Pith. sign in

REVIEW 5 major objections 7 minor 50 references

AirLLM: Diffusion Policy-based Adaptive LoRA for Remote Fine-Tuning of LLM over the Air

T0 review · 5 major / 7 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 AirLLM claims that treating LoRA rank allocation as a wireless-aware sequential decision problem and solving it with a PPO-plus-diffusion hierarchy cuts transmitted parameters by up to 12.5% while improving task accuracy by 0.69% over…

desk verdict A promising but under-specified RL-plus-diffusion approach to LoRA rank allocation; the core idea deserves review, but the empirical claims need major revision. read the letter →

arxiv 2507.11515 v1 pith:FTRYG7DG submitted 2025-07-15 cs.LG cs.AIcs.CL

Shiyi Yang , Xiaoxue Yu , Rongpeng Li , Jianhang Zhu , Zhifeng Zhao , Honggang Zhang This is my paper · ORCID
classification cs.LGcs.AIcs.CL
keywords remotefine-tuninglow-rankadaptationdiffusionpolicyreinforcementlearningrankallocationwirelesscommunicationparameter-efficientPPO
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

AirLLM tries to establish that the per-layer rank budgets of low-rank adaptation (LoRA) can be learned as a wireless-aware sequential decision problem, instead of being set by fixed or heuristic rules. The paper models the rank configuration as a structured vector over all LoRA-inserted projections, and solves the resulting high-dimensional decision problem with a two-level policy: a PPO (Proximal Policy Optimization) agent issues a coarse rank prior from channel and data states, and a DDIM (Denoising Diffusion Implicit Model) denoising module refines it into per-layer, per-projection ranks. The reward combines the fine-tuning loss with the communication cost of transmitting the rank-sized updates, so the policy is trained to trade accuracy against bandwidth. On an OPT-1.3B model fine-tuned on SST-2 over simulated AWGN channels, AirLLM reports up to 0.69% higher accuracy than the AdaLoRA baseline while transmitting up to 12.5% fewer parameters, with 20-30% faster convergence than vanilla PPO. The central bet is that hierarchical coarse-to-fine policy generation makes high-dimensional rank allocation tractable where plain PPO becomes unstable.

What carries the argument

The load-bearing object is the rank vector rt = {r(t)_l,m} over L layers and six projection modules per layer, tied to the SVD parameterization W(t)_l,m = W(0)_l,m + P(t)_l,m Λ(t)_l,m Q(t)_l,m, so that each rank value directly sets both trainable parameters and transmitted bytes. The paper converts rank allocation from a combinatorial search into a continuous denoising problem: PPO outputs a coarse latent prior aout_t, and a conditional DDIM iteratively denoises a Gaussian latent into a clean rank vector, which is then rounded and clipped into {0,...,rmax}^{6L}. The denoiser is trained with a hybrid loss combining noise-reconstruction error with policy reward, under classifier-free guidance. This machinery is what makes high-dimensional, structured actions tractable.

What would settle it

Run AirLLM's training loop but freeze each chosen rank configuration and fine-tune it to convergence; if the low-cost loss used in Eq. (10) does not predict converged accuracy, or if evaluating that loss is so expensive that it reflects an older configuration, the reward signal is invalid and the reported gains would not transfer to a real deployment.

Watch

Extended reading notes

Core claim

The paper's central claim is that communication-efficient remote fine-tuning can be formulated as a Markov decision process over rank vectors, and that a hierarchical PPO-DDIM policy solves it better than either component alone or than heuristic baselines. The state includes SNR, bandwidth, lexical entropy, and out-of-vocabulary rate; the action is the next rank vector across all six LoRA-inserted projections of every layer; the reward is negative task loss minus a weighted transmission cost. The two policy levels are trained alternately, with DDIM using classifier-free guidance and a hybrid denoising-plus-reward objective so that the refined actions stay aligned with PPO rewards. The reported results place AirLLM above AdaLoRA on both objectives in the tested setup: higher accuracy (up to 0.69%) and smaller transmitted parameter count (up to 12.5%), with faster convergence and more stable behavior under varying SNR than vanilla PPO.

