7, no. 3, pp. 1–1, 2019, doi: https://doi.org/10.1109/tbdata.2019.2921572

    2019

  20. [28]

    Design of coherence- aware channel indication and prediction for rate adaptation,

    Y . Du, P. Huang, Y . Shi, D. Rajan, and J. Camp, “Design of coherence- aware channel indication and prediction for rate adaptation,” EURASIP Journal on Wireless Communications and Networking, vol. 2019, no. 1, Aug. 2019, doi: https://doi.org/10.1186/s13638-019-1517-y

    2019 doi

  21. [29]

    Bayesian Forecasting and Dynamic Models,

    E. R. Ziegel, M. West, and J. Harrison, “Bayesian Forecasting and Dynamic Models,” Technometrics, vol. 39, no. 4, p. 433, Nov. 1997, doi: https://doi.org/10.2307/1271526

    1997 doi

  22. [30]

    Exact Expressions for Kullback–Leibler Divergence for Univariate Distributions,

    V . Nawa and S. Nadarajah, “Exact Expressions for Kullback–Leibler Divergence for Univariate Distributions,” Entropy, vol. 26, no. 11, p. 959, Nov. 2024, doi: https://doi.org/10.3390/e26110959

    2024 doi

  23. [31]

    MME: A comprehensive evaluation benchmark for multimodal large language models,

    C. Fu, P. Chen, Y . Shen, Y . Qin, M. Zhang, X. Lin, Z. Qiu, W. Lin, J. Yang, X. Zheng, K. Li, X. Sun, and R. Ji, “MME: A comprehensive evaluation benchmark for multimodal large language models,” arXiv preprint arXiv:2306.13394, 2023

    2023 arXiv

  24. [32]

    Learn to explain: Multimodal reasoning via thought chains for science question answering,

    P. Lu, S. Mishra, T. Xia, L. Qiu, K.-W. Chang, S.-C. Zhu, O. Tafjord, P. Clark, and A. Kalyan, “Learn to explain: Multimodal reasoning via thought chains for science question answering,” in Proc. NeurIPS, 2022, pp. 2507–2521

    2022

  25. [33]

    EdgeViT: Efficient visual modeling for edge computing,

    Z. Chen, et al., “EdgeViT: Efficient visual modeling for edge computing,” in Proc. Int. Conf. Wireless Algorithms, Syst., Appl. , Cham: Springer Nature Switzerland, 2022, pp. 1–12

    2022

  26. [34]

    DINOv2: Learning Robust Visual Features without Supervision,

    M. Oquab et al., “DINOv2: Learning Robust Visual Features without Supervision,” arXiv.org, Apr. 14, 2023. https://arxiv.org/abs/2304.07193

    2023 arXiv

  27. [35]

    Carion, F

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirchhoff, Z. Dolla ´e, R. Girshick, M. Ott, and S. J ´egou, ”End-to-end object detection with transformers,” ECCV , pp. 213–229.2020,

    2020

  28. [36]

    ”Segment anything.” Proceedings of the IEEE/CVF international conference on computer vision

    Kirillov, Alexander, et al. ”Segment anything.” Proceedings of the IEEE/CVF international conference on computer vision. 2023

    2023

  29. [37]

    ”Pix2struct: Screenshot parsing as pretraining for visual language understanding.” International Conference on Machine Learning

    Lee, Kenton, et al. ”Pix2struct: Screenshot parsing as pretraining for visual language understanding.” International Conference on Machine Learning. PMLR, 2023

    2023

  30. [38]

    DePlot: One-shot visual language reasoning by plot-to-table translation,

    F. Liu et al., “DePlot: One-shot visual language reasoning by plot-to-table translation,” arXiv (Cornell University), Jan. 2023, doi: https://doi.org/10.18653/v1/2023.findings-acl.660

    2023 doi

  31. [39]

    H. Wei, et al., ”Vary: Scaling up the vision vocabulary for large vision - language model,” in European Conference on Computer Vision, Cham, Switzerland: Springer Nature Switzerland, 2024

    2024

  32. [40]

    Zhang, Y

    S. Zhang, Y . Xu, N. Usuyama, et al., ”Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image - text pairs,” arXiv preprint arXiv:2303.00915, 2023

    2023 arXiv

  33. [41]

    MRKL systems: A modular, neuro-symbolic architecture that combines large language models, external knowledge sources and discrete reasoning,

    E. Karpas et al., “MRKL systems: A modular, neuro-symbolic architecture that combines large language models, external knowledge sources and discrete reasoning,” arXiv preprint arXiv:2205.00445 , 2022

    2022 arXiv

  34. [42]

    Resource manage- ment with deep reinforcement learning,

    H. Mao, M. Alizadeh, I. Menache, and S. Kandula, “Resource manage- ment with deep reinforcement learning,” in Proc. ACM HotNets, 2016, pp. 50–56

    2016

  35. [43]

    Human-level control through deep reinforcement learning,

    V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran, D. Wierstra, S. Legg, and D. Hassabis, “Human-level control through deep...

  36. [44]

    Adaptive computation time for recurrent neural networks,

    A. Graves, “Adaptive computation time for recurrent neural networks,” in Proc. ICML, 2016, pp. 3500–3509

    2016

Pith tools

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