Pith. sign in

REVIEW 4 major objections 7 minor 45 references

MORES turns latent LLM reasoning into a wireless service: devices and edge servers share recurrent steps, and a semantic MoE DRL agent lifts system throughput about 18% over plain SAC.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-10 12:48 UTC pith:HJFXO3SJ

load-bearing objection Competent edge-systems packaging of latent recurrent reasoning with a clear simulated MoE-SAC gain; the 18% number rides on offline accuracy surfaces the online loop never re-runs end-to-end. the 4 major comments →

arxiv 2607.08116 v1 pith:HJFXO3SJ submitted 2026-07-09 cs.NI

MORES: Mobile Reasoning-as-a-Service via Distributed LLM Inference-Time Scaling

Guanchen Liu , Hongyang Du , Kaibin Huang This is my paper
classification cs.NI
keywords large language modelsinference-time scalinglatent reasoningmobile edge computingdeep reinforcement learningmixture of expertswireless networksdistributed inference
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Large language models can reason better at inference time by running more computation rather than growing parameters, but both explicit chain-of-thought and implicit latent recursion are too heavy for phones and other edge devices. MORES treats that extra reasoning as a service over the wireless network. It focuses on implicit latent reasoning: a device runs the encoder, a few local recurrent updates, and the decoder, while an edge server continues the same recurrent unit on the intermediate latent state when deeper thinking is needed. Intermediate features can be pruned before uplink so communication stays cheap. How many recurrent steps to request and how hard to prune become a joint scheduling decision under device energy, server recurrence budget, and per-token latency caps. The paper solves that decision with a Soft Actor-Critic agent whose actor and critic use a semantic Mixture-of-Experts design: a lightweight router classifies the query type, deterministic experts specialize by task and channel, and a shared expert carries common knowledge. On a mix of math, code, and commonsense benchmarks with three channel qualities, that agent raises average correct-answer throughput by roughly 18% over conventional SAC under the same budgets.

Core claim

The paper claims that recursive latent reasoning can be cleanly partitioned across a wireless device–server link, and that a semantic MoE-based SAC agent that jointly chooses the total recurrent depth and the uplink pruning rate can exploit task and channel heterogeneity to raise long-run system throughput by about 18% relative to ordinary SAC under device energy, server recurrence, and latency constraints.

What carries the argument

Distributed latent recursion plus a semantic MoE-SAC controller: the device applies the prelude encoder and a fixed number of local recurrent steps, ships the (optionally pruned) latent representation and state, the server continues the same recurrent unit, and the device decodes; the controller’s semantic router and deterministic expert branches set the total recurrent steps r and pruning rate ρ per request.

Load-bearing premise

That you can throw away a non-trivial fraction of the latent features on the uplink and still leave enough semantic state for the server’s continued recurrence to produce accurate answers at the rates the offline accuracy curves suggest.

What would settle it

Run the same recurrent-depth model end-to-end on a real wireless link with the paper’s pruning rates (up to ~8–15%) and quantization; if measured answer accuracy falls far below the offline GSM8K/MBPP/HellaSwag curves in Figs. 7–9, or if the learned (r, ρ) policy no longer beats SAC on throughput, the joint scheduler’s claimed trade-off fails.

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

If this is right

  • Edge devices can request adjustable reasoning depth from a server without shipping full chain-of-thought tokens, cutting both communication volume and explicit intermediate output.
  • Task-type and channel-aware expert routing lets a single DRL policy specialize resource choices for math, code, and commonsense queries under poor, fair, and good channels.
  • System throughput becomes a schedulable quantity under explicit device-energy and server-recurrence budgets rather than an uncontrolled byproduct of fixed model depth.
  • The same partitioning idea extends, in principle, to any recurrent latent reasoning stack whose intermediate state is compact enough to prune and transmit.