Load-bearing premise

The reward signal in Eq. (10) assumes the task loss can be evaluated cheaply and faithfully right after each rank change, so that the rank policy receives a signal that actually reflects the configuration it chose.

Editorial extensions

If this is right

  • If the central claim holds, rank allocation for remote fine-tuning no longer needs hand-tuned layer-importance heuristics; it can be learned end-to-end from channel and data signals.
  • Because the action is a structured vector, the same hierarchical policy should transfer to larger models by scaling the denoising network, not by redesigning the decision problem.
  • The reward coupling means the policy can reallocate ranks on the fly as SNR fluctuates, which is exactly the regime of cloud-to-edge transmission.
  • The reported convergence gain over vanilla PPO suggests that diffusion refinement also reduces the sample complexity of learning high-dimensional rank policies.

Reading between the lines

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

  • A stricter communication accounting would include the bytes needed to send the learned policy itself and any per-step rank vectors; the paper counts only the LoRA update parameters, so the 12.5% saving is an upper bound on end-to-end transmission savings.
  • Because the state already carries lexical entropy and OOV rate, AirLLM naturally extends to curriculum-style training, where rank budgets are adjusted as the data distribution shifts during fine-tuning.
  • The same coarse-to-fine action generation could be applied to other structured compression choices, such as quantization bit-widths, pruning masks, or mixture-of-experts routing, where the action is a high-dimensional structured vector.
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 / 7 minor

Summary. The paper proposes AirLLM, a hierarchical reinforcement-learning framework for adaptive LoRA rank allocation in cloud-to-edge remote fine-tuning of LLMs. It formulates rank adaptation as an MDP whose state includes SNR, bandwidth, lexical entropy, OOV rate, and current ranks; a PPO actor produces coarse actions that a conditional DDIM refines into discrete rank vectors. The reward in Eq. (10) combines a fine-tuning loss surrogate with a transmission-cost term. Experiments on SST-2 with OPT-1.3B compare AirLLM with AdaLoRA and with PPO-only/DDIM-only ablations under varying SNR and rank budgets, reporting up to 0.69% higher accuracy and 12.5% fewer transmitted parameters at rmax=64, plus a claimed 20-30% training-efficiency gain over vanilla PPO.

Significance. If the underlying RL environment is correctly specified, AirLLM is a plausible and well-motivated mechanism for communication-aware PEFT: it makes channel state and data complexity part of the rank-allocation decision, combines PPO's stability with DDIM's high-dimensional action modeling, and provides explicit complexity analyses. The paper also includes useful ablations over PPO/DDIM backbones, prediction types, reward coefficients, diffusion steps, and noise schedules. The headline claim that AirLLM strictly dominates AdaLoRA on both accuracy and transmitted parameters is falsifiable and would be practically valuable if supported by the experiments. However, the current manuscript leaves the environment dynamics underspecified, contains a reward-indexing inconsistency, and reports results from a single seed, so the central empirical claims cannot yet be validated.

