acceptodds
Under review as a conference paper at ICLR 2027

How Can LLM Agents Become Wireless Communication Algorithm Engineers?

10 views

Abstract

Wireless communication relies on effective algorithms, yet discovering high-performing algorithms remains challenging. The design space is large, while end-to-end feedback provides limited information about which components should be changed. Moreover, coupled components can change each other's utility as the algorithm evolves, making earlier feedback outdated. We benchmark current LLM-based wireless algorithm discovery agents across four wireless algorithm design tasks and find an important limitation in their search behavior. Existing agents can explore new designs and improve performance through experimentation, but they tend to accumulate and tune algorithm components rather than question whether previously accepted components are still necessary. We find that effective wireless algorithm discovery therefore requires not only exploring new structures, but also proactively and strategically challenging the current design by removing, replacing, or simplifying inherited components to test whether they still contribute to performance. To enable this behavior, we propose Breathe, a wireless algorithm discovery framework that explicitly coordinates when an agent should explore new designs and when it should challenge the current algorithm structure. Across all wireless algorithm design tasks, Breathe consistently improves over the state-of-the-art agent system under the same experimental budget, achieving a task-level macro-average performance improvement of 20.35%.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
LLM Agents, Autonomous Research, Wireless Algorithm Design

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_BspgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates LLM agents for wireless algorithm discovery under iterative experimentation. The authors find a key gap: existing LLM agents tend to accumulate and tune algorithm components but rarely question whether current accepted components are useful when the surrounding stages change. To address this gap, the authors propose a novel framework, namely BREATHE, which explicitly coordinates parallel exploration with periodic challenges to inherited algorithm components, thereby enabling the agent to develop, reassess, simplify, re-explore, and revise the structure as the algorithm evolves. To evaluate BREATHE’s performance, the authors test four wireless tasks (OFDM synchronization and tracking, channel estimation, channel equalization, MU-MIMO precoding) and three orchestrator models (Qwen3.6-27B, Llama-4-Scout-109B, or GPT-OSS-120B). Experimental results show that BREATHE can achieve better performance than existing agent baselines.

Critical strengths
  1. The topic of this paper—evaluating components of wireless systems through iterative “parallel exploration” and “component challenging”—is promising for improving communication performance. This topic is timely and of practical value.
  2. The experiments are comprehensive. The authors consider four wireless tasks (OFDM synchronization and tracking, channel estimation, channel equalization, and MU-MIMO precoding), and employ three orchestrators. Extensive experimental results validate the superiority of the BREATHE framework over existing agent baselines across diverse tasks.
Critical weaknesses
  1. The introduction of the BREATHE framework in Section 4 lacks clarity, particularly the implementation details of "component challenge" and “component recheck.” For example, how do we ensure that a candidate edit has removed a component without incurring unintended side effects? The relevant parts are not sufficiently explained.

  2. The BREATHE framework relies on LLMs to perform component evaluation and "challenging," rather than explicit mathematical analysis. While LLMs can enhance certain aspects of wireless systems, this framework’s reliability cannot be guaranteed.

Questions for authors

After reading this manuscript, I have four questions as follows:

  1. Section 3.2 evaluates the performance of four methods—direct LLM, coding agent, AutoResearch, and AI Telco Engineer—on four wireless tasks. How reliable and robust are these experimental results? Given that LLMs can produce different outputs even with the same inputs, have the authors taken this issue into account? Further, how does this issue affect the effectiveness and robustness of the BREATHE framework?
  2. The introduction of the BREATHE framework in Section 4 is not clear yet. Could the authors clarify the detailed procedures of BREATHE, particularly the “component challenge” and “component recheck” parts? Besides, how does the framework address the limitations of existing AI Telco Engineer agent methods?
  3. Section 4.2 sets 10-round implementation assignments (3/2/3/2) to successive explore-challenge-reexplore-recheck blocks. Is this setting (3/2/3/2) the optimal one? If not, how do we set the round assignments in practical systems to ensure the effectiveness of BREATHE?
  4. In Table 2 and Fig. 3, the authors compare the performance of BREATHE and other baseline methods. Although BREATHE achieves the best task performance, the costs (e.g., computational cost and memory usage) are not presented. Did the authors conduct experiments under the same resource constraints?