Where Pith is reading between the lines

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

  • If latent pruning remains faithful under real channel noise and quantization, MORES-style services become a practical alternative to shipping long explicit CoT sequences over cellular or Wi-Fi links.
  • The same semantic-router + shared-expert pattern could be reused for multimodal or multi-user edge inference where task mix and radio conditions vary even faster than in the paper’s three-benchmark, three-channel setting.
  • Once recurrence budgets are priced, operators could treat reasoning depth as a billable network resource rather than only as a model hyperparameter.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes MORES, a device–edge cooperative framework for implicit (latent) LLM reasoning over wireless networks. Latent recurrent updates are partitioned so devices run a prelude encoder, a fixed number of local recurrent steps re, and a coda decoder, while servers continue remaining recurrent steps; intermediate latent features may be uplink-pruned at rate ρ. The authors formulate a joint scheduling problem (P1) that maximizes average answer correctness U under device energy, server recurrence budget, and per-token TBT latency constraints, with decision variables (rk, ρk). They solve it with a semantic MoE-augmented SAC agent whose experts are deterministically gated by task type (via a semantic router) and scenario, and report roughly 18% higher U than plain SAC (e.g., U≈33 vs 28 at E_dev_max=80, G_max=3×10^5) on a composite of GSM8K, MBPP, and HellaSwag under a simulated wireless model.

Significance. If the evaluation of δk under pruning and wireless transfer is faithful, the work is a solid systems contribution at the intersection of inference-time scaling and edge–cloud wireless inference. Partitioning recursive latent reasoning (rather than explicit CoT tokens) is a natural fit for communication-efficient cooperative inference, and the joint (r, ρ) control problem is well motivated by the measured task heterogeneity in Figs. 6–9. The semantic MoE-SAC design, code release, and systematic sweeps over energy, recurrence budget, and re are concrete strengths. The significance is primarily systems/algorithmic rather than a new learning-theoretic result; it would be of interest to the wireless/edge-AI community provided the simulator-to-model fidelity gap is closed.

major comments (4)
  1. The central 18% throughput claim (Figs. 10–12; abstract) rests on U = (1/K) Σ δk (Eq. 15) with reward ut = δt. Figs. 7–9 report offline accuracy vs (r, ρ) on GSM8K/MBPP/HellaSwag, but §5 never states whether the DRL environment obtains δk by re-running the same recurrent-depth backbone end-to-end on the actually pruned, quantized (q=16) uplink features under the wireless model of Eqs. (11)–(14), or by table lookup from those offline curves. If the latter, the agent optimizes a surrogate that may systematically over-value high-ρ / low-r actions; the reported gain is then a property of the simulator, not of the latent service. Please clarify the online δk pipeline and, if lookup is used, re-evaluate with true pruned-latent inference (or ablate the gap).
  2. MDP state (Eq. 20) is st = {t, G_res,t, E_dev_res,t} and §3.4 states that channel gain hk is unknown to device and server. Yet the MoE experts and Fig. 13 are conditioned on task–channel scenarios (poor/fair/good), and the semantic router (§4.2) injects task type. How channel (and task) enter the policy at decision time is not specified in the MDP; if they are side information to the router/gating only, the formulation should say so and justify that this does not contradict the “unknown channel” assumption. Without this, expert specialization in Fig. 13 is hard to reconcile with the stated observation model.
  3. Uplink pruning (§3.4, Eq. 11) is defined only as a scalar fraction ρk of “latent information” removed before transmission. For server-side continuation zi = R(e, zi−1) (Eq. 2 / Alg. 1) to remain well-defined, the paper must specify which coordinates are dropped (structured vs random), how the server reconstructs or masks the incomplete (e, z_re), and whether the offline accuracy surfaces in Figs. 7–9 use that same operator. Without a concrete pruning operator tied to the recurrent unit, the accuracy–cost trade-off that (P1) and the MoE agent exploit is under-specified.
  4. Baselines in §5 are limited to conventional SAC. For a joint discrete (r, ρ) scheduling claim, at least one non-RL policy (fixed (r, ρ), task-conditional greedy from Figs. 7–8, or myopic energy/latency heuristic) is needed to show that the 18% is not merely “SAC vs weaker SAC.” Adding these would substantially strengthen the load-bearing experimental claim.
minor comments (7)
  1. “Throughput” U is average correctness, not requests per unit time. The term is defined in Eq. (15) but remains easy to misread in the abstract and §5; consider “average success rate” or “correctness throughput” in titles of Figs. 10–12.
  2. Action grids (§5.1): r ∈ {8,…,32}, ρ ∈ {0,…,8%} (and offline curves go to 15%). State that the DRL action set is discrete and whether continuous SAC is projected onto this grid.
  3. Complexity analysis (Eqs. 27–28) claims ~2× cost from MoE; useful, but report wall-clock actor/critic step time vs SAC to confirm the asymptotic claim under the actual network sizes used.
  4. Fig. 3 (token-wise convergence depth) is suggestive but not quantitatively linked to the choice of re or the action set; a short caption note on how it motivates re would help.
  5. Notation: αk,l is introduced for tokens processed per round, then used in energy/latency; a one-line reminder that only the first round processes nk tokens would reduce reader load in §3.3–3.4.
  6. Related work on latent/recurrent-depth reasoning and edge split inference is adequate; a brief pointer to any prior work on feature pruning of intermediate LLM activations (beyond [27],[28]) would round out §2.2.
  7. Typos/style: “T o address” / “T o optimize” (abstract/intro spacing), “Soft Actor-Critic (SAC) algorithm” repeated; “Reasoning-as-aService” missing hyphen in abstract title line.

