acceptodds
Withdrawn from ICLR 2027

RevoNAS: Reflective Evolutionary Exploration for Neural Architecture Search

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent progress in leveraging large language models (LLMs) has enabled Neural Architecture Search (NAS) to generate new architecture not limited from manually predefined search space. Nevertheless, LLM-driven generation remains challenging: the token-level design loop is discrete and non-differentiable, preventing constructive feedback from smoothly guiding architectural improvement. These methods, in turn, commonly suffer from mode collapse into redundant structures or drift toward infeasible designs, resulting in unnecessary generation and evaluation costs. We introduce RevoNAS, a reflective evolutionary orchestrator that effectively bridges LLM-based reasoning with feedback-aligned architectural search. First, RevoNAS presents a Multi-round Multi-expert Consensus to transfer isolated design rules into meaningful architectural clues. Then, Adaptive Reflective Exploration adjusts the degree of exploration leveraging reward variance; it explores when feedback is uncertain and refines when the improvement direction becomes sufficiently clear. Finally, Pareto-guided Evolutionary Selection effectively promotes architectures that jointly optimize accuracy, efficiency, latency, confidence, and structural diversity. On NAS-Bench-201, RevoNAS achieves state-of-the-art accuracies of 95.22, 76.38, and 50.72 on CIFAR-10, CIFAR-100, and ImageNet16-120, respectively, while evaluating only five candidate architectures. Further empirical analyses support its effectiveness and practical applicability.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Neural Architecture Search, Large Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_sykEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes RevoNAS, a neural network search architecture based on large language models, designed to address redundancy in the generation of LLM-based architecture search methods. The proposed method consists of three components, namely Multi-round Multi-expert Consensus, Adaptive Reflective Design Exploration, and Pareto-guided Evolutionary Selection. Experiments on NAS-Bench-201 and other benchmarks show that RevoNAS achieves promising results.

Critical strengths
  1. This article addresses a key issue in the NAS based on LLM, particularly the challenges of generating redundant architectures, invalid designs, and low exploration efficiency under a limited computational budget.

  2. This article is well-written and very well-organized.

  3. On the NAS-BENCH201 benchmark, the proposed method achieved highly competitive results with a very small number of architecture queries.

Critical weaknesses
  1. The novelty of the proposed framework is very limited. The three components proposed by the author have already received a considerable amount of research [1-4]. It seems that the author merely combined them.

[1] Yang Z, Zeng W, Jin S, et al. Nader: Neural architecture design via multi-agent collaboration[C]//2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025: 4452-4461.

[2] Ji Z, Zhu G, Yuan C, et al. RZ-NAS: Enhancing LLM-guided Neural Architecture Search via Reflective Zero-Cost Strategy[C]//ICML. 2025.

[3] Chen A, Dohan D, So D. Evoprompting: Language models for code-level neural architecture search[J]. Advances in neural information processing systems, 2023, 36: 7787-7817.

[4] Lu Z, Whalen I, Boddeti V, et al. Nsga-net: neural architecture search using multi-objective genetic algorithm[C]//Proceedings of the genetic and evolutionary computation conference. 2019: 419-427.

  1. The ablation experiment lacks a control group for Multi-expert Multi-round Consensus, making it difficult to determine the role of this component.

  2. The author claims that only a very small number of architecture queries are needed to achieve competitive results. I would like to ask whether, with such a small number of architecture evaluations (for example, 5 in NASBENCH-201), adaptive adjustment of exploration and exploitation is possible.

Questions for authors

See Weaknesses above.

Reviewer_r3m3OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

In this paper, authors propose RevoNAS in order to address redundant and infeasible architectures in LLM-driven NAS. It combines literature-based multi-expert consensus, reward variance controlled reflective exploration, and multi-objective evolutionary selection. With five evaluated architectures per dataset, it reports test accuracies of 95.22%, 76.38%, and 50.72% on CIFAR-10/-100 and ImageNet16-120, respectively. Additional experiments examine component ablations, broader tasks, and architectural transfer. The main contribution is the integration of these mechanisms into a low-budget architecture-design pipeline.

Critical strengths
  1. The proposed method combining reusable architectural inspirations with accuracy, resource costs, and structural novelty is a reasonable approach to reduce unproductive LLM-generated proposals. The parent-child structural similarity measure also makes the novelty objective more concrete than a purely textual diversity criterion.

  2. The experiment results, e.g., in Table 2, shows consistent improvements from combining ARDE and PES across all three classification datasets. The appendix also documents unsuccessful inspirations and implementation failures, providing useful visibility into limitations rather than reporting only successful designs.