Additionally, I have some comments on the paper writing and organization:

  1. The Introduction of the paper is not clear, particularly the third and fourth paragraphs that introduce the paper design and experimental results. Key innovations and results should be highlighted.
  2. The related work section introduces existing studies of LLM agents for wireless communications, but mainly lists some references without detailed descriptions. Although the authors introduce some works in the Appendix, it is better to clearly introduce important references in the main paper.
  3. The paper’s organization needs improvement. In Section 3.2, Table 1 summarizes the experimental scaffolding of all evaluated methods, including the proposed BREATHE. However, BREATHE is developed in Section 4. It is unsuitable to present the experimental scaffolding before introducing the framework itself.
  4. The location of Table 2 is not appropriate. The results of Table 2 are introduced on page 4, but the table is presented on page 7.
3 comments
AuthorsOpenReview ↗

Response by Authors (Part 1/3)

We sincerely thank the reviewer for their time, careful reading, and constructive comments. We address each concern below.

1. Reliability and robustness of the Section 3.2 results (Q1)

Yes, We took into account that LLMs can produce different outputs even with the same inputs. We handle this at two levels for every method, including BREATHE. Within a search, four workers independently implement each idea (except in the coding-agent baseline), and the best implementation by discovery score represents that idea, reducing reliance on a single implementation. Across searches, we repeat the complete discovery process three times per method, which captures variation in idea generation, implementation, and the orchestrator’s decisions. Each run’s selected algorithm is then frozen and evaluated on 20 unseen seeds.

Under this protocol, BREATHE achieves the best mean among the evaluated research loops in all twelve task–orchestrator comparisons. BREATHE is also more robust: averaged over orchestrators and tasks, its run-to-run standard deviation is 33.9% lower than that of AI Telco Engineer. Furthermore, although individual LLM outputs vary, the agents’ preferences during search remain relatively consistent.

2. LLMs versus mathematical analysis (W2)

BREATHE does not use the LLM as the final judge of whether a wireless algorithm component is useful; it uses the LLM to reason about possible designs, while controlled experiments provide the evidence. Wireless algorithms operate under complex and environment-dependent channel conditions, whose distributions and interactions are often difficult to capture with a tractable analytical model. Consequently, many practical design choices cannot be determined from a closed-form math analysis alone. Instead, human engineers typically combine domain knowledge and approximate reasoning to formulate hypotheses, test them under representative channel conditions, and use experimental feedback to revise the design. BREATHE aims to reproduce this workflow with an agent.

Specifically, the orchestrator uses wireless-domain knowledge to reason about the current algorithm, identify a component worth questioning, and propose an intervention. Workers implement the intervention, while the task evaluator compares the resulting candidate with its parent under matched conditions. Only candidates that achieve a strictly better score are accepted; an LLM-generated rationale alone can never justify a change. After discovery, the selected algorithm is frozen and evaluated on unseen seeds without invoking an LLM.

Mathematical analysis remains complementary to this process. When suitable assumptions yield an analytical characterization, it can explain why a discovered component works and under what conditions. However, whether these assumptions adequately capture the actual channel and whether a component remains beneficial after interacting with the rest of the algorithm often still require empirical validation. Indeed, the poor performance of direct LLM generation in Section 3.2 shows that prior knowledge and reasoning alone are insufficient without feedback from the target environment. BREATHE therefore uses the LLM to bring engineering priors and reasoning into the search, while relying on experiments—not the LLM itself—to validate those hypotheses.

AuthorsOpenReview ↗

Response by Authors (Part 2/3)

3. Detailed procedures of BREATHE (Q2 & W1)

We apologize that Section 4 did not describe these procedures clearly. We clarify them below and will revise Section 4 accordingly.

  • Complete BREATHE workflow. BREATHE starts from the basic baseline. In each exploration round, the orchestrator proposes multiple hypotheses and designs experiments to test them, primarily adding components to help the pipeline handle complex signals. The best candidate replaces the incumbent only if it improves performance, and subsequent exploration builds on the incumbent. All other candidates and results are also saved for the orchestrator to review. During challenge, the orchestrator proposes multiple hypotheses about components in the incumbent and designs separate controlled experiments to test their contributions. Each candidate targets one component; the best improving candidate is retained. This block lasts two rounds, with the second starting from the updated incumbent. Re-exploration then develops this incumbent through multiple hypotheses and experiments. Finally, recheck applies the same controlled testing procedure to identify any harmful components remaining in the final solution; no further exploration follows.

  • Avoiding unintended side effects. Before evaluation, the orchestrator reviews each candidate’s code and the worker’s implementation summary against the parent. Removing or replacing a component may change tensor shapes or interfaces at its boundaries, so workers may make the minor adjustments needed to keep the surrounding pipeline executable; these are part of the intervention. Any other change counts as an unintended side effect, and the orchestrator returns the candidate for reimplementation. Each worker has at most two retries, after which its candidate is discarded. Only candidates that pass this check are scored.