Circularity Check

0 steps flagged

No significant circularity; the ~18% throughput gain is an empirical DRL comparison under an explicitly formulated MDP, not forced by definition, fit, or self-citation.

full rationale

The paper's central claim is experimental: under problem (P1) maximizing average correctness U = (1/K) sum δ_k subject to device energy, server recurrence budget G_max and TBT latency, a semantic MoE-SAC agent that chooses recurrent steps r_k and pruning rate ρ_k yields ~18% higher U than plain SAC (e.g., U≈33 vs 28 at E_dev_max=80, G_max=3e5). U is defined directly from the binary correctness indicator δ (Eq. 15) and is measured in simulation; the reward of the SAC agent is likewise set to u_t = δ_t. Offline accuracy-vs-(r,ρ) curves (Figs. 7–9) supply the environment's δ mapping, but this is ordinary simulator construction, not a tautological reduction of the reported gain. There is no uniqueness theorem, ansatz, or fitted parameter that is later re-labeled a prediction; self-citations appear only in the related-work survey and do not underwrite the numerical claim. The derivation chain (latent recurrence partition → joint scheduling MDP → MoE-SAC policy → measured throughput) is therefore self-contained against its own experimental benchmarks and exhibits none of the six circularity patterns.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 2 invented entities

This is a systems/simulation paper. The central claim rests on modeling choices (partitionable latent recurrence, pruning fidelity, linear compute latency, block-fading rates, fixed power/bandwidth), many hand-set numerical budgets and costs, and the architectural postulate that deterministic task-semantic expert routing improves SAC under heterogeneous tasks and channels. No physical free constants are fitted to nature; free parameters are engineering knobs that shape the reported 18% gap.

