Token-Aware QoE Optimization for Multimodal AIGC Inference Tasks in Space-Air-Ground Integrated Networks

Linghao Zhang; Bo Xu; Haitao Zhao; Jinlong Sun; Jun Xiong; Xianbin Wang

doi:10.1109/TMC.2026.3743972

1/4 The three-tier architecture for multimodal AIGC inference (Fig. 1).

2/4 The inference pipeline and end-to-end streaming latency model (Fig. 2).

3/4 The two-timescale optimization framework (Fig. 3).

IEEE TMC2026Early AccessFirst author

Token-Aware QoE Optimization for Multimodal AIGC Inference Tasks in Space-Air-Ground Integrated Networks

Linghao Zhang, Bo Xu, Haitao Zhao, Jinlong Sun, Jun Xiong, Xianbin Wang

IEEE Transactions on Mobile Computing, pp. 1–18, 2026

IEEE Xplore

+31.9%
Per-user QoE over TD3: +69.2%
Per-user QoE over Greedy

Abstract

The space-air-ground integrated network (SAGIN) can provide ubiquitous multimodal artificial intelligence-generated content (AIGC) services. In SAGIN, lightweight task-specific models hosted on uncrewed aerial vehicles (UAVs) provide low-latency inference, while full-scale foundation models deployed on low Earth orbit (LEO) satellites provide high-quality inference. However, efficient AIGC inference in SAGIN is hindered by the tight coupling between computation and cross-tier transmission under multi-timescale network dynamics. In this paper, we maximize the long-term per-user quality of experience (QoE), which captures both the generation quality and the end-to-end streaming latency of heterogeneous multimodal AIGC requests. To this end, we develop a unified token-level latency model that characterizes the time-to-first-token (TTFT) and the time-per-output-token (TPOT) along multi-tier offloading paths. We then formulate the joint optimization of model caching, UAV deployment, inference offloading, semantic compression, and resource allocation as a two-timescale mixed-integer nonlinear program (MINLP). For the slow-timescale subproblem, a knowledge-driven planner generates and evaluates candidate UAV deployment and model caching configurations from historical demand priors. For the fast-timescale subproblem, we propose a conditional diffusion twin delayed deep deterministic policy gradient (CD-TD3) algorithm with a hybrid action space, which optimizes offloading decisions, compression ratios, and resource allocation in real time. Simulation results show that the proposed framework improves the per-user QoE by 31.9% over TD3 and by 69.2% over the Greedy heuristic.

Cite

Linghao Zhang, Bo Xu, Haitao Zhao, Jinlong Sun, Jun Xiong, Xianbin Wang, “Token-Aware QoE Optimization for Multimodal AIGC Inference Tasks in Space-Air-Ground Integrated Networks”, IEEE Transactions on Mobile Computing, pp. 1–18, 2026, doi: 10.1109/TMC.2026.3743972.

@article{zhang2026tokenaware,
  author = {Zhang, Linghao and Xu, Bo and Zhao, Haitao and Sun, Jinlong and Xiong, Jun and Wang, Xianbin},
  title = {Token-Aware QoE Optimization for Multimodal AIGC Inference Tasks in Space-Air-Ground Integrated Networks},
  journal = {IEEE Transactions on Mobile Computing},
  year = {2026},
  pages = {1--18},
  doi = {10.1109/TMC.2026.3743972},
  note = {Early Access},
}

What breaks

Generative AI answers are produced token by token, so a user feels two delays: the time to the first token (TTFT) and the time per output token (TPOT). Most offloading work treats a request as a bit-stream and a local-or-remote choice. That misses how prompt length drives both delays, and how much of a long image or video prompt is redundant.

In a remote area with no base stations the choice is real. A small model on a drone answers fast at lower quality. A full foundation model on a LEO satellite answers better and later. And a drone can cache only a few models.

Three-tier SAGIN architecture for multimodal AIGC inference — Plate 1 The three-tier architecture: full foundation models in orbit, cached lightweight models and semantic encoders on UAVs, multimodal requests on the ground. Fig. 1 of the paper.

The idea

Treat everything as tokens. One continuous ratio η compresses the whole input sequence, so it sets the uplink payload and the prefill work together, and therefore TTFT. It also costs quality, and the scheduler can price that cost.

n ~_{u}^{in} = η_{u} (n_{u}^{txt} + n_{u}^{vis}), η_{u} \in [η_{m i n}, 1]

One ratio compresses the input. It sets both the payload on the link and the work in prefill.