Addressing AI Telco Engineer’s limitation. We find that AI Telco Engineer develops algorithms mainly by tuning and recombining components, but does not reassess inherited ones: across 36 trajectories, none of its 1,800 manager proposals suggested removing an inherited component. Component challenge and recheck address this gap by testing inherited components during search and after re-exploration.

4. Optimality of the 3/2/3/2 allocation (Q3)

Under current models, 3/2/3/2 is among the best schedules we tested, and schedules following the same principle perform comparably. What matters is a general principle, which also tells how to set the rounds in practice.

BREATHE alternates exploration and challenge, and each stage needs enough consecutive rounds: (i) enough exploration before challenging, so that there are developed components to reassess; (ii) enough consecutive challenge rounds to reassess inherited components; and (iii) enough exploration after reassessment to develop the revisions. The 3/2/3/2 schedule (EEECCEEECC; E: exploration, C: challenge) is one implementation of this principle for a ten-round budget.

We evaluated alternative ten-round schedules on four tasks with GPT-OSS-120B as the orchestrator. Schedules satisfying (i)–(iii) (EECCEEEECC, EEEECCEECC, and EECCCEECCC) perform comparably to BREATHE, differing by at most 6%, consistent with the Qwen3.6-27B results in Sec. 4.2. Schedules violating them recover only 21–35% of BREATHE's improvement compared with AI Telco Engineer, and trajectory inspection traces each loss to a violated property. The evenly spaced ECECECECEC (21.04%) violates (i) and (ii), and EEEECEEEEC (34.91%) isolates single challenge rounds, violating (ii). Of two random schedules, ECCECCCEEE (29.77%) concentrates challenges early, violating (i), and EEEECECCCC (33.11%) concentrates them late, violating (iii).

In practice, the rounds can be set by following (i)–(iii) instead of searching for an exact optimum. In our experiments, schedules worked well when they kept at least two exploration rounds before the first challenge block, at least two consecutive challenge rounds, and at least two exploration rounds after it. For larger budgets, more explore–challenge cycles can be added following the same principle.

AuthorsOpenReview ↗

Response by Authors (Part 3/3)

5. Resource constraints and costs (Q4)

All research loops were run under the same resource constraints. BREATHE and the research-loop baselines use matched search budgets of 50 hypotheses and 200 implementation attempts, with the same evaluator and the same per-trial time and memory limits. Since matched implementation budgets do not imply equal token usage or runtime cost, we additionally measured the cost of the discovery process and of the resulting algorithms.

Discovery cost. BREATHE used 15.8% fewer tokens than AI Telco Engineer (106.45M vs. 126.40M) on average for each search.

Cost of the discovered algorithms. We benchmarked the final algorithms on the same RTX A6000 GPU using each task’s original evaluator batch size. The table reports median steady-state latency and incremental peak GPU memory during algorithm execution.

TaskBREATHE LatencyAI Telco Engineer LatencyBREATHE MemoryAI Telco Engineer Memory
Channel Estimation50.01 ms137.18 ms142.24 MiB1699.90 MiB
Channel Equalization3.78 ms8.01 ms20.51 MiB22.84 MiB
MU-MIMO Precoding0.57 ms2.40 ms24.63 MiB26.85 MiB
OFDM Sync./Tracking25.42 ms203.76 ms2.93 MiB2.95 MiB

BREATHE’s final algorithms have lower latency on all four tasks and lower or comparable GPU memory usage, with the largest memory reduction on channel estimation. Together with the token measurements, these results show that BREATHE’s performance gains under matched budgets come with lower discovery and execution costs.

Writing and organization

Thank you for these suggestions. We propose the following revisions.

1. Introduction. We will rewrite paragraphs 3–4 so that the gap, the method, and the results are each stated directly, and highlight the following:

  • Gap. Existing agents improve algorithms by tuning and recombining components, but rarely test whether inherited components are still useful: none of 1,800 AI Telco Engineer manager proposals did so.
  • Key innovation. BREATHE schedules controlled component challenges between exploration blocks, so that successful revisions seed further exploration.
  • Key result. BREATHE achieves the best mean held-out performance in all twelve task–orchestrator comparisons, with a 20.35% macro-average improvement over AI Telco Engineer under the same search budget.