free parameters (6)
  • c1, c2 (compute latency coefficients)
    T_comp = c1·r·α + c2 with c1=5e-4, c2=0 chosen for the simulator; they directly set the latency–depth trade-off in (P1).
  • Ep, Er, Ec (per-token device energy costs)
    Set to 2.5e-4, 3e-4, 2.5e-4 J; determine device energy constraint bindingness and thus feasible (r,ρ).
  • Tmax, Gmax, E_dev_max
    Hard budgets (e.g., Tmax=0.025, Gmax in {2e5,3e5,4e5}, E_dev_max in {60,80,100}) define the feasible region where throughput is measured.
  • Discrete action grids for r and ρ
    r∈{8..32}, ρ∈{0..8%} (and offline curves to 15%) are hand-chosen; the policy can only optimize inside this grid.
  • Channel gain set {0.4, 0.7, 1.2} and η, powers, bandwidths
    Wireless rates and energy follow from these fixed simulation choices; heterogeneity is synthetic, not measured traces.
  • SAC entropy temperature τ and MoE architecture sizes
    Training dynamics and expert capacity depend on unreported or lightly reported hyperparameters that affect the 18% gap.
axioms (6)
  • domain assumption Latent recurrent reasoning of the form z_i = R(e, z_{i-1}) can be split after r_e device steps and continued on a server with only (e, z_re) exchanged.
    Core of Algorithm 1 and §3.2; taken from latent/recurrent-depth literature and treated as operationally true.
  • domain assumption Uplink pruning of a fraction ρ of latent information plus q-bit quantization preserves enough semantics for continued reasoning, with accuracy as in Figs. 7–9.
    Enters communication model §3.4 and the joint action space; without it, ρ is not a free accuracy–cost knob.
  • domain assumption Block-fading Shannon rates with fixed power/bandwidth and unknown h_k adequately model the wireless cost of latent exchange.
    Eqs. (7)–(14); standard MEC simplification used as ground truth for scheduling.
  • ad hoc to paper Average correctness U under energy, recurrence, and TBT constraints is the right long-term objective for “system throughput.”
    Problem (P1); equates reasoning quality with binary δ_k averaged over K=100 mixed requests.
  • ad hoc to paper Deterministic semantic task routing into MoE experts plus a shared expert improves SAC under task/channel heterogeneity.
    §4.2–4.3 architectural claim validated only by the paper’s own training curves and Table 2.
  • standard math Standard MDP/SAC theory (entropy-regularized policy improvement, soft Bellman updates).
    §4.1 cites Haarnoja et al.; used as the learning backbone.
invented entities (2)
  • MORES (Mobile Reasoning-as-a-Service) cooperative latent-reasoning service no independent evidence
    purpose: Name and system abstraction for on-demand server continuation of device-side latent recurrence over wireless links.
    Framework label for the split encoder/recurrent/decoder deployment; evidence is only the authors’ simulator and code.
  • Semantic MoE-SAC with deterministic task/scenario gating and shared-expert fusion no independent evidence
    purpose: Specialize actor/critic feature extractors by task while fusing shared features for joint (r,ρ) decisions.
    Architectural construct in §4.3; performance claims rest on internal ablations (Figs. 10–13, Table 2), not external replications.

pith-pipeline@v1.1.0-grok45 · 25033 in / 3972 out tokens · 51521 ms · 2026-07-10T12:48:42.374531+00:00 · methodology

0 comments
read the original abstract

Inference-time scaling has emerged as an effective approach for enhancing the capabilities of Large Language Models (LLMs), addressing the growing demand for stronger reasoning without increasing model size. This novel form of LLM scaling comprises two representative approaches: explicit reasoning, which generates intermediate chain-of-thought tokens during an explicit thinking phase, and implicit reasoning, which iteratively updates hidden states in the latent space without producing explicit outputs. Despite their effectiveness, both paradigms incur substantial computational and memory overhead, raising challenges for deployment on resource-constrained edge devices. To address these issues, we propose a Mobile Reasoning-as-aService (MORES) framework that treats reasoning as a computational service accessible to edge devices over wireless networks. Focusing on implicit reasoning, we leverage its recursive structure to partition hiddenstate updates between edge devices and servers, enabling cooperative inference that allows devices to access additional cloud computation on demand. To optimize long-term performance, we formulate a joint computation and communication scheduling problem and solve it using a semantic Mixture-of-Experts (MoE)-based Deep Reinforcement Learning (DRL) algorithm to address heterogeneity in wireless conditions and task demands. The agent adaptively allocates resources by adjusting the number of recurrent steps and the transmission pruning rate, while a semantic router enables high-speed gating for real-time expert selection. Experimental results show that the proposed method achieves an approximately 18% improvement in system throughput over the baseline Soft Actor-Critic (SAC) algorithm. Our code is available at https://github.com/NICE-HKU/MORES.