major comments (5)
  1. [III.B and IV.A.1] The reward in Eq. (10) is written as R_t = -U(r_t) - lambda*eta(r_t), but the MDP definition states that the action a_t produces r_{t+1}=a_t and that the state s_t contains r_t. Algorithm 1 (lines 12-13) deploys the new rank configuration and then computes the reward using Eq. (10). As written, R_t evaluates the old rank configuration r_t rather than the action just taken, so the advantage estimator in Eq. (11) and the DDIM reward term in Eq. (19) are not coherent objectives for the executed action. The authors should redefine R_t = -U(r_{t+1}) - lambda*eta(r_{t+1}), or otherwise clarify the indexing so that the reward is a function of the deployed action.
  2. [III.B and Algorithm 1] U(r_t) is the sole task-quality signal in the reward, but the paper does not specify the environment dynamics that produce it. There is no statement of how many fine-tuning gradient steps run between consecutive PPO decisions, whether U(r_t) is measured on a model freshly fine-tuned under the newly deployed rank or on the ongoing trajectory, how the SVD factors in Eq. (2) are resized or re-initialized when the rank changes, or which subset of D is used for the loss. If U(r_t) is a raw mid-trajectory loss, it will trend downward regardless of rank choice, confounding the GAE advantages in Eq. (11) and making the policy optimize training progress rather than rank allocation. Please specify the environment protocol precisely, including the reward computation and the warm-start versus fresh-start semantics.
  3. [Table II and Table IV] The main quantitative claim of strict dominance over AdaLoRA rests on Table IV, which reports single-point accuracy values with no error bars. Table II lists a single random seed (42). PPO and diffusion training are stochastic, and the reported accuracy differences are small relative to typical run-to-run variance for 1.3B-parameter fine-tuning (e.g., 0.46 percentage points at rmax=8 and 0.69 percentage points at rmax=64). Report mean and standard deviation or confidence intervals over at least three seeds, and state whether the differences are statistically significant; otherwise the 'consistently outperforms' claim is not supported.
  4. [Abstract, Section I, and Section V] The claimed 'over 30% training efficiency gains compared to vanilla PPO' in the Abstract and Introduction is not measured anywhere in Section V. Table VI reports total inference time for different DDIM inference steps, but there is no comparison of wall-clock time, number of environment steps, or sample efficiency between AirLLM and vanilla PPO. Please add a direct efficiency measurement (e.g., time or episodes to reach a target reward or accuracy) or remove the claim.
  5. [Eq. (10), Eq. (19), and Table IV] The 12.5% transmission reduction is an explicit term in the reward that the agents are trained to maximize (lambda*eta(r_t) in Eq. (10) and R_t in Eq. (19)), so reporting it as an independently discovered benefit of AirLLM is misleading. The paper should present the transmission saving as verification that the reward shaping works, not as an emergent advantage. To argue for a genuine advantage over AdaLoRA, compare methods at matched parameter budgets or provide a Pareto frontier of accuracy versus transmitted parameters; note also that U(r_t) is the same training loss that AdaLoRA optimizes, which limits the strength of any accuracy-only comparison.
minor comments (7)
  1. [Section I and Table IV] The abstract and Introduction say 'up to 12.5%' transmission reduction, but Table IV at rmax=64 corresponds to (32,399,048-28,234,160)/32,399,048 approximately 12.85%; please reconcile the numbers.
  2. [Section V.B] The text says Fig. 4 shows AirLLM achieves 0.36%-0.46% higher accuracy than AdaLoRA, but the plotted differences vary from about 0.1 to 0.47 percentage points across the SNR values shown; describe the actual observed range.
  3. [Table III] The table layout under the 'AirLLM' column is ambiguous: it appears to list PPO, DDIM, PPO+DDPM, PPO+DDIM/MLP, and PPO+DDIM/U-Net as separate variants, but the header does not clearly indicate which values correspond to which method.
  4. [Eq. (19)] The hybrid loss writes -kappa*E[R_t]; since R_t can be negative, please state explicitly whether the diffusion update maximizes the reward term via gradient ascent or uses a sign convention that keeps the loss well-defined.
  5. [Eq. (9) and Eq. (18)] The action space is defined as {0,1,...,rmax}^{6L}, but Eq. (18) rounds and clips continuous outputs; the paper does not describe how the total transmission-time constraint in Eq. (9) is enforced after this rounding.
  6. [Eq. (8) and Table I] In Table I, eta(r_t) is described as communication cost, but Eq. (8) also divides by C_t*T_max, making it dimensionless; please clarify the units and interpretation.
  7. [Fig. 7 and Table VI] The label 'DDPM Training Steps' is confusing because DDPM is a generative model, not a training-step count; rename it to 'diffusion training steps'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: AirLLM's reported gains are empirical outcomes of an explicitly defined reward, not predictions equivalent to its inputs.