2. Related work. We propose replacing the first paragraph of Related work with this:

LLM agents for wireless communications. TeleQnA evaluates telecommunications knowledge, TelecomGPT develops domain-specific LLMs, and NetLLM adapts LLMs to networking tasks. Other studies address wireless hardware design and network control or replace signal-processing blocks with neural models. For physical-layer algorithm discovery, AI Telco Engineer couples an LLM orchestrator with Sionna to iteratively propose, implement, and evaluate algorithms. Its follow-up extends autonomous design to OTFS and pilot-free OFDM receivers. We analyze whether these searches reassess inherited components as the surrounding pipeline changes. BREATHE introduces controlled component tests and develops successful revisions through further exploration.

Citations will be retained.

3. Table 1 and method ordering. Introducing BREATHE-specific mechanisms before the framework may make the presentation harder to follow. We apologize for this oversight. We will move Table 1 and its definitions of challenge, recheck, and re-exploration to the end of Section 4.2. Section 3.2 will retain only the baseline setup.

4. Table 2 placement. We combined the results to keep comparisons together, as Section 5.1 also uses them for the overall evaluation. However, the reviewer is right that this places the table too far from the earlier discussion. We will split it as shown below: Table A beside Section 3.2 and Table B beside Section 5.1, with a “Best of Table A” row connecting the comparisons.

Table A. Classical methods, LLMs, and coding agents.

CategoryMethodEst. ↓Equal. ↑Prec. ↑OFDM ↑
ClassicalBasic............
Strong............
LLMQwen3.6-27B............
...............
Coding AgentCodex (GPT-5.5)............
...............

Table B. Research-loop comparison.

OrchestratorMethodEst. ↓Equal. ↑Prec. ↑OFDM ↑
—Best of Table A............
Qwen3.6-27BAutoResearch............
AI Telco Engineer............
BREATHE............
..................

If you find these revisions suitable, we will incorporate them into the manuscript. We sincerely appreciate the reviewer’s thoughtful comments and constructive guidance, which have helped us refine our work and clarify it.

Reviewer_1gGsOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces BREATHE, a new paradigm for LLM-based agents for network algorithm design allowing for creative destruction. This new framework introduces a new exploration degree by challenging previously added components instead of only searching for iterative refinements built on the existing set of components. BREATHE is benchmarked against state of the art network algorithm design agentic frameworks.

Critical strengths
  1. The paper addresses a real, fairly general problem LLM-based algorithm discovery. Accumulating components without questioning previous design choices seems like an important limitation.

  2. The overall solution and intuition is simple and elegant, introducing a form a creative destruction into algorithm exploration. The reported improvements over the state-of-the-art agent system are also convincing.

Critical weaknesses
  1. The comparison is mostly limited to LLM-based approaches. It would be interesting to compare BREATHE with more classical search strategies, such as genetic programming or evolutionary search, even though setting a fair and comparable experimental setting could be challenging.

  2. The explore/challenge cycle is rather rigid, with a static schedule. It is unclear whether this is the most effective way to allocate the exploration budget.

  3. It is both theoretically and empirically somewhat unclear whether BREATHE enables substantially deeper structural exploration or mainly improves local search efficiency within a fixed budget. Does it enable more efficient local, fixed budget allocation or does it allow for higher order optimization schemes (ie. removing a component slightly degrades performance for the given step but opens new optimization paths that allow avoiding a current local minimum)?

Questions for authors
  1. Does BREATHE enable second-order improvements, where removing a component opens the way to substantially better algorithm structures, or does it mainly improve local search efficiency? (related to Weakness n.3) It would be an interesting topic to discuss in the paper.

  2. Could a more flexible/adaptive DESTROY scheduler improve performance?

  3. How would BREATHE compare with evolutionary or genetic programming approaches under a similar experimental budget?

4 comments
AuthorsOpenReview ↗

Response by Authors (Part 1/2)

We sincerely appreciate the reviewer’s time and thoughtful feedback. We respond to each comment and question below.

1. Second-order improvements (Q1 & W3)

From our experiments, we observe that BREATHE enables second-order improvements, where removing a component opens the way to substantially better algorithm structures. For the case in Weakness 3, where removal first slightly degrades performance, BREATHE has a limitation, but its replacement action provides a way around it.

Second-order improvements. In a channel-estimation trajectory (Sec. 5.2), removing an inherited hard delay-domain mask reduces discovery-time NVE from 9.8571 to 5.2286. Renewed exploration then introduces error-adaptive soft thresholding and reaches 2.6857. This new design uses sparsity differently: it shrinks taps according to estimation uncertainty instead of excluding them with a fixed support. The agent developed it only after the mask was removed, whereas the AI Telco Engineer search in Sec. 3.3 kept tuning its mask. Overall, re-exploration after challenges produced further improvement in 33 of 36 BREATHE trajectories.