Figures

Figures reproduced from arXiv: 2607.08116 by Guanchen Liu, Hongyang Du, Kaibin Huang.

Figure 1
Figure 1. Figure 1: Comparison of reasoning modes between device [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: System model for distributed latent LLM reasoning. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Token-wise reasoning depth required for full conver [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of the semantic routing process, where the [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Histogram of token count distributions of the GSM8K, MBPP, and HellaSwag datasets. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Accuracy vs. pruning rate under different recurrent steps. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Accuracy vs. recurrent steps under different pruning rates. [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Example GSM8K problem illustrating the effect of recurrent steps ( [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Throughput over training epochs for MoE and baseline DRL models across different device energy budgets. [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Throughput over training epochs for MoE and baseline DRL models across different server recurrence budgets. [PITH_FULL_IMAGE:figures/full_fig_p012_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Throughput over training epochs for MoE and baseline DRL models across different initial edge iterations. [PITH_FULL_IMAGE:figures/full_fig_p012_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Expert decision distributions across distinct task and channel conditions. [PITH_FULL_IMAGE:figures/full_fig_p013_13.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

45 extracted references · 45 canonical work pages · 14 internal anchors

  1. [1]

    Language models are few-shot learners,

    T. Brown, B. Mann, N. Ryderet al., “Language models are few-shot learners,” inProc. Adv. Neural Inf. Process. Syst., vol. 33, 2020, pp. 1877–1901. 14

    work page 2020

  2. [2]

    OpenAI GPT-5 System Card

    A. Singh, A. Fry, A. Perelmanet al., “OpenAI GPT-5 system card,” arXiv preprint arXiv:2601.03267, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  3. [3]

    DeepSeek-V4: Towards highly efficient million-token context intelligence,

    DeepSeek-AI, “DeepSeek-V4: Towards highly efficient million-token context intelligence,” [Online]. Available: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro, 2026

    work page 2026

  4. [4]

    A new era of intelligence with Gemini 3,

    Google, “A new era of intelligence with Gemini 3,” [Online]. Available: https://blog.google/products/gemini/gemini-3, 2025

    work page 2025

  5. [5]

    Large language model (LLM) for telecommunications: A comprehensive survey on principles, key techniques, and opportunities,

    H. Zhou, C. Hu, Y. Yuanet al., “Large language model (LLM) for telecommunications: A comprehensive survey on principles, key techniques, and opportunities,”IEEE Commun. Surveys T uts., vol. 27, no. 3, pp. 1955–2005, 2024

    work page 1955

  6. [6]

    An empirical analysis of compute-optimal large language model training,

    J. Hoffmann, S. Borgeaud, A. Menschet al., “An empirical analysis of compute-optimal large language model training,” inProc. Adv. Neural Inf. Process. Syst., vol. 35, 2022, pp. 30 016–30 030

    work page 2022

  7. [7]

    Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters

    C. Snell, J. Lee, K. Xu, and A. Kumar, “Scaling LLM test-time compute optimally can be more effective than scaling model parameters,”arXiv preprint arXiv:2408.03314, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  8. [8]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    D. Guo, D. Yang, H. Zhang, J. Song, P . Wang, Q. Zhu, R. Xu, R. Zhang, S. Ma, X. Biet al., “DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning,”arXiv preprint arXiv:2501.12948, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  9. [9]

    QwQ-32B: Embracing the power of reinforcement learning,

    Qwen Team, “QwQ-32B: Embracing the power of reinforcement learning,” [Online]. Available: https://qwenlm.github.io/blog/ qwq-32b, 2025

    work page 2025

  10. [10]

