Agentic Autoresearch for Cell-Edge Power Control: Radically Redefining the Researcher's Role

Ahmad Khan; Akram Bin Sediq; Sara Azadegi Naeini; Raviraj S. Adve

Source-linked AI summary

Agentic Autoresearch for Cell-Edge Power Control: Radically Redefining the Researcher's Role

Ahmad Khan, Akram Bin Sediq, Sara Azadegi Naeini, Raviraj S. Adve

arXiv:2608.26093v1Aug 26, 2026cs.LGcs.ITeess.SY

TL;DR

The paper asks whether an autonomous agent can replace human specification of learned wireless resource-management algorithms. It applies autoresearch to difficult multicell cell-edge percentile power control, granting the agent authority over the system design and evaluating fixed-budget experiments with safeguards. The resulting controller reaches 99.5% of a converged reference in one pass at roughly 600× lower inference cost, while recovering exact max-min structure at the minimum percentile.

Problem
Wireless learning-to-optimize systems traditionally require humans to specify the architecture, objective, and training recipe, while cell-edge SLqP power control is non-convex, non-smooth, and strongly NP-hard away from its max-min vertex.
Method
The paper uses autoresearch: an agent edits the training script while controlling the architecture, inputs, outputs, loss, and task-sampling law, then retains changes using a hash-pinned evaluator and fixed metric.
Results
99.5% of the converged minorization-maximization reference was reached in one forward pass, with roughly 600× lower inference cost and 94% of the initial gap closed over eighty-one experiments.
Takeaways & Limitations
The campaign recovered interpretable structure: its output parameterization reproduces the exact max-min-optimal allocation at the minimum percentile for every trained-weight value.
Takeaways & Limitations
The findings are simulator-based, budget-dependent, limited to the cell-edge band (0, 0.25], and reflect one stochastic search draw.

Abstract

from arXiv · show

Designing machine learning algorithms for wireless resource management is labour-intensive: the architecture, the loss function and the training recipe are all specified by hand. We demonstrate that this design layer can be surrendered to an autonomous agent in its entirety. We adopt the autoresearch protocol, in which an AI coding agent edits a training script, runs a fixed-budget experiment, and retains or discards the change according to a single immutable metric. We grant the agent authority over the architecture family, the input representation, the output parameterization, the loss function and the task-sampling law, and set it a target chosen for its difficulty: sum-least-percentile-rate power control across a multicell network. The formulation targets cell-edge throughput and is non-convex, non-smooth and strongly NP-hard away from its max-min vertex. Safeguards render the results trustworthy: a hash-pinned evaluator, an enforced inference contract and a pre-registered falsifier per experiment. In eighty-one unattended experiments over twenty-six hours, the agent reached $99.5\%$ of a converged minorization-maximization reference in one fixed-cost inference pass, at roughly $600\times$ lower inference cost, closing $94\%$ of the gap from its first working architecture, with one parameter set serving every network size and percentile target. It recovered provable structure rather than tuned constants: the output parameterization it discovered reproduces the exact max-min-optimal allocation at the minimum percentile, for every value of the trained weights.

I. INTRODUCTION

The paper extends autoresearch from searching designer-fixed spaces to autonomously redesigning learned wireless resource-management systems. It tests this shift on difficult cell-edge percentile power control, where a single amortized model is practically valuable because per-instance optimization is too slow.

Motivation: Autoresearch traditionally searches spaces fixed in advance by human designers, whereas this paper lets an agent redesign the learned system itself.The agent controls the architecture family, input representation, output parameterization, loss function, and task-sampling law.
Motivation: RRM provides a scalar metric, a fast simulator, and a mutable algorithmic artifact for an autonomous research loop.The paper organizes this into an inner weight-training loop and an outer model-, feature-, and objective-redesign loop.
Target problem: SLqP directly optimizes throughput at the weakest user percentile, making it a cell-edge target with industrial relevance.The target is challenging because users are coupled through an order statistic and the problem is non-convex, non-smooth, and strongly NP-hard away from its max-min vertex.
Target problem: A single amortized map is sought because iterative per-instance solvers are too slow for real scheduling timescales.The model is intended to serve network sizes and in-band percentile targets through fixed-cost inference.

B. Contributions of This Paper

The paper contributes a hardened protocol and grants autonomous design authority across the learned system, applying it to multicell percentile power control. Its campaign reports broad deployment, large efficiency gains, recovered structure, and released reproducibility artifacts.