full rationale

The derivation chain is self-contained. The RL reward (Eq. 10) is explicitly defined as the negative of the optimization objective (Eq. 9), and the DDIM training loss (Eq. 19) uses that same reward; this is a closed-loop policy-optimization design, not a circular derivation. The headline claims—0.69% accuracy gain and 12.5% parameter reduction over AdaLoRA at rmax=64—are benchmark results (Table IV, Fig. 4) obtained by comparing the trained policy's transmitted parameter counts and SST-2 accuracies against an external baseline, AdaLoRA [6]. The transmitted-parameter term eta(rt) is indeed a literal component of the reward, so reducing it is the intended objective rather than an independent discovery; however, the paper does not present this as a first-principles prediction, and the accuracy improvement is not forced by the reward definition. No load-bearing self-citations or imported uniqueness claims appear. The main caveat is an experimental-specification gap: Section III.B and Algorithm 1 (lines 12-13) never specify how many LLM gradient steps occur between rank decisions or how U(rt) is measured after deploying a new rank configuration, which makes the reward signal hard to reproduce; this is a correctness and reproducibility concern, not a circularity.

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

The framework relies on standard RL and diffusion math from cited works, plus domain assumptions about the wireless channel and data complexity features. The free parameters lambda, kappa, the early-stopping threshold, and the chosen noise schedule are tuned by the authors and affect the reported trade-offs.

free parameters (4)
  • lambda (reward balance factor) = 0.01, 0.1, 1
    Weights task loss vs communication cost in Eq. (10); the paper selects 0.1 as best for accuracy in Table V(a), so results depend on this choice.
  • kappa (diffusion reward weight) = 0.1
    Balances noise-prediction loss and policy reward in Eq. (19); no sensitivity analysis is provided.
  • early stopping reward threshold = min[-lambda, -0.5]
    Ad hoc convergence criterion in Section V-A that differs between RL and non-RL baselines.
  • noise schedule beta = scaled-linear
    Chosen after comparing cos, linear, scaled-linear in Fig. 8; the schedule that performs best on the test setting is used, a post hoc selection.
assumptions (5)
  • standard math PPO with GAE provides a valid policy gradient estimator (Schulman et al. 2017)
    Invoked as the coarse policy optimizer in Section IV-A.2 without modification.
  • standard math DDIM/DDPM theory with CFG training produces valid conditional samplers (Ho et al. 2020, Song et al. 2021, Ho and Salimans 2021)
    Used as the refinement module in Section IV-B; correctness is assumed from prior work.
  • domain assumption Wireless channel is slow-fading AWGN with perfect receiver CSI
    Equations (4)-(5) model the link; real channels can be more complex, but this is a stated modeling choice.
  • domain assumption Lexical entropy and OOV rate are sufficient state features for data complexity
    State s_t includes H_t and rho_t as the only data-complexity signals (Section IV-A.1); no ablation shows they are necessary.
  • domain assumption The fine-tuning loss U(r_t) can be evaluated accurately and cheaply after each rank change
    The entire reward (Eq. 10) and therefore the policy depends on this loss, yet the paper never specifies the loss evaluation procedure in the RL loop.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AirLLM: Diffusion Policy-based Adaptive LoRA for Remote Fine-Tuning of LLM over the Air." pith.science (2026). https://pith.science/paper/FTRYG7DG

@misc{pith2026250711515,
  author       = {Pith},
  title        = {Pith review of: AirLLM: Diffusion Policy-based Adaptive LoRA for Remote Fine-Tuning of LLM over the Air},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FTRYG7DG}},
  note         = {Machine review of arXiv:2507.11515}
}
read the original abstract

Operating Large Language Models (LLMs) on edge devices is increasingly challenged by limited communication bandwidth and strained computational and memory costs. Thus, cloud-assisted remote fine-tuning becomes indispensable. Nevertheless, existing Low-Rank Adaptation (LoRA) approaches typically employ fixed or heuristic rank configurations, and the subsequent over-the-air transmission of all LoRA parameters could be rather inefficient. To address this limitation, we develop AirLLM, a hierarchical diffusion policy framework for communication-aware LoRA adaptation. Specifically, AirLLM models the rank configuration as a structured action vector that spans all LoRA-inserted projections. To solve the underlying high-dimensional sequential decision-making problem, a Proximal Policy Optimization (PPO) agent generates coarse-grained decisions by jointly observing wireless states and linguistic complexity, which are then refined via Denoising Diffusion Implicit Models (DDIM) to produce high-resolution, task- and channel-adaptive rank vectors. The two modules are optimized alternatively, with the DDIM trained under the Classifier-Free Guidance (CFG) paradigm to maintain alignment with PPO rewards. Experiments under varying signal-to-noise ratios demonstrate that AirLLM consistently enhances fine-tuning performance while significantly reducing transmission costs, highlighting the effectiveness of reinforcement-driven, diffusion-refined rank adaptation for scalable and efficient remote fine-tuning over the air.