    OpenAI o1 System Card

    A. Jaech, A. Kalai, A. Lereret al., “OpenAI o1 system card,”arXiv preprint arXiv:2412.16720, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  11. [11]

    Scaling up test-time compute with latent reasoning: A recurrent depth ap- proach,

    J. Geiping, S. McLeish, N. Jain, J. Kirchenbauer, S. Singh, B. Bar- toldson, B. Kailkhura, A. Bhatele, and T. Goldstein, “Scaling up test-time compute with latent reasoning: A recurrent depth ap- proach,” inProc. Adv. Neural Inf. Process. Syst., vol. 38, 2026, pp. 41 340–41 391

    work page 2026

  12. [12]

    A Survey on Latent Reasoning

    R.-J. Zhu, T. Peng, T. Chenget al., “A survey on latent reasoning,” arXiv preprint arXiv:2507.06203, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  13. [13]

    Distilling step-by-step! Out- performing larger language models with less training data and smaller model sizes,

    C.-Y. Hsieh, C.-L. Li, C.-K. Yeh, H. Nakhost, Y. Fujii, A. Ratner, R. Krishna, C.-Y. Lee, and T. Pfister, “Distilling step-by-step! Out- performing larger language models with less training data and smaller model sizes,” inProc. Findings Assoc. Comput. Linguist., 2023, pp. 8003–8017

    work page 2023

  14. [14]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V . Le, D. Zhouet al., “Chain-of-thought prompting elicits reasoning in large language models,” inProc. Adv. Neural Inf. Process. Syst., vol. 35, 2022, pp. 24 824–24 837

    work page 2022

  15. [15]

    Large language models are zero-shot reasoners,

    T. Kojima, S. S. Gu, M. Reid, Y. Matsuo, and Y. Iwasawa, “Large language models are zero-shot reasoners,” inProc. Adv. Neural Inf. Process. Syst., vol. 35, 2022, pp. 22 199–22 213

    work page 2022

  16. [16]

    Reasoning or not? A comprehensive evaluation of reasoning LLMs for dialogue summarization,

    K. Jin, Y. Wang, L. Santos, T. Fang, X. Yang, S. K. Im, and H. G. Oliveira, “Reasoning or not? A comprehensive evaluation of reasoning LLMs for dialogue summarization,”Expert Syst. Appl., p. 129831, 2025

    work page 2025

  17. [17]

    Implicit Chain of Thought Reasoning via Knowledge Distillation

    Y. Deng, K. Prasad, R. Fernandez, P . Smolensky, V . Chaudhary, and S. Shieber, “Implicit chain of thought reasoning via knowledge distillation,”arXiv preprint arXiv:2311.01460, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  18. [18]

    Training Large Language Models to Reason in a Continuous Latent Space

    S. Hao, S. Sukhbaatar, D. Su, X. Li, Z. Hu, J. Weston, and Y. Tian, “Training large language models to reason in a continuous latent space,”arXiv preprint arXiv:2412.06769, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  19. [19]

    A survey on mobile edge computing: The communication perspective,

    Y. Mao, C. You, J. Zhang, K. Huang, and K. B. Letaief, “A survey on mobile edge computing: The communication perspective,”IEEE Commun. Surveys T uts., vol. 19, no. 4, pp. 2322–2358, 2017

    work page 2017

  20. [20]

    Federated learning of large language models with parameter-efficient prompt tuning and adaptive optimization,

    T. Che, J. Liu, Y. Zhou, J. Ren, J. Zhou, V . Sheng, H. Dai, and D. Dou, “Federated learning of large language models with parameter-efficient prompt tuning and adaptive optimization,” in Proc. Conf. Empirical Methods Nat. Lang. Process., 2023, pp. 7871– 7888

    work page 2023

  21. [21]

    OpenFedLLM: Training large language models on decentralized private data via federated learning,

    R. Ye, W. Wang, J. Chai, D. Li, Z. Li, Y. Xu, Y. Du, Y. Wang, and S. Chen, “OpenFedLLM: Training large language models on decentralized private data via federated learning,” inProc. ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining, 2024, pp. 6137– 6147

    work page 2024

  22. [22]