The first token has to cross the whole chain in series. Later tokens stream through a pipeline, so they are limited by its slowest stage.

TTFT_{u} = T_{u \to k}^{G2A} + T_{k \to u}^{A2G} + (1 - x_{u}) TTFT_{u}^{UAV} + x_{u} [T^{A2S} + T^{ISL, ↑} + TTFT_{u}^{sat} + T^{ISL, ↓} + T^{S2A}]

TPOT_{u} = {max {TPOT_{u}^{UAV}, T_{k \to u}^{A2G}} max {TPOT_{u}^{sat}, T^{ISL, ↓}, T^{S2A}, T^{A2G}} x_{u} = 0 x_{u} = 1

The first-token delay is a serial sum. The per-token delay is a pipeline bottleneck.

QoE_{u} = w_{q} ln (1 + Q_{u}) - w_{1} TTFT_{u} - w_{2} TPOT_{u}

A concave reward for quality, with diminishing returns, minus penalties for both delays.

Inference pipeline and end-to-end streaming latency model — Plate 2 The inference pipeline and the end-to-end streaming latency model, for UAV-side and satellite-side inference. Fig. 2 of the paper.

Run the model

Send one request to both tiers

RequestUAV modelCompression η Satellite load ρ UAV compute share

UAV · 150 m

First token
Per token
Quality
QoE

LEO · 550 km

First token
Per token
Quality
QoE

This runs the paper’s analytic model live, with parameters from its Tables I and II and supplement (each request type uses the midpoint of its token range). To keep the demo honest and simple it covers one user, with a mean channel, nominal link rates and no inter-satellite routing. It is not the learned scheduler, and these numbers are not results from the paper. In the paper’s simulation the scheduler compresses satellite-bound vision inputs to about η = 0.44.

The system

Three tiers and two timescales. Every 120 s a knowledge-driven planner proposes UAV positions and model caches from recent demand, scores each candidate by rolling out the fast scheduler, and refines by alternating search. Every 2.4 s slot a conditional-diffusion TD3 scheduler (CD-TD3) picks an offload tier and a compression ratio for each user, and closed-form rules allocate bandwidth and compute.

Two-timescale optimization framework — Plate 3 The fast-timescale CD-TD3 scheduler (left) and the slow-timescale knowledge-driven planner (right), coupled through configuration provisioning and value feedback. Fig. 3 of the paper.

What it found

The evaluation is a simulation: 5 UAVs, 30 users, 6 satellites and a library of 10 UAV models, with decode costs set from NVIDIA Jetson AGX Thor rates and quality coefficients from model cards. Results are over 30 held-out configurations. The proposed framework reaches the highest per-user QoE, 31.9% above TD3 and 69.2% above the Greedy heuristic.

Per-user QoE

Proposed1.798
CD-TD3-Fixed1.507
TD31.363
PPO1.174
Greedy1.063

From the evaluation file behind the paper's Table III.

Scheme	QoE	TTFT (s)	TPOT (s/token)	Quality
Proposed	1.798	0.873	0.0272	0.811
CD-TD3-Fixed	1.507	1.225	0.0271	0.833
TD3	1.363	0.806	0.0522	0.727
PPO	1.174	0.917	0.0363	0.638
Greedy	1.063	0.903	0.0368	0.599

Where the gain comes from is visible in that table. Against the fixed-configuration variant the proposed scheme cuts TTFT from 1.225 s to 0.873 s. In the learned behaviour, 35.7% of requests with a cached model still go to the satellite voluntarily, and vision inputs headed there are compressed to about η = 0.44.

Removing one component at a time shows what carries the result. Per-user factorization matters most, then the diffusion actor, then compression.

QoE lost when one component is removed

w/o Per-User−22.8%
w/o Diffusion−15.8%
w/o Compression−9.2%
w/o Refinement−4.6%
w/o ConfigGen−0.9%

Retrained variants, from the evaluation file behind the paper's Fig. 7.

The advantage is largest when resources are tight. With only 10 GB of UAV storage the proposed scheme exceeds Greedy by 67.3%, narrowing to 37.7% at 30 GB.

Aggregate slot QoE against UAV storage

From the sensitivity sweep behind the paper's Fig. 8(b).

The scheduler is cheap enough to run in real time. One decision for all 30 users takes 0.67 ms on average on a single CPU core (Apple M3 Max, PyTorch), and 0.93 ms at the 99th percentile.