Figures

Figures reproduced from arXiv: 2507.11515 by the authors.

Figure 2
Figure 2. 1) Conditional Diffusion Refinement Guided by PPO Priors: We begin by briefly introducing the forward and reverse processes of the DDPM and its deterministic variant DDIM, which achieves faster inference with similar generative quality [18]. Notably, DDIM shares the same training objective and noise schedule as DDPM, with the main distinction residing in the sampling procedure used during inference. a) Forward Diffu… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 41 canonical work pages

  1. [1]

    , L} contains a self-attention (SA) block and an FFN block

    LLM Parameter-Efficient Fine-tuning: We consider a pre-trained Transformer-based LLM M composed of L layers, where each layer l ∈ {1, . . . , L} contains a self-attention (SA) block and an FFN block. Without loss of generality, we exemplify the PEFT of LLM by AdaLoRA [6]. In the SA block, AdaLoRA modules (i.e., projection matrices) are inserted into the f...

    work page

  2. [2]

    Wireless Communication Channel Model: For the re- mote fine-tuning of LLMs, we assume a typical cloud-edge deployment setting where the fine-tuning process is conducted on the cloud server, and the resulting low-rank parameters are transmitted to resource-constrained local edge devices. We model the wireless transmission of fine-tuned parameters from the ...

    work page

  3. [3]

    Computational Complexity Analyses: To circumvent the inefficiency of directly generating high-dimensional discrete actions via PPO, we formulate the policy as a continuous vector generator and discard the softmax operation of PPO, resulting an O(6L · (rmax + 1)) saving for the computational complexity. On the other hand, the fine-grained decision- making ...

    work page

  4. [4]

    Scaling down to scale up: A guide to parameter-efficient fine-tuning,

    V . Lialin, V . Deshpande, and A. Rumshisky, “Scaling down to scale up: A guide to parameter-efficient fine-tuning,” arXiv preprint arXiv:2303.15647, 2023. 11

    arXiv 2023

  5. [5]

    To enable adaptive rank configuration under real-world constraints such as wireless bandwidth and task complexity, we formulate the rank allocation as an MDP (S, A, R)

    MDP Formulation for Rank Allocation: RL provides a dynamically adaptive decision-making paradigm for remote fine-tuning of large models, with its core advantage lying in real-time strategy optimization through environmental feed- back. To enable adaptive rank configuration under real-world constraints such as wireless bandwidth and task complexity, we for...

    work page

  6. [6]

    AirLLM: Diffusion Policy-based Adaptive LoRA for Remote Fine-Tuning of LLM over the Air

    decompose updates into low-rank matrices or dynamically allocate rank budgets based on Singular Value Decomposition (SVD). While effective in reducing computational load, they primarily focus on the training process itself, often overlooking the deployment constraints of real-world systems. S. Yang, X. Yu, R. Li, and J. Zhu are with the College of Informa...

    work page Pith review arXiv 2025

  7. [7]

    Beforehand, we briefly present the key ingre- dients related to PPO

    PPO for Coarse-Grained Policy Optimization: As noted above, we adopt PPO to train a feasible policy for dynamic rank allocation. Beforehand, we briefly present the key ingre- dients related to PPO. In particular, the return at time t is defined as ˜Rt = P∞ k=0 γkRt+k, where γ ∈ (0, 1) discounts future rewards. The state-value function Vϕ(st) = Eπ[ ˜Rt|st]...

    work page

  8. [8]

    Notably, DDIM shares the same training objective and noise schedule as DDPM, with the main distinction residing in the sampling procedure used during inference

    Conditional Diffusion Refinement Guided by PPO Priors: We begin by briefly introducing the forward and reverse processes of the DDPM and its deterministic variant DDIM, which achieves faster inference with similar generative quality [18]. Notably, DDIM shares the same training objective and noise schedule as DDPM, with the main distinction residing in the...

    work page

Show all 50 references
  1. [9]

    Hybrid Training Objective and Optimization.: To en- hance controllability without relying on external classifiers, the training adopts the CFG strategy [33]. Specifically, the conditional diffusion model is trained to minimize a hybrid loss that combines reconstruction fidelit...