    SplitLLM: Hierarchical split learning for large language models over wireless network,

    S. Zhang, G. Cheng, Z. Li, and W. Wu, “SplitLLM: Hierarchical split learning for large language models over wireless network,” inProc. IEEE Global Commun. Conf. Workshops, 2024, pp. 1–6

    work page 2024

  23. [23]

    SplitLoRA: A Split Parameter-Efficient Fine-Tuning Framework for Large Language Models

    Z. Lin, X. Hu, Y. Zhang, Z. Chen, Z. Fang, X. Chen, A. Li, P . Vepakomma, and Y. Gao, “SplitLoRA: A split parameter- efficient fine-tuning framework for large language models,”arXiv preprint arXiv:2407.00952, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  24. [24]

    FedsLLM: Federated split learning for large language models over commu- nication networks,

    K. Zhao, Z. Yang, C. Huang, X. Chen, and Z. Zhang, “FedsLLM: Federated split learning for large language models over commu- nication networks,” inProc. Int. Conf. Ubiquitous Commun., 2024, pp. 438–443

    work page 2024

  25. [25]

    EdgeShard: Efficient LLM inference via collaborative edge computing,

    M. Zhang, X. Shen, J. Cao, Z. Cui, and S. Jiang, “EdgeShard: Efficient LLM inference via collaborative edge computing,”IEEE Internet Things J., vol. 12, no. 10, pp. 13 119–13 131, 2024

    work page 2024

  26. [26]

    CE-CoLLM: Efficient and adaptive large lan- guage models through cloud-edge collaboration,

    H. Jin and Y. Wu, “CE-CoLLM: Efficient and adaptive large lan- guage models through cloud-edge collaboration,” inProc. IEEE Int. Conf. Web Services, 2025, pp. 316–323

    work page 2025

  27. [27]

    DILEMMA: Joint LLM Quantization and Distributed LLM Inference Over Edge Computing Systems

    M. Hosseinzadeh and H. Khamfroush, “DILEMMA: Joint LLM quantization and distributed LLM inference over edge computing systems,”arXiv preprint arXiv:2503.01704, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  28. [28]

    The larger the merrier? Efficient large AI model inference in wireless edge networks,

    Z. Lyu, M. Xiao, J. Xu, M. Skoglund, and M. Di Renzo, “The larger the merrier? Efficient large AI model inference in wireless edge networks,”IEEE J. Sel. Areas Commun., 2025

    work page 2025

  29. [29]

    Mixture-of- experts for distributed edge computing with channel-aware gating function,

    Q. Song, S. Jing, S. Zhang, S. Zhang, and C. Huang, “Mixture-of- experts for distributed edge computing with channel-aware gating function,” inProc. IEEE Int. Conf. Commun. Workshops, 2025, pp. 1353–1358

    work page 2025

  30. [30]

    Theory of mixture-of-experts for mobile edge computing,

    H. Li and L. Duan, “Theory of mixture-of-experts for mobile edge computing,” inProc. IEEE INFOCOM, 2025, pp. 1–10

    work page 2025

  31. [31]

    WDMoE: Wireless distributed mixture of experts for large language models,

    N. Xue, Y. Sun, Z. Chen, M. Tao, X. Xu, L. Qian, S. Cui, W. Zhang, and P . Zhang, “WDMoE: Wireless distributed mixture of experts for large language models,”IEEE T rans. Wireless Commun., 2025

    work page 2025

  32. [32]

    Lyapunov-guided deep reinforcement learning for Stable online computation of- floading in mobile-edge computing networks,

    S. Bi, L. Huang, H. Wang, and Y.-J. A. Zhang, “Lyapunov-guided deep reinforcement learning for Stable online computation of- floading in mobile-edge computing networks,”IEEE T rans. Wire- less Commun., vol. 20, no. 11, pp. 7519–7537, 2021

    work page 2021

  33. [33]