Protocol: The protocol places every layer of a learned system under autonomous design, extending the agent’s mandate to the inductive bias itself.This contrasts with prior wireless agentic work that designs symbolic code or tunes a human-specified pipeline.
Campaign: Eighty-one unattended experiments over twenty-six hours closed 94% of the gap to a converged reference solver at roughly 600× lower inference cost.A single controller serves every K and in-band percentile.
Discovered structure: The discovered output parameterization pins the model to the exact max-min optimum at the minimum percentile for every trained-weight value.The paper presents this as recovered interpretable structure rather than merely tuned hyperparameters.
Reproducibility: The authors release relevant files, the complete experiment log, weights, and scripts needed to reproduce the champion.The paper separates its sections into problem, protocol, discovered solution, results, and lessons for adoption.
Problem formulation: The target problem is multicell downlink power control with B interfering cells, K users per cell, and KB users overall.User channels are collected in H, while allocated powers are collected in p and bounded by Pmax.
Problem formulation: The SLqP objective sums the weakest Kq rates, with Kq = ⌈q KB⌉, and this work focuses on Kq/(KB) ∈ (0, 0.25].At Kq = 1 the problem becomes max-min-rate power control; for Kq > 1 it is non-convex and strongly NP-hard, and generally non-smooth.

A. Protocol

The campaign uses an immutable evaluator, one editable training script, and a research charter governing fixed-budget agent experiments. Each cycle records a falsifiable hypothesis, evaluates the resulting model on a pinned grid, and commits only improvements beyond a calibrated noise band.

Campaign structure: The campaign uses three strictly separated files: an immutable evaluator, an editable training script, and a research charter.The separation defines the judge, the mutable artifact, and the experiment goal and rules.
Campaign structure: The evaluator implements a certified seven-cell wrapped-channel model, scores candidates on a fixed held-out set, and verifies its SHA-256 hash every iteration.The agent cannot edit this judge.
Evaluation: The benchmark contains seventeen (K, percentile) pairings, and HELDOUTSCORE averages model-to-full-power SLqP ratios across pinned realizations.The converged minorization-maximization reference sets the bar at 1.485, while the evaluator enforces a ten-second inference budget over the grid.
Outer loop: Each outer-loop cycle states one hypothesis and a pre-registered falsifier before launching a complete 2000-step Adam training run.The resulting model is then evaluated on the pinned grid, and the outcome is appended to the log.
Retention and exploration: A change is committed only when HELDOUTSCORE improves beyond the calibrated ±0.0005 noise band; otherwise it is rejected.The protocol also limits exploration to at most six architecture families and protects newly opened families from immediate reversion.

B. On the Necessity of the Safeguards

The champion combines classical fixed-point power-control structure with learned components, while safeguards constrain evaluation and preserve the intended fixed-cost inference setting.

Classical scaffolding: The champion contains exact classical algebra alongside trained encoder and scalar output-head components.Teal stages have no trainable parameters, while amber stages carry all trained weights.
Classical scaffolding: The recursion’s all-ones target yields the max-min-rate optimum, and its positive-interference-map structure guarantees convergence from any positive initialization.These properties establish the classical behavior embedded in the learned system.
Classical scaffolding: The Kq-clipped anchor balances users above the full-power threshold while reducing exactly to the max-min solution at Kq = 1.It was constructed from full-power SINRs and their Kqth smallest value.
Classical scaffolding: The campaign’s contribution was placing classical mathematics inside the inference pipeline rather than inventing the underlying fixed-point method.This placement makes the recursion usable both for feature generation and for mapping emitted profiles to powers.
Classical scaffolding: The fixed-point recursion supplies a differentiable, forty-pass route from target SINR profiles to powers and also generates input features.Its fixed-cost form replaces per-instance iterative solver components inside the inference pipeline.
Learned representation: The champion’s 24 input features are closed-form functions of (H, Kq), including order-statistic and set-coupling blocks that encode the SLqP utility’s ordinal structure.None evaluates the rate or SLqP objective, and one parameter set serves every K in the trained range.

2) Learned components:

The encoder combines permutation-equivariant cell-mediated message passing with global attention whose logits use both directions of interference information.

Learned components: The encoder is permutation-equivariant and combines cell-mediated message passing with a global attention path.The message-passing path reflects that powers enter only through per-cell totals.
Learned components: Attention logits are biased per head and round by both aggressor-to-victim and victim-to-aggressor log-gains.Both directions are supplied explicitly because the corresponding interference relations are distinct.
Learned components: The representation preserves directional interference information rather than treating the two cross-cell effects as interchangeable.The passage states that one direction does not determine the other.

3) The clamp and an exactness guarantee:

A cut clamp constrains the emitted profile before the fixed-point power map, yielding an exact max-min guarantee at the minimum percentile.

The clamp: The cut clamp pins every above-threshold user to the threshold before the profile is converted into powers.This clamp was introduced at experiment 49.
Exactness guarantee: For Kq = 1, the model returns the max-min-optimal allocation for every value of the trained parameters.Therefore, the minimum-percentile evaluation column cannot regress with parameter changes.
Exactness guarantee: The clamped profile is uniform, and positive rescaling invariance makes it equivalent to the all-ones profile that produces the max-min optimum.This explains why the allocation is independent of both the emitted profile magnitude and θ.
Exactness guarantee: The forty-pass implementation reproduces the max-min optimum to a relative error of 3.2 × 10^-7 on held-out data.The reparameterization is described as lossless, and direct-output failure modes are structurally unreachable.