  2. [10]

    Pre-trained models for natural language processing: A survey,

    X. Qiu, T. Sun, Y . Xu, Y . Shao, N. Dai, and X. Huang, “Pre-trained models for natural language processing: A survey,” Sci. China Technol. Sci., vol. 63, no. 10, pp. 1872–1897, 2020

    2020

  3. [11]

    Notably, we also study the performance differences when using MLP and U-Net as backbones of DDIM

    Performance Comparison: Table III first compares the proposed AirLLM with AdaLoRA, while several variants of AirLLM with different RL algorithms are also leveraged. Notably, we also study the performance differences when using MLP and U-Net as backbones of DDIM. On the other h...

  4. [12]

    (10), regulates the trade-off between task accuracy and communication cost

    Performance Sensitivity Studies: Next, we examine the impact of the reward balancing coefficient λ, which as defined in Eq. (10), regulates the trade-off between task accuracy and communication cost. Fig. 5 shows λ = 0 .1 or λ = 0 .01 achieves superior, competitive balance, na...

  5. [13]

    GPT-4 technical report,

    OpenAI, “GPT-4 technical report,” OpenAI, Tech. Rep. arXiv:2303.08774, 2023, arXiv preprint arXiv:2303.08774. [Online]. Available: https://arxiv.org/abs/2303.08774

    2023 arXiv

  6. [14]

    The proposed RL agent observes both data and channel state information [15] and dynamically adjusts rank budgets to balance model accuracy and transmission cost

    framework to learn adaptive rank allocation policies. The proposed RL agent observes both data and channel state information [15] and dynamically adjusts rank budgets to balance model accuracy and transmission cost. However, standard RL algorithms still struggle with the under...

  7. [15]

    Deepseek-v2 technical report,

    DeepSeek-AI, “Deepseek-v2 technical report,” DeepSeek-AI, Tech. Rep. arXiv:2405.04434, 2024, arXiv preprint arXiv:2405.04434. [Online]. Available: https://arxiv.org/abs/2405.04434

    2024 arXiv

  8. [17]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in Proc. NeurIPS, Virtual Edition, Dec. 2020

    2020

  9. [18]

    LoRA: Low-Rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-Rank adaptation of large language models,” in Proc. ICLR, Virtual Edition, Apr. 2022

    2022

  10. [19]

    AdaLoRA: Adaptive budget allocation for parameter-efficient fine-tuning,

    Q. Zhang, Z. Liu, J. Fu, H. Dong, X. Han, P. Zhang, Y . Sun, H. Tian, H. Wu, and H. Wang, “AdaLoRA: Adaptive budget allocation for parameter-efficient fine-tuning,” in Proc. ICLR , Kigali, Rwanda, May 2023

    2023

  11. [20]

    Adaptation in cloud resource configuration: a survey,

    A. R. Hummaida, N. W. Paton, and R. Sakellariou, “Adaptation in cloud resource configuration: a survey,” J. Cloud Comput. , vol. 5, no. 1, p. 7, 2016

    2016

  12. [21]

    Curriculum learning for natural language understanding,

    B. Xu, L. Zhang, Z. Mao, Q. Wang, H. Xie, and Y . Zhang, “Curriculum learning for natural language understanding,” in Proc. ACL , Virtual Edition, Jul. 2020

    2020

  13. [22]

    Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding,

    S. Han, H. Mao, and W. J. Dally, “Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding,” in Proc. ICLR, San Juan, Puerto Rico, May 2016

    2016

  14. [23]

    Goldsmith, Wireless Communications

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

    2005

  15. [24]

