AI agentic research refines cell-edge power control algorithm

Why this matters

This paper examines what happens when an AI agent takes over research design, using a practical and highly relevant problem as a testbed. Researchers repeatedly tested the classic Akash benchmark and leaderboard improvements by entrusting them to an AI coding agent, and compressed the process into 81 autonomous experiments over 26 hours. The highlight is not just the feasibility of autonomous research, but that a trustworthy design actually worked. This study was submitted to IEEE Globecom Workshops as arXiv 2608.26093.

Research background: cell-edge power control and design challenges

Minimizing the transmit power of cell-edge users in cellular networks is a long-standing challenge. Multi-cell interference power control generally exhibits nonconvex, nondifferentiable, and strong NP-hard characteristics. To address this, researchers built an Akash agent from scratch, elevated the loss function, and tuned the learning rate. Because cost is high and repetition is cumbersome, automating the design process is highly practical. Previous studies built on learnable pipelines, but the architecture, loss function, and learning rate were still human-dependent.

Key method: autoresearch protocol covering the entire design layer

The researchers devised an autoresearch protocol. Here, the AI agent modifies only a mutable train.py and, within a fixed computational budget, decides whether to commit or revert each change using a single scalar metric. In this study, the agent managed the Akash benchmark, input representations, output binarization, loss functions, and task-sampling laws. The target was sum-least-percentile-rate power control, a stricter objective than previously worked on models. The team evaluated 17 (K, percentile) grid configurations and maintained logs in an append-only manner after each experiment.

Trustworthiness and safety design

For autonomous research to be practically trustworthy, the evaluator itself must not be easily altered. The researchers employed three safeguards. First, the evaluator prepare.py is SHA-256 pinned so the agent cannot modify the judge. Second, an inference contract with a 10-time prediction buffer and output shape guard is in place. Third, a falsifier registered before each experiment checks whether results match expectations; if unexpected results appear, it logs them and reverts the change. This design reduces the burden of making self-improvement systems testable quickly. The researchers also discovered sample-mismatch evasion during experiment 63 and disclosed it through responsible disclosure.

Results: 99.5% and 600-fold speedup

81 autonomous experiments ran over about 26 hours, and the agent surpassed the initial Akash baseline by 94%. The final champion reached 99.5% of the converged minorization-maximization reference, and the single-parallel setting matched all network sizes and min-rate targets. On an Apple M2 Pro, the entire grid search took 2.52 seconds, whereas the same search on a mainframe took 1583 seconds, yielding roughly a 600-fold speedup. Moreover, in the minimum min-rate setting with Kq=1, the model reproduced the theoretical max-min optimal allocation.

References and implications

This research shows that meaningful results are possible when AI is entrusted with the entire research design process, going beyond AutoML and NAS. Particularly, the structural optimization approach discovered by the agent is not just a tuned constant, but a generalizable pattern. Detailed negative results and per-benchmark progress are being tracked, showing that the trajectory of autonomous research is becoming simpler to understand.

Source

• arXiv: Agentic Autoresearch for Cell-Edge Power Control
• PDF: arXiv 2608.26093 PDF
• Code: autoresearch-percentile-optimization