Removals that degrade performance. BREATHE’s acceptance rule is locally greedy: in each round, the best candidate replaces the current algorithm only if it performs better. A removal that slightly degrades performance for the given step is therefore not kept, so BREATHE cannot pass through such a temporarily worse intermediate to escape a local minimum. Accepting the degraded removal anyway does not help, since later rounds tend to add the removed component back as the most direct way to recover performance, and forbidding this would hurt search when the component is genuinely useful.

Replacement offers a way around this limitation. It evaluates removal and substitution as one candidate, with the orchestrator using domain knowledge to choose what could take the component’s place. For example, in a channel-equalization trajectory, removing the inherited pilot-residual loading estimator alone drops the discovery score from 5.0755 to 0.9371. Replacing it with a simpler received-power-based loading rule instead reaches 5.4429 (+7.24%), keeping the surrounding iterative LMMSE framework. BREATHE thus moves past a poor removal-only intermediate without accepting it. As the reviewer suggests, we will add this discussion and the limitation to the paper.

2. Fixed and adaptive scheduling (Q2 & W2)

A more flexible DESTROY scheduler does not consistently improve performance: the difference from the fixed schedule ranged from −3.54% to +1.79%. The reason is that BREATHE’s gains depend on giving exploration and challenge enough rounds, which the fixed schedule already does.

BREATHE alternates exploration (E) and challenge (C), and each stage needs enough consecutive rounds: (i) enough exploration before challenging, so that there are developed components to reassess; (ii) enough consecutive challenge rounds to reassess inherited components; and (iii) enough exploration after reassessment to develop the revisions. The fixed EEECCEEECC schedule is one implementation. Under the same ten-round budget, we evaluated alternative schedules on four tasks with GPT-OSS-120B. Schedules satisfying (i)–(iii) (EECCEEEECC, EEEECCEECC, and EECCCEECCC) performed comparably to BREATHE, differing by at most 6%.

Schedules violating them recovered only 21–35% of BREATHE’s improvement over AI Telco Engineer, and trajectory inspection traces each loss to a violated property. The evenly spaced ECECECECEC (21.04%) violates (i) and (ii): components are challenged before they have developed, and exploration resumes before reassessment is complete. EEEECEEEEC (34.91%) isolates single challenge rounds, violating (ii) and leaving many components untested.

To test whether adaptivity helps beyond this, we implemented a flexible scheduler for exploration and challenge (DESTROY) blocks. Instead of using fixed block lengths, it ends a block early when one of three conditions is met: component reassessment is complete, exploration has converged, or no improvement occurs for more than two consecutive rounds. The saved rounds are reallocated to later blocks, so the total budget stays at ten rounds. Even with this flexibility, the difference from the fixed EEECCEEECC schedule ranged from −3.54% to +1.79%. Once exploration and challenge have enough rounds to satisfy (i)–(iii), adjusting the schedule brings no consistent additional benefit.

AuthorsOpenReview ↗

Response by Authors (Part 2/2)

3. Comparison with evolutionary search (Q3 & W1)

We agree that comparing with evolutionary or genetic programming approaches helps clarify what the LLM orchestrator contributes. A fair comparison faces two difficulties. First, these approaches need a well-defined search space: a fixed set of components and rules for combining them. In our tasks, no such space is available in advance. Components are open-ended code, and new ones appear during search. Second, applying genetic programming directly to receiver code would mutate raw program syntax. Random edits usually break such multi-stage programs, for example through mismatched tensor shapes between stages, so most candidates fail for non-algorithmic reasons.

To address both, we implemented Typed Component Evolution (TCE), a conventional (μ+λ)(\mu+\lambda) evolutionary algorithm. For the search space, TCE uses a fixed library of the 50 components that occur most frequently in our BREATHE and AI Telco Engineer searches, giving it domain priors from both systems. A genotype is a component combination; a phenotype is its code implementation. The evolutionary controller makes all search decisions without an LLM orchestrator; the same GPT-OSS-120B workers only turn each genotype into code, so code-level errors do not dominate the comparison. Evaluation follows BREATHE’s protocol.

TCE uses μ=λ=20\mu=\lambda=20 phenotypes over ten generations. The first generation adds one component to the root algorithm. Each later generation selects parents by size-three tournaments and produces five genotypes: four by mutation (adding, deleting, or replacing one compatible component) and one by crossover. Four workers implement each genotype, giving 20 phenotypes per generation and 200 per task, matching BREATHE’s budget. The best 20 phenotypes from parents and offspring survive; after ten generations, the best one by discovery score is frozen and evaluated on unseen seeds.

