AirLLM: Wireless Adaptive Fine-Tuning of LLMs

• AirLLM is a framework that fine-tunes large language models remotely by adapting LoRA ranks based on wireless channel conditions and task complexity.
• It employs a hierarchical approach combining PPO for coarse rank allocation with DDIM for projection-wise refinement, ensuring structured adaptation.
• The system optimizes the trade-off between model accuracy and communication cost by reducing transmitted parameters while maintaining performance.

AirLLM is a communication-aware remote fine-tuning framework for LLMs in cloud–edge settings, where a model is adapted in the cloud and only the fine-tuned low-rank parameters are transmitted over a wireless channel to resource-constrained edge devices. Its central claim is that, in this setting, the LoRA or AdaLoRA rank configuration should not be treated as a static training hyperparameter; rank becomes a dynamic control variable that must respond jointly to wireless conditions and task or data complexity. AirLLM formulates this control problem as hierarchical sequential decision-making, using a Proximal Policy Optimization agent to generate coarse rank-allocation priors and a Denoising Diffusion Implicit Models module to refine them into projection-wise rank vectors over all LoRA-inserted matrices (Yang et al., 15 Jul 2025).

1. Communication-aware remote fine-tuning setting

AirLLM addresses a narrow but important deployment bottleneck: running or fine-tuning LLMs directly on edge devices is difficult because modern LLMs have large memory footprints, high compute demands, and substantial communication requirements when updates must be synchronized or delivered remotely. Even parameter-efficient fine-tuning only partially resolves this bottleneck, because after cloud-side tuning the low-rank adapter parameters must still be transmitted to the edge. Standard LoRA typically uses a fixed rank $r$ across layers, while AdaLoRA adapts rank heuristically based on parameter importance during training; neither is designed for wireless remote delivery, where channel quality fluctuates and communication is itself part of the optimization problem (Yang et al., 15 Jul 2025).

The framework assumes a pre-trained Transformer-based LLM $\mathscr{M}$ with $L$ layers. LoRA-style low-rank modules are inserted into six projections per layer,

$$\mathcal{M}_{l} = \{ \mathbf{W}_l^Q, \mathbf{W}_l^K, \mathbf{W}_l^V, \mathbf{W}_l^O, \mathbf{W}_l^\text{fc1}, \mathbf{W}_l^\text{fc2} \}.$$

Rank is therefore allocated per inserted projection, not globally, producing a structured rank vector over all $6L$ adapted matrices. This design makes communication-aware adaptation intrinsically high-dimensional: each layer and each projection can receive a different rank budget, and the optimal choice depends on both linguistic difficulty and current wireless capacity (Yang et al., 15 Jul 2025).

A key conceptual shift is the paper’s definition of communication cost. In AirLLM, communication cost is the amount of LoRA or AdaLoRA parameters that must be transmitted over the air, normalized by achievable channel rate and latency budget. Better channels support larger transmitted rank budgets; poorer channels force more aggressive compression. This suggests that rank selection should be viewed not merely as a model-capacity choice, but as a cross-layer control signal coupling task quality to wireless resource availability.

2. LoRA parameterization, wireless model, and optimization objective

AirLLM builds on AdaLoRA’s SVD-style parameterization. For module $m$ in layer $l$, the fine-tuned weight at training step $t$ is

$$\mathbf{W}^{(t)}_{l,m} = \mathbf{W}_{l,m}^{(0)} + \mathbf{P}^{(t)}_{l,m} \mathbf{\Lambda}^{(t)}_{l,m} \mathbf{Q}^{(t)}_{l,m},$$

with

$$\mathbf{P}^{(t)}_{l,m} \in \mathbb{R}^{d_h \times r^{(t)}_{l,m}}, \quad \mathbf{Q}^{(t)}_{l,m} \in \mathbb{R}^{r^{(t)}_{l,m} \times d_h}.$$

The trainable parameter count contributed by one module is $\mathscr{M}$0, and the total trainable parameter count across all inserted modules is

$\mathscr{M}$1

The formulation also includes an orthogonality regularizer,

$\mathscr{M}$2

These expressions make the communication burden explicit: rank is directly proportional to the number of transmitted low-rank parameters (Yang et al., 15 Jul 2025).

The wireless system model is cloud-to-edge transmission over a slow-fading AWGN channel. A codeword $\mathscr{M}$3 is transmitted and the edge receives

$\mathscr{M}$4

where $\mathscr{M}$5 is the slow-fading gain and $\mathscr{M}$6. Assuming perfect CSI at the receiver, the average achievable rate is

$\mathscr{M}$7

For module $\mathscr{M}$8, transmission time is modeled as

$\mathscr{M}$9

with $L$0 for FP32, subject to a total latency budget and rank bounds. The normalized communication cost of a rank configuration $L$1 is

$L$2

The optimization objective is

$L$3

with constraints enforcing the latency budget and $L$4. Here $L$5 is the fine-tuning loss under rank configuration $L$6, used as a surrogate for task quality, and $L$7 balances accuracy against transmission burden. AirLLM therefore treats parameter-efficient fine-tuning as a joint adaptation-and-delivery problem rather than a purely training-side problem (Yang et al., 15 Jul 2025).