4) Training objective:

The champion trains against a normalized SLqP objective plus one-sided distillation, using diverse task sampling while excluding the already guaranteed Kq = 1 case.

Training objective: The champion loss combines the true SLqP objective, normalized by each batch’s full-power SLqP, with a distillation term weighted by αT = 1.0.Distillation uses cached teacher profiles from short local searches initialized at the anchor wclip.
Training objective: Objective normalization was introduced at experiment 2 and produced the campaign’s single largest gain.The stated rationale is to prevent gradients from small-Kq settings from being starved by Kq scaling.
Task sampling: At each of 2000 optimizer steps, eight independent (K, Kq) tasks are sampled and their gradients are averaged before one Adam update.K is uniform on {1, . . . , 10}, while percentile targets vary continuously across the band.
Task sampling: Kq = 1 is excluded from training because Proposition 2 guarantees optimal output for any parameter values.The sampling law therefore focuses optimization on the remaining settings.

5) Task sampling and optimization:

The campaign reached near-reference performance through diagnostic, structure-revealing experiments rather than parameter sweeps. Its final system combines one-pass inference with fixed algebraic refinement and substantially lower cost.

Milestones: 1.4775 versus 1.4850 yielded 99.5% of the reference on the pinned grid after one network forward pass and fixed algebraic iterations.No per-instance optimization was used.
Milestones: Inference over the entire grid took 2.52 s versus 1583 s for the reference solver, a roughly 600× speedup within the ten-second contract.The measurements were obtained on an Apple M2 Pro.
Trajectory: The largest gains came from diagnosing gradient scaling and output-map resolution, not from descent-noise or parameter sweeps.Experiment 2 addressed raw-objective gradient scaling with Kq, while experiment 81 identified output resolution as limiting.
Negative results: Added capacity, objective softening, percentile re-weighting, and distillation repeatedly failed under independent probes, including distillation from the certified reference solver.The self-trained student outscored the distilled model, supporting training across many realizations over imitating a per-instance optimizer.
Architecture progression: Successive architecture families exposed objective coupling, the value of max-min structure as an input, and the need for attention-based all-pairs comparison.The target-SINR reparameterization made the classical policy the zero point of the search space before attention supplied the remaining gains.

A. The Strongest Known Benchmark, and What the Plateau Means

The campaign plateaued at the strongest reported performance for this problem while lowering inference cost, indicating saturation rather than a demonstrated shortfall. That interpretation is bounded by the problem’s classical anchor, simulator-based evaluation, stochastic budget, and cell-edge scope.

Benchmark: Prior transforms and a hand-designed self-supervised learner provide the benchmark context that the campaign approached rather than clearly exceeded.The cited transform methods converged to closely matched values, while the hand-designed learner trailed across six cell-edge settings.
Plateau: The final dozen experiments reverted qualitatively new mechanisms, leaving parameter-level refinements while performance saturated at the strongest level reached on this problem.The resulting system had roughly 600× lower inference cost than methods of comparable quality.
Interpretation: Across related work, convergence toward a classical solution occurs where that solution is already best available, whereas novel structure appears on tasks lacking such a solution.This supports reading classical convergence as diagnostic of the problem rather than as a method limitation.
Scope: The emerged hybrid depends on a closed-form solution at one vertex, prior-work context supplied to the agent, and an inference contract penalizing per-instance iteration.A problem without these anchors may yield a different design.
Scope: All scores are simulator scores with no measured data, and gains within the ±0.0005 noise band are not claimed.The evaluator implements the cited model and is certified against its reference implementation.
Scope: The findings are budget-dependent and stochastic, while conclusions target the cell-edge band (0, 0.25] rather than metrics such as sum-rate.The agent’s log is supported by version history and a pinned evaluator rather than prose alone.

VI. CONCLUSIONS

The paper argues that autonomous autoresearch can redesign wireless learning systems while matching strong benchmarks, and that trustworthy unattended search depends on explicit safeguards and precise task specification. Its conclusions emphasize model choice, randomness, and charter design as practical determinants of the search process.

Within half a percent of the strongest known benchmark, the agent redesigned the system under an immutable evaluator at roughly 600× lower inference cost.The contribution includes one parameter set spanning network sizes and percentile targets, plus a provable exactness guarantee.
The initial design demands the stronger model, whereas a faster model can drive the iterative loop once the protocol is fixed.The evaluator, seed script, and charter were built separately from the harness and outer-loop driver.
A single campaign does not identify a unique or globally best design because sampled hypotheses and early divergences can redirect the search.A comparable campaign using a similar charter and prompts produced a different architecture and lower score.
Unattended search requires firm, unambiguous obligations, including a family cap, falsifier requirement, and prohibition on editing the evaluator.The charter must specify both what the agent must attempt and what it must not do.