TCE runs are reported individually, as one run’s test failure would otherwise dominate the mean.

MethodChannel Estimation ↓Channel Equalization ↑MU-MIMO Precoding ↑OFDM Sync./Tracking ↑
TCE (run 1, seed 20261107)173.9634.2928.4730.558
TCE (run 2, seed 20261108)11.1964.1638.4200.543
TCE (run 3, seed 20261109)12.3244.2188.4330.477
AI Telco Engineer (GPT-OSS-120B)9.089 ± 0.5995.230 ± 0.2378.288 ± 0.3260.511 ± 0.204
BREATHE (GPT-OSS-120B)7.947 ± 0.6425.579 ± 0.1658.505 ± 0.0920.777 ± 0.020
BREATHE (Qwen3.6-27B)6.856 ± 0.7046.199 ± 0.2038.516 ± 0.0630.797 ± 0.101
BREATHE (Llama-4-Scout-109B)8.325 ± 1.7685.791 ± 0.3708.540 ± 0.0530.718 ± 0.027

TCE falls below BREATHE on all four tasks. TCE selects component combinations by fitness alone, without reasoning about their physical consistency or how the components interact. It can combine individually standard components in ways that do not make sense for signal processing. For example, in run 1 of channel estimation, TCE applies delay-domain Top-K tap selection after frequency-domain interpolation of comb-pilot estimates. Interpolation creates spurious delay-domain replicas and broadens the true paths, so tap selection can keep artifacts and discard weak genuine paths. BREATHE’s LLM orchestrator instead reasons about each design before it is implemented, which steers the search toward coherent receiver designs.

In the end, we want to thank the reviewer again for the careful feedback and valuable suggestions, which have helped us improve the study and communicate our contributions more clearly.

Reviewer_1gGsOpenReview ↗

Response to rebuttal

Thank you for your response and for the additional experiments. Your answer about second order improvements is clear. I imagine that, for a complex problem with problematic local minima,a multistep Replacement mechanism with look-ahead could enable "second-order improvements" while staying in the BREATHE framework.

I will update my ranking.

AuthorsOpenReview ↗

Response to Reviewer 1gGs

Thank you for your response. We sincerely appreciate the time and care you have devoted to reviewing our paper and engaging in this constructive discussion. Your thoughtful comments and suggestions have helped us strengthen our work and improve its presentation.

Reviewer_LuruOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies the LLM agents for wireless communication algorithm discovery and identifies the limits of preserving the existing components without validating that they are still efficient under current circumstances. In order to solve this problem, they proposed BREATHE, which follows a manually specified schedule of parallel exploration, controlled component challenges, renewed exploration, and a final component recheck. The validation covers four physical layer tasks and three orchestrator models, improving average performance over the evaluated research-loop baselines, with particularly large gains on OFDM synchronisation and tracking. The paper presents a useful engineering method and informative application results. However, the central method is relatively simple, and the current evaluation lacks sufficient evidence of its contribution over straightforward baseline modifications. The reviewer's main concern lies not in the simplicity alone but also in the absence of evidence establishing why the particular workflow is required.

Critical strengths
  1. This paper identifies a specific gap in current LLM-based wireless algorithm discovery agents. It reports that existing LLM-based agents that leverage prior knowledge to propose initial algorithms and use experimental feedback to improve them through parameter tuning and component recombination, can retain unnecessary or harmful components that limit further performance gains. The analysis of 36 AI Telco Engineer trajectories, where none of 1,800 manager proposals explicitly suggested removing an inherited component, gives a concrete basis for the problem statement. The channel-estimation and OFDM examples further explain how repeated tuning can leave redundancy or harmful interactions untested.

  2. BREATHE was evaluated across varied tasks. The evaluation covers four physical-layer tasks and three orchestrator models, uses matched search budgets, and evaluates frozen algorithms on 20 paired unseen seeds. BREATHE achieves the best mean in all twelve task–orchestrator comparisons, with particularly strong OFDM results.

Critical weaknesses
  1. The methodology contribution appears limited. The core method is a fixed algorithm development schedule (3/2/3/2 allocation across explore, challenge, re-explore, and recheck blocks). This is a reasonable engineering heuristic, but the paper does not sufficiently establish a substantial methodological advancement. The alternative schedule experiments provide some sensitivity analysis, but do not explain why the proposed organization is necessary. A simple method can be contributive and valuable. However, from the reviewer's perspective, the evidence needs to be more clearly established with a more generalisable empirical insight.