3. Hierarchical diffusion policy

AirLLM formulates adaptive rank allocation as a Markov decision process. The state is

$L$8

where $L$9 is lexical entropy and $$\mathcal{M}_{l} = \{ \mathbf{W}_l^Q, \mathbf{W}_l^K, \mathbf{W}_l^V, \mathbf{W}_l^O, \mathbf{W}_l^\text{fc1}, \mathbf{W}_l^\text{fc2} \}.$$0 is OOV rate of the fine-tuning data at step $$\mathcal{M}_{l} = \{ \mathbf{W}_l^Q, \mathbf{W}_l^K, \mathbf{W}_l^V, \mathbf{W}_l^O, \mathbf{W}_l^\text{fc1}, \mathbf{W}_l^\text{fc2} \}.$$1. These encode linguistic complexity: lexical entropy captures diversity and unpredictability in the token distribution, while OOV rate measures vocabulary mismatch between the task data and the pretraining vocabulary. The action directly defines the next rank configuration,

$$\mathcal{M}_{l} = \{ \mathbf{W}_l^Q, \mathbf{W}_l^K, \mathbf{W}_l^V, \mathbf{W}_l^O, \mathbf{W}_l^\text{fc1}, \mathbf{W}_l^\text{fc2} \}.$$2

and the reward is

$$\mathcal{M}_{l} = \{ \mathbf{W}_l^Q, \mathbf{W}_l^K, \mathbf{W}_l^V, \mathbf{W}_l^O, \mathbf{W}_l^\text{fc1}, \mathbf{W}_l^\text{fc2} \}.$$3

The paper argues that directly learning in $$\mathcal{M}_{l} = \{ \mathbf{W}_l^Q, \mathbf{W}_l^K, \mathbf{W}_l^V, \mathbf{W}_l^O, \mathbf{W}_l^\text{fc1}, \mathbf{W}_l^\text{fc2} \}.$$4 is too high-dimensional for vanilla PPO, because reward differences across neighboring actions become small and sparse, destabilizing advantage estimation (Yang et al., 15 Jul 2025).

The proposed remedy is a hierarchical diffusion policy. PPO first produces a coarse prior $$\mathcal{M}_{l} = \{ \mathbf{W}_l^Q, \mathbf{W}_l^K, \mathbf{W}_l^V, \mathbf{W}_l^O, \mathbf{W}_l^\text{fc1}, \mathbf{W}_l^\text{fc2} \}.$$5, and a conditional DDIM module then refines that prior into a high-resolution structured rank vector. In the diffusion stage, a clean rank vector $$\mathcal{M}_{l} = \{ \mathbf{W}_l^Q, \mathbf{W}_l^K, \mathbf{W}_l^V, \mathbf{W}_l^O, \mathbf{W}_l^\text{fc1}, \mathbf{W}_l^\text{fc2} \}.$$6 is progressively noised according to

$$\mathcal{M}_{l} = \{ \mathbf{W}_l^Q, \mathbf{W}_l^K, \mathbf{W}_l^V, \mathbf{W}_l^O, \mathbf{W}_l^\text{fc1}, \mathbf{W}_l^\text{fc2} \}.$$7

with closed-form marginal

$$\mathcal{M}_{l} = \{ \mathbf{W}_l^Q, \mathbf{W}_l^K, \mathbf{W}_l^V, \mathbf{W}_l^O, \mathbf{W}_l^\text{fc1}, \mathbf{W}_l^\text{fc2} \}.$$8

The reverse denoiser conditions on the PPO prior through

$$\mathcal{M}_{l} = \{ \mathbf{W}_l^Q, \mathbf{W}_l^K, \mathbf{W}_l^V, \mathbf{W}_l^O, \mathbf{W}_l^\text{fc1}, \mathbf{W}_l^\text{fc2} \}.$$9

Inference uses DDIM sampling for faster refinement, and after $6L$0 denoising steps the refined continuous latent is discretized as

$6L$1

The paper characterizes this division of labor succinctly: PPO provides a low-resolution structural guess, and DDIM fills in a coherent, projection-wise rank assignment over all $6L$2 dimensions (Yang et al., 15 Jul 2025).

The DDIM module is trained under the Classifier-Free Guidance paradigm to maintain alignment with PPO rewards, but the practical training objective given in the paper is a hybrid DDIM loss,

$6L$3

This makes the diffusion model explicitly reward-aware rather than purely reconstructive.

