TP-MPPO, a scheduling method for wireless edge inference, recorded the highest reward and higher simulated goodput than four benchmark algorithms, according to a preprint. Goodput counts requests completed within their end-to-end latency target, known as a service-level objective or SLO. In this case, the method coordinated task offloading with the allocation of wireless bandwidth.
A narrow definition of success
The paper treats the task as an optimization problem aimed at maximizing goodput while meeting SLO and memory-capacity constraints. It jointly optimizes which tasks are offloaded and how bandwidth is assigned. A request contributes to the main objective only when it satisfies its end-to-end latency target.
TP-MPPO alternates two phases until convergence. Maskable PPO handles task offloading; a closed-form calculation allocates uplink bandwidth, and a greedy procedure allocates downlink bandwidth. For that downlink subproblem, the paper describes the procedure as optimal because selected requests contribute equally to goodput and are prioritized by their smallest bandwidth requirement.
Inside the simulated network
The main simulation considered 3 edge nodes and 15 users randomly deployed within a 300 m radius circular area. Further tests changed node and user counts, GPU provisioning and model size. The comparison set was MPPO, TP-PPO, Rewardless and Heuristic.
To populate the modeled requests, input lengths, SLO targets, maximum output lengths and realized output lengths were independently sampled uniformly from their respective discrete sets. Input lengths were 128, 256 or 512 tokens; SLO targets were 3, 5 or 9 seconds; maximum outputs were 128, 256, 512 or 1,024 tokens; and realized output lengths used 0.5 or 0.75 of the maximum.
What the comparisons showed
TP-MPPO converged within 200 episodes and had the highest reward in the reported comparison. Its reward was 33.3% higher than MPPO, 87.5% higher than TP-PPO and 71.4% higher than Rewardless.
Across various network configurations, the paper reported higher goodput for TP-MPPO than for the benchmarks. At 6 nodes and 20 users, its goodput ratio was 5.4% higher than MPPO, 9.5% higher than TP-PPO, 7.1% higher than Heuristic and 10.2% higher than Rewardless. At 3 nodes and 30 users, the margins were 3%, 7.3%, 6.5% and 7%, respectively.
Model size and GPU resources
In the Llama-30B experiment, TP-MPPO's reported goodput ratio rose from 42.0% to 60.7% as GPU resources increased. The reported pattern differed by model size: lighter models had lower latency and higher throughput and goodput, with a smaller gap between throughput and goodput, while Llama-7B showed a larger gap.
Where the evidence stops
The results are bounded by the simulation design. The evidence covers synthetic configurations, and the listed head-to-head comparison is limited to MPPO, TP-PPO, Rewardless and Heuristic. The reported margins therefore describe performance under those modeled conditions; they do not establish how TP-MPPO would perform on untested algorithms, hardware, model families or a deployed wireless system.
The supplied document is labeled arXiv version 1 and dated 26 Aug 2026. It says the work was supported by Shanghai Municipal Science and Technology Commission Foundation grants 25DP1500300 and 24DP1501001.
Paper data and sources
Original title: Goodput Maximization for Large Language Model Edge Inference: A Two-Phase Maskable PPO Approach
Authors: Xiaojing Chen, Qi Zhang, Wei Ni et al.
Journal/Repository: arXiv
Status: Peer-reviewed
First online: 2026-08-26
DOI: 10.1109/lwc.2026.3718001
Original paper · Full text