  2. The comparison between simple baselines is missing. As the authors are aiming to use BREATHE to solve the design issue, baselines explicitly prompted to perform controlled component ablations should be taken into consideration. This would make it clear whether the gains are contributed by BREATHE's specific design or can be obtained through straightforward changes to the workflow.

  3. More ablation tests should be conducted. The current experimental design shows the utility of the complete workflow rather than the contribution of each individual mechanism. Although BREATHE-extend compares the complete continuation strategy with continued exploration, it still cannot isolate the benefit of controlled ablation from that of restarting or reorganizing search. More ablation tests should be carried out in order to make the conclusions stronger.

  4. The paper could more clearly position BREATHE relative to work on search-policy self-improvement. For example, Dream-RSI (Recursive Self-Improvement through Evolving Worlds, arXiv:2609.14858) explicitly treats the exploration policy as an editable object and improves it using replay of accumulated discovery histories. As BREATHE introduces a manually designed procedure, discussing the distinction would help clarify the scope and contribution of BREATHE.

Questions for authors
  1. How much of the improvement can be achieved by simply modifying the baseline prompts to request controlled component ablations at regular intervals?
  2. Does the proposed schedule outperform evenly spaced or randomly scheduled challenges under the same experimental budget?
  3. What is the individual contribution of restarting the algorithm from the incumbent, independent of component challenges?
  4. How should BREATHE be positioned relative to approaches that optimize the exploration policy itself, such as Dream-RSI?
4 comments
AuthorsOpenReview ↗

Response by Authors (Part 1/2)

We sincerely thank the reviewer for their time and valuable comments. We address the comments and questions below.

1. Comparison with evenly spaced and random schedules (W1 & Q2)

BREATHE follows a general strategy: exploration and challenge alternate iteratively. Exploration develops the algorithm pipeline, challenge reassesses inherited components, and further exploration develops the resulting revisions. For this strategy to work, each stage needs enough consecutive rounds: (i) enough exploration before challenging, so that there are developed components to reassess; (ii) enough consecutive challenge rounds to reassess inherited components; and (iii) enough exploration after reassessment to develop the revisions. The 3/2/3/2 schedule (EEECCEEECC; E: exploration, C: challenge) is one implementation of this strategy for a ten-round budget.

We tested this with current LLM models, evaluating alternative ten-round schedules on four tasks with GPT-OSS-120B as the orchestrator. Schedules satisfying (i)–(iii) (EECCEEEECC, EEEECCEECC, and EECCCEECCC) perform comparably to BREATHE, differing by at most 6%.

Schedules violating them recover only 21–35% of BREATHE's improvement compared with AI Telco Engineer, and trajectory inspection traces each loss to a violated property. The two evenly spaced schedules fall short: ECECECECEC (21.04%) violates (i) and (ii), as components are challenged before they have developed and exploration resumes before reassessment is complete; EEEECEEEEC (34.91%) isolates single challenge rounds, violating (ii) and leaving many components untested. Of the two random schedules, ECCECCCEEE (29.77%) concentrates challenges early, violating (i), so few developed components remain to reassess; EEEECECCCC (33.11%) concentrates them late, violating (iii), so little exploration remains to develop the revisions.

These results show that satisfying (i)–(iii) matters more than the exact allocation. Under current models, 3/2/3/2 is among the best schedules tested, consistent with the Qwen3.6-27B results in Sec. 4.2, and schedules keeping this balance perform comparably. For other models or budgets, the allocation can be adjusted following (i)–(iii).

2. Prompting baseline for ablations (W2 & Q1)

To test whether BREATHE’s gains can be obtained by simply prompting the baseline, we modified AI Telco Engineer to explicitly instruct the orchestrator to perform ablations at regular intervals. We ran this prompt-only variant three times on each of the four tasks with both Qwen3.6-27B and GPT-OSS-120B (24 runs in total). This prompt-only baseline recovered, on average, only 17.96% of BREATHE’s improvement over the original AI Telco Engineer baseline. The table reports means ± standard deviations across three independent runs.