4. Alternating optimization and computational profile

Training alternates between PPO and DDIM. At each environment step, PPO outputs a coarse action $6L$4; a Gaussian latent is initialized; DDIM denoising runs backward from $6L$5 to $6L$6; and the final latent is rounded and clipped to form the executable rank vector $6L$7. The environment then returns the next state and reward, and tuples $6L$8 are stored. PPO is updated first with the collected trajectories; holding PPO fixed, DDIM is then trained from noisy samples generated from clean action targets $6L$9. Both modules share the same replay data, so the optimization is alternating but cooperative (Yang et al., 15 Jul 2025).

The paper also gives a complexity argument for the hierarchy. Vanilla PPO over all discrete heads incurs output complexity involving $m$0. By reformulating PPO as a continuous coarse vector generator and leaving structured synthesis to diffusion, AirLLM claims savings of $m$1 from discarding softmax over all discrete heads. DDIM with a residual MLP denoiser has per-step complexity $m$2, so total inference complexity over $m$3 steps is $m$4. A U-Net denoiser is also analyzed and is more expensive (Yang et al., 15 Jul 2025).

This architecture implies a separation between global reward alignment and structured output synthesis. A plausible implication is that AirLLM is best read not as a generic RL policy or a generic diffusion controller, but as a two-stage structured-action generator designed for combinatorial rank allocation under wireless constraints.

5. Experimental evaluation

The experiments use OPT-1.3B, a 24-layer decoder-only Transformer, on SST-2 binary sentiment classification. LoRA adapters are inserted into all linear projection layers. Wireless transmission uses an AWGN channel with 100 MHz bandwidth, 1 s latency, and SNRs from $m$5 dB to $m$6 dB. Training runs for 15,000 steps, with evaluation every 100 steps and early stopping if reward exceeds $m$7 and accuracy stagnates for five evaluations. Baselines include AdaLoRA, random rank allocation, standalone PPO, standalone DDIM, PPO+DDPM, PPO+DDIM with MLP denoiser, and PPO+DDIM with U-Net denoiser (Yang et al., 15 Jul 2025).

The main comparison shows that hierarchical variants outperform standalone PPO or DDIM:

Method	Accuracy
PPO	94.50%
DDIM	94.72%
PPO+DDPM	95.41%
PPO+DDIM/MLP	95.53%
PPO+DDIM/U-Net	95.30%
AdaLoRA	95.18%

The strongest reported variant is PPO+DDIM/MLP, improving over AdaLoRA by $m$8 in that comparison. Elsewhere the paper reports gains of up to $m$9 higher accuracy and $l$0 fewer transmitted parameters versus AdaLoRA, especially at $l$1 (Yang et al., 15 Jul 2025).

$l$2	Accuracy: AdaLoRA / AirLLM	Transmitted parameters: AdaLoRA / AirLLM
8	95.18% / 95.64%	4,718,592 / 4,235,648
16	94.72% / 95.30%	9,437,184 / 8,471,296
32	95.41% / 95.87%	18,874,368 / 16,942,592
64	94.84% / 95.53%	32,399,048 / 28,234,160
128	94.72% / 95.19%	63,870,464 / 56,664,832

At $l$3, the headline result is a $l$4 accuracy gain and about $l$5 fewer transmitted parameters. In the channel-robustness study, AirLLM consistently outperforms AdaLoRA over SNRs from $l$6 dB to $l$7 dB, with gains around $l$8–$l$9. The paper attributes this to channel-aware rank adaptation rather than static or heuristic allocation (Yang et al., 15 Jul 2025).

Several ablations further define the method’s empirical profile. PPO-only and DDIM-only perform worst, supporting the hierarchical design. PPO+DDPM performs well, but PPO+DDIM converges faster, reflecting DDIM’s lower sampling cost than DDPM. The MLP denoiser performs slightly better than U-Net while being computationally lighter. The reward coefficient $t$0 also matters: reported accuracies are $t$1 for $t$2, $t$3 for $t$4, and $t$5 for $t$6, making $t$7 the best setting in the paper. For diffusion parameterization, both $t$8-prediction and $t$9-prediction reach $$\mathbf{W}^{(t)}_{l,m} = \mathbf{W}_{l,m}^{(0)} + \mathbf{P}^{(t)}_{l,m} \mathbf{\Lambda}^{(t)}_{l,m} \mathbf{Q}^{(t)}_{l,m},$$0, outperforming $$\mathbf{W}^{(t)}_{l,m} = \mathbf{W}_{l,m}^{(0)} + \mathbf{P}^{(t)}_{l,m} \mathbf{\Lambda}^{(t)}_{l,m} \mathbf{Q}^{(t)}_{l,m},$$1-prediction at $$\mathbf{W}^{(t)}_{l,m} = \mathbf{W}_{l,m}^{(0)} + \mathbf{P}^{(t)}_{l,m} \mathbf{\Lambda}^{(t)}_{l,m} \mathbf{Q}^{(t)}_{l,m},$$2. On diffusion-step counts, 1000 training steps and 50 DDIM inference steps achieve $$\mathbf{W}^{(t)}_{l,m} = \mathbf{W}_{l,m}^{(0)} + \mathbf{P}^{(t)}_{l,m} \mathbf{\Lambda}^{(t)}_{l,m} \mathbf{Q}^{(t)}_{l,m},$$3 accuracy; with 600 DDPM training steps, timing is 907 s for 30 DDIM inference steps, 1037 s for 50, and 1101 s for 100. The paper also reports 20–30% convergence acceleration via DDIM sampling and over 30% training efficiency gain compared with vanilla PPO (Yang et al., 15 Jul 2025).