Critical weaknesses
  1. Algorithm 1 uses confidence for parent selection, while Appendix B.3 explicitly defines that confidence using test samples. Based on my understanding, the test inputs appear to influence search, the paper must establish that selection actually uses validation data only. Although it's not proof of test-label leakage, it's a material ambiguity in the held-out evaluation. Besides, the mean maximum softmax probability alone doesn't establish robustness, and greater confidence can accompany incorrect predictions.

  2. Table 1 seems mix fixed-space NAS results with unrestricted architectural redesign, thus, exceeding the listed oracle doesn't itself demonstrate a better search algorithm within the benchmark space.

  3. Section 3.3 says instability increases exploration, but Equation 3 decreases exploration as reward variance increases and describes high variance as convergence. Besides, reward variance alone doesn't establish convergence or estimation uncertainty. The central mechanisms seems to be insufficiently justified and experimentally isolated.

Questions for authors
  1. Can authors provide further details regarding which splits supply accuracy rewards, confidence scores, parent selection, and hyper-parameter tuning?

  2. Is Appendix B.3's reference to test sampels a reporting-only definition or the actual selection procedure?

  3. What exactly counts toward the five evaluations? Can author provide per-run accounting of generated, rejected, trained, and selected candidates?

  4. Are there equal-budget controls isolating MMC, adaptive exploration, and Pareto filtering?

  5. What exact settings produced Table 1?

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

This paper proposes RevoNAS, an LLM-based neural architecture search framework that combines multi-agent collaboration with reflective evolutionary search. The framework consists of three components: Multi-round Multi-expert Consensus (MMC), which extracts architectural insights from existing literature; Adaptive Reflective Design Exploration (ARDE), which adjusts exploration and exploitation based on reward variance; and Pareto-guided Evolutionary Selection (PES), which selects promising architectures considering accuracy, efficiency, latency, confidence, and structural diversity. The authors report strong results on NAS-Bench-201, achieving 95.22%, 76.38%, and 50.72% test accuracy on CIFAR-10, CIFAR-100, and ImageNet16-120, respectively, with only five architecture evaluations. Additional experiments investigate search efficiency, architectural diversity, and transferability.

Critical strengths

I think the overall framework is interesting. Combining multi-agent architectural reasoning with feedback-driven evolutionary search is a promising direction, especially given the limitations of manually defined search spaces. The three proposed components also fit together reasonably well, providing a structured approach to architecture generation and refinement.

The improvements over existing LLM-based NAS methods are encouraging. I also appreciate the additional experiments on NAS-Bench-360, different LLM backbones, and architecture transfer, which help demonstrate the potential applicability of the approach.

Critical weaknesses

One thing that caught my attention is that RevoNAS exceeds the reported NAS-Bench-201 oracle accuracy on all three datasets (Table 1). This suggests that the searched architectures may not belong to the original benchmark search space, making direct comparisons with conventional NAS methods potentially unfair. Also, the claimed search budget of only five architectures is somewhat unclear, as the full cost of LLM interactions, unsuccessful generations, and multi-agent consensus is not sufficiently accounted for.

Questions for authors
  1. Could the authors clarify why RevoNAS outperforms the NAS-Bench-201 oracle? Are all methods evaluated under comparable search spaces and training protocols?

  2. Is the confidence score in Eq. (19) computed using test samples during architecture selection? If not, which data split is actually used?

Reviewer_s1sSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This submission, called RevoNAS, uses LLM agents to propose and refine neural network architectures through multi-expert consensus, reflective exploration, and Pareto-based selection. The paper reports promising results across several tasks. However, the benefit of multiple agents (e.g. GPT-4o in the manuscript) over a strong single agent (e.g. Astra-6) remains unclear.

Critical strengths
  1. A concrete architecture design framework. RevoNAS combines multi-expert consensus, feedback-driven exploration, and Pareto-based selection into an iterative design process. It considers accuracy, model size, and latency, making its objectives relevant to practical architecture design.

  2. Useful component analysis. Table 2 examines ARDE and PES separately and together, providing evidence for the contributions of exploration and selection. Table 3 further shows how selection weights affect accuracy, parameter count, and latency, helping explain the resulting trade-offs.

  3. Evaluation across diverse tasks, with additional analyses. Beyond CIFAR and ImageNet16-120, the paper evaluates four NAS-Bench-360 tasks, including fluid-state and genomic sequence prediction. The appendix adds detection and segmentation transfer experiments, comparisons across underlying LLMs, and examples of discovered architectures. These provide a broader view of the framework’s behavior.

Critical weaknesses
  1. The benefit of multiple agents is not established. Table 2 studies exploration and selection, but does not isolate multi-agent collaboration. The key missing comparison is an iterative single agent using the same strong LLM, the same information, and matched token and architecture-evaluation budgets. A single-shot baseline does not answer this question. If a single stronger agent can solve the specific task with less time and less token