OrchestratorMethodChannel Estimation ↓Channel Equalization ↑MU-MIMO Precoding ↑OFDM Sync./Tracking ↑
Qwen3.6-27BPrompt-only8.212 ± 1.9836.062 ± 0.1188.271 ± 0.1830.495 ± 0.134
AI Telco Engineer8.895 ± 1.8186.078 ± 0.1108.123 ± 0.2140.389 ± 0.183
BREATHE6.856 ± 0.7046.199 ± 0.2038.516 ± 0.0630.797 ± 0.101
GPT-OSS-120BPrompt-only8.878 ± 0.8325.305 ± 0.1828.295 ± 0.2310.555 ± 0.116
AI Telco Engineer9.089 ± 0.5995.230 ± 0.2378.288 ± 0.3260.511 ± 0.204
BREATHE7.947 ± 0.6425.579 ± 0.1658.505 ± 0.0920.777 ± 0.020

Trajectory inspection reveals two failures. First, interventions are not performed correctly: in 23 of the 24 trajectories, at least one challenge removed or replaced several components at once, confounding their effects so that the outcome cannot be attributed to the designated component. Second, the challenge budget is inappropriately allocated: fourteen runs settled into ECECECECEC, six into EEEECEEEEC, and four performed only a single late challenge round. These patterns cut short development or reassessment.

The gap lies not in requesting ablations but in carrying them out correctly and effectively, which prompting alone cannot simply achieve. BREATHE enforces both requirements: each candidate edit is checked against its parent for changes beyond the designated component before evaluation, and the schedule provides consecutive exploration and challenge blocks long enough for development and reassessment, followed by further exploration to develop the revisions.

AuthorsOpenReview ↗

Response by Authors (Part 2/2)

3. Contribution of restarting alone (W3 & Q3)

To isolate the contribution of restarting, we ran a restart-only variant that keeps BREATHE’s schedule but replaces the challenge rounds (R4–5 and R9–10) with ordinary exploration, while still restarting exploration from the current incumbent at R6, using Qwen3.6-27B as the orchestrator with three independent runs per task. The table reports mean performance. Gains are measured relative to AI Telco Engineer, with positive values indicating improvement for both higher- and lower-is-better metrics.

TaskAI Telco EngineerBREATHERestart-onlyBREATHE GainRestart-only GainBREATHE Gain Retained
OFDM Tracking ↑0.3890.7970.381+0.408−0.008<0%
Channel Estimation ↓8.8956.8568.658+2.039+0.23711.6%
Channel Equalization ↑6.0786.1996.030+0.121−0.048<0%
MU-MIMO Precoding ↑8.1238.5168.113+0.393−0.010<0%

Restarting alone contributes little: it retains 11.6% of BREATHE’s gain on channel estimation and does not improve over AI Telco Engineer on the other three tasks. Trajectory inspection shows why. Restarting only moves the starting point of search to the current best algorithm. That algorithm still contains its inherited components, including harmful ones, so later exploration keeps building on them. Challenges are what remove such components and change the structure that exploration starts from. For example, in channel estimation, BREATHE’s challenge removed an inherited hard delay-domain mask, and renewed exploration developed error-adaptive soft thresholding (Sec. 5.2).

Re-exploration nonetheless matters once that structure has been revised: in the full BREATHE workflow, it produced further improvement in 33 of 36 trajectories. The two mechanisms are therefore complementary. BREATHE’s gains cannot be simply obtained by restarting search while leaving inherited components unexamined.

4. Positioning relative to Dream-RSI (W4 & Q4)

The key difference is how each method reaches a better algorithm. Dream-RSI improves the exploration policy by replaying its accumulated discovery history, so it learns from algorithms that search has already produced. BREATHE’s challenge instead creates a new algorithm by removing or replacing one component of the current algorithm. This matters because a component that helped earlier may become unnecessary or harmful after other components change, and the better algorithm without it may never have appeared in the history.

For example, suppose search builds a channel estimator: it starts with component AA, then adds BB, then adds CC, giving AA, A+BA+B, and A+B+CA+B+C. Each step improves performance, so all three algorithms enter the history. Once BB and CC are present, AA may become harmful, and the best algorithm is B+CB+C. B+CB+C never appeared during search, so returning to an earlier algorithm cannot give it. BREATHE reaches B+CB+C by challenging AA: it removes AA from A+B+CA+B+C while keeping BB and CC, and compares the result with A+B+CA+B+C.

The two methods are complementary. Each challenge adds a measured result on one inherited component to the history, which could make replay-based policy evaluation more informative. We will cite and discuss Dream-RSI in the related work to clarify the scope of our contribution.

In the end, we sincerely thank the reviewer again for the thoughtful and constructive suggestions, which have helped us strengthen both our work and its presentation.

Thank you for the responses and the additional experiments. I would like maintain my rating as it is already positive.

AuthorsOpenReview ↗

Response to Reviewer Luru

Thank you for your response and for your positive assessment. We sincerely appreciate the time and care you have devoted to reviewing our paper and engaging in this discussion. Your thoughtful comments and suggestions have helped us strengthen our work and clarify its contributions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.