    Channel characteristics and transmission performance for various channel configurations at 60 GHz,

    H. Yang, P. F. Smulders, and M. H. Herben, “Channel characteristics and transmission performance for various channel configurations at 60 GHz,” Eurasip J. Wirel. Commun. Netw. , vol. 2007, no. 1, p. 019613, 2007

    2007

  16. [25]

    Feed-forward neural networks,

    G. Bebis and M. Georgiopoulos, “Feed-forward neural networks,” IEEE Potentials, vol. 13, no. 4, pp. 27–31, 1994

    1994

  17. [26]

    Real-time millimeter-wave MIMO channel sounder for dynamic directional me- asurements,

    C. U. Bas, R. Wang, S. Sangodoyin, D. Psychoudakis, T. Henige, R. Monroe, J. Park, C. J. Zhang, and A. F. Molisch, “Real-time millimeter-wave MIMO channel sounder for dynamic directional me- asurements,” IEEE Trans. Veh. Technol., vol. 68, no. 9, pp. 8775–8789, 2019

    2019

  18. [27]

    Szepesvári, Algorithms for reinforcement learning

    C. Szepesvári, Algorithms for reinforcement learning. Springer nature, 2022

    2022

  19. [28]

    Adaptive sampling and joint semantic- channel coding under dynamic channel environment,

    Z. Qi, Y . Feng, and Z. Qin, “Adaptive sampling and joint semantic- channel coding under dynamic channel environment,” arXiv preprint arXiv:2502.07236, 2025

    2025 arXiv

  20. [29]

    Diffusion policy policy optimization,

    A. Z. Ren, T. Yu, C. Finn, and S. Levine, “Diffusion policy policy optimization,” in Proc. ICLR, Singapore, Apr 2025

    2025

  21. [30]

    Deep reinforcement learning that matters,

    P. Henderson, R. Islam, P. Bachman, J. Pineau, D. Precup, and D. Meger, “Deep reinforcement learning that matters,” inProc. AAAI, New Orleans, LA, USA, Feb. 2018

    2018

  22. [31]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” in Proc. ICLR, Vienna, Austria, May 2021

    2021

  23. [32]

    Proxi- mal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proxi- mal policy optimization algorithms,” arXiv preprint arXiv:1707.06347 , 2017

    2017 arXiv

  24. [33]

    PEFT-U: Parameter-efficient fine-tuning for user personalization,

    C. Clarke, Y . Heng, L. Tang, and J. Mars, “PEFT-U: Parameter-efficient fine-tuning for user personalization,” arXiv preprint arXiv:2407.18078 , 2024

    2024 arXiv

  25. [34]

    Wireless com- munication is modeled as an AWGN channel with 100 MHz bandwidth, 1s latency, and SNR levels ranging from −5 dB to 15 dB

    (i.e., a binary classification dataset) with the OPT- 1.3B model [35] (a 24-layer Transformer decoder) equipped with LoRA adapters for all linear projection layers. Wireless com- munication is modeled as an AWGN channel with 100 MHz bandwidth, 1s latency, and SNR levels rangin...

  26. [35]

    Intelligent cloud-edge collaborations assisted energy-efficient power control in heterogeneous networks,

    L. Zhang, J. Peng, J. Zheng, and M. Xiao, “Intelligent cloud-edge collaborations assisted energy-efficient power control in heterogeneous networks,” IEEE Trans. Wirel. Commun., vol. 22, no. 11, pp. 7743–7755, 2023

    2023

  27. [36]

    T. S. Rappaport, Wireless Communications: Principles and Practice . Pearson Education India, 2020

    2020

  28. [37]

    dLoRA: Dyna- mically orchestrating requests and adapters for LoRA/LLM serving,

    B. Wu, R. Zhu, Z. Zhang, P. Sun, X. Liu, and X. Jin, “dLoRA: Dyna- mically orchestrating requests and adapters for LoRA/LLM serving,” in Proc. OSDI, Santa Clara, CA, USA, Jul. 2024

    2024

  29. [38]

    R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction. MIT Press, 2018

    2018

  30. [39]