A qualitative result concerns learned rank distribution. Under a scaled-linear $$\mathbf{W}^{(t)}_{l,m} = \mathbf{W}_{l,m}^{(0)} + \mathbf{P}^{(t)}_{l,m} \mathbf{\Lambda}^{(t)}_{l,m} \mathbf{Q}^{(t)}_{l,m},$$4 schedule, higher average ranks appear in lower layers and larger ranks are allocated to $$\mathbf{W}^{(t)}_{l,m} = \mathbf{W}_{l,m}^{(0)} + \mathbf{P}^{(t)}_{l,m} \mathbf{\Lambda}^{(t)}_{l,m} \mathbf{Q}^{(t)}_{l,m},$$5 and $$\mathbf{W}^{(t)}_{l,m} = \mathbf{W}_{l,m}^{(0)} + \mathbf{P}^{(t)}_{l,m} \mathbf{\Lambda}^{(t)}_{l,m} \mathbf{Q}^{(t)}_{l,m},$$6, which the authors say aligns with decoder-only model behavior and AdaLoRA’s importance findings. This suggests that AirLLM learns structured, nonuniform allocations rather than trivial global rank shrinkage.

6. Interpretation, scope, and nomenclature

The paper’s main explanatory claim is that fixed-rank LoRA ignores cross-layer heterogeneity, while heuristic rank adjustment such as AdaLoRA responds to parameter-importance signals but not to wireless delivery constraints or real-time data complexity. PPO can ingest channel and data signals but struggles in the full combinatorial rank space; DDIM can model coherent high-dimensional outputs but lacks a reward-grounded global objective. AirLLM combines the two: PPO provides a reward-aligned global prior over adaptation intensity, and DDIM imposes structured refinement over all $$\mathbf{W}^{(t)}_{l,m} = \mathbf{W}_{l,m}^{(0)} + \mathbf{P}^{(t)}_{l,m} \mathbf{\Lambda}^{(t)}_{l,m} \mathbf{Q}^{(t)}_{l,m},$$7 rank dimensions (Yang et al., 15 Jul 2025).

Its limitations are also clear. Experiments are restricted to SST-2 with OPT-1.3B, so generalization to QA, instruction tuning, or summarization is not demonstrated. The channel model is idealized as slow-fading AWGN with perfect CSI at the receiver. The alternating PPO–diffusion stack adds computational overhead compared with standard AdaLoRA. The method also depends on handcrafted state features such as lexical entropy and OOV rate, and the paper leaves open whether these remain sufficient on more diverse tasks. Future work explicitly mentioned by the authors includes multi-agent collaboration and adaptive noise scheduling for more extreme channel conditions (Yang et al., 15 Jul 2025).

Despite its name, AirLLM is not an aviation-domain foundation model. It is an “over the air” remote fine-tuning method. This distinguishes it from UAV-specific distributed architectures such as Aero-LLM (Dharmalingam et al., 5 Feb 2025), aviation knowledge systems such as AviationLLM (Wan et al., 17 Jun 2025), and aviation safety-analysis prototypes around non-towered airports (Darrell et al., 12 May 2026). It is also distinct from other works that use the acronym AIR, including Activation- and Influence-Aware Ranks for compression (Harder et al., 18 Jun 2026) and Adaptive Interleaved Reasoning with Code in MLLMs (Han et al., 22 Jun 2026). The closest systems-level neighbors are over-the-air distributed inference frameworks that use AirComp for tensor-parallel LLM execution over wireless networks (Zhang et al., 18 Feb 2025, Zhang et al., 19 Mar 2025). AirLLM differs from those efforts by focusing on remote fine-tuning and communication-aware LoRA rank adaptation rather than on collaborative inference.

In that narrower sense, AirLLM occupies a specific place in the PEFT and wireless-LLM literature: it recasts adapter rank as a context-dependent control variable and couples reinforcement learning with diffusion-based structured action synthesis to optimize accuracy–communication trade-offs for cloud-to-edge LLM adaptation (Yang et al., 15 Jul 2025).