    Offloading and resource alloca- tion with general task graph in mobile edge computing: A deep reinforcement learning approach,

    J. Yan, S. Bi, and Y. J. A. Zhang, “Offloading and resource alloca- tion with general task graph in mobile edge computing: A deep reinforcement learning approach,”IEEE T rans. Wireless Commun., vol. 19, no. 8, pp. 5404–5419, 2020

    work page 2020

  34. [34]

    Meta- reinforcement learning with mixture of experts for generalizable multi access in heterogeneous wireless networks,

    Z. Liu, X. Wang, C. Feng, X. Sun, W. Zhan, and X. Chen, “Meta- reinforcement learning with mixture of experts for generalizable multi access in heterogeneous wireless networks,”IEEE T rans. Commun., vol. 74, pp. 870–885, 2025

    work page 2025

  35. [35]

    Mixture of Experts for Network Optimization: A Large Language Model-enabled Approach

    H. Du, G. Liu, Y. Lin, D. Niyato, J. Kang, Z. Xiong, and D. I. Kim, “Mixture of experts for network optimization: A large language model-enabled approach,”arXiv preprint arXiv:2402.09756, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  36. [36]

    Mixture of Experts in a Mixture of RL settings

    T. Willi, J. Obando-Ceron, J. Foerster, K. Dziugaite, and P . S. Castro, “Mixture of experts in a mixture of RL settings,”arXiv preprint arXiv:2406.18420, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  37. [37]

    Toward scalable generative AI via mixture of experts in mobile edge networks,

    J. Wang, H. Du, D. Niyato, J. Kang, Z. Xiong, D. I. Kim, and K. B. Letaief, “Toward scalable generative AI via mixture of experts in mobile edge networks,”IEEE Wireless Commun., 2024

    work page 2024

  38. [38]

    Orca: A distributed serving system for Transformer-based generative models,

    G.-I. Yu, J. S. Jeong, G.-W. Kim, S. Kim, and B.-G. Chun, “Orca: A distributed serving system for Transformer-based generative models,” inProc. USENIX OSDI, 2022, pp. 521–538

    work page 2022

  39. [39]

    R. S. Sutton and A. G. Barto,Reinforcement learning: An introduction. MIT press Cambridge, 1998

    work page 1998

  40. [40]

    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,” inProc. Int. Conf. Mach. Learn. (ICML), 2018, pp. 1861–1870

    work page 2018

  41. [41]

    Semantic routing for enhanced performance of LLM-assisted intent-based 5G core network management and orchestration,

    D. M. Manias, A. Chouman, and A. Shami, “Semantic routing for enhanced performance of LLM-assisted intent-based 5G core network management and orchestration,” inProc. IEEE Global Commun. Conf. (GLOBECOM), 2024, pp. 2924–2929

    work page 2024

  42. [42]

    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,”arXiv preprint arXiv:1701.06538, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  43. [43]

    Training Verifiers to Solve Math Word Problems

    K. Cobbe, V . Kosaraju, M. Bavarian, M. Chen, H. Jun, L. Kaiser, M. Plappert, J. Tworek, J. Hilton, R. Nakano, C. Hesse, and J. Schulman, “Training verifiers to solve math word problems,” arXiv preprint arXiv:2110.14168, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  44. [44]

    Program Synthesis with Large Language Models

    J. Austin, A. Odena, M. Nye, M. Bosma, H. Michalewski, D. Do- han, E. Jiang, C. Cai, M. Terry, Q. Leet al., “Program synthesis with large language models,”arXiv preprint arXiv:2108.07732, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  45. [45]

    Hel- laSwag: Can a machine really finish your sentence?

    R. Zellers, A. Holtzman, Y. Bisk, A. Farhadi, and Y. Choi, “Hel- laSwag: Can a machine really finish your sentence?” inProc. ACL, 2019, pp. 4791–4800

    work page 2019