    Reinforcement learning from suboptimal demonstrations based on reward relabeling,

    Y . Peng, J. Zeng, Y . Hu, Q. Fang, and Q. Yin, “Reinforcement learning from suboptimal demonstrations based on reward relabeling,” Expert Syst. Appl., vol. 255, p. 124580, 2024

    2024

  31. [40]

    Learning continuous control policies by stochastic value gradients,

    N. Heess, H. Soyer, A. Saxena, T. Joachims, T. Degris, P. M. Pilarski, A. J. Ballard, D. Wierstra, and J. Peters, “Learning continuous control policies by stochastic value gradients,” in Proc. NeurIPS , Montreal, Quebec, Canada, Dec. 2015

    2015

  32. [41]

    Latent space po- licies for hierarchical reinforcement learning,

    T. Haarnoja, K. Hartikainen, P. Abbeel, and S. Levine, “Latent space po- licies for hierarchical reinforcement learning,” in Proc. ICML. PMLR, 2018, pp. 1851–1860

    2018

  33. [42]

    Off-policy reinforcement learning with high dimensional reward,

    D. N. Lee and M. R. Kosorok, “Off-policy reinforcement learning with high dimensional reward,” arXiv preprint arXiv:2408.07660 , 2024

    2024 arXiv

  34. [44]

    What matters in on-policy reinforcement learning? A large-scale empirical study,

    M. Andrychowicz, A. Raichuk, P. Sta ´nczyk, M. Orlowski, L. Espeholt, R. Marinier, M. Zi˛ eba, J. Kay, Y . Tassa, N. Heesset al., “What matters in on-policy reinforcement learning? A large-scale empirical study,” in Proc. ICLR, Vienna, Austria, May 2021

    2021

  35. [45]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” in Proc. CVPR, New Orleans, LA, USA, Jun. 2022

    2022

  36. [46]

    Classifier-free diffusion guidance,

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,” in Proc. NeurIPS, Virtual Edition, Dec. 2021

    2021

  37. [47]

    Recursive deep models for semantic compositionality over a sentiment treebank,

    R. Socher, A. Perelygin, J. Wu, J. Chuang, C. D. Manning, A. Y . Ng, and C. Potts, “Recursive deep models for semantic compositionality over a sentiment treebank,” in Proc. EMNLP, Seattle, W A, USA, Oct. 2013

    2013

  38. [48]

    OPT: Open pre-trained transformer language models,

    S. Zhang, S. Roller, N. Goyal, M. Artetxe, M. Chen, S. Chen, G. Dewan, M. Diab, J. Dodge, X. L. Maestre, X. V . Lin, T. Mihaylov, M. Ott, S. Shleifer, K. Shuster, D. Simig, P. S. Koura, A. Singh, H. Schwenk, and L. Zettlemoyer, “OPT: Open pre-trained transformer language model...

    2022 arXiv

  39. [49]

    Early stopping and non- parametric regression: an optimal data-dependent stopping rule,

    G. Raskutti, M. J. Wainwright, and B. Yu, “Early stopping and non- parametric regression: an optimal data-dependent stopping rule,” J. Mach. Learn. Res. , vol. 15, no. 1, p. 335–366, Jan. 2014

    2014

  40. [50]

    Diffusion Policy: Visuomotor policy learning via action diffusion,

    C. Chi, S. Feng, Y . Du, Z. Xu, E. Cousineau, B. Burchfiel, and S. Song, “Diffusion Policy: Visuomotor policy learning via action diffusion,” in Proc. RSS, Daegu, Republic of Korea, Jul. 2023

    2023

  41. [51]

    Wider and deeper LLM networks are fairer LLM evaluators,

    X. Zhang, B. Yu, H. Yu, Y . Lv, T. Liu, F. Huang, H. Xu, and Y . Li, “Wider and deeper LLM networks are fairer LLM evaluators,” arXiv preprint arXiv:2308.01862, 2023

    2023 arXiv

  42. [2023]

    Available: https://arxiv.org/abs/2303.18223

    [Online]. Available: https://arxiv.org/abs/2303.18223

    arXiv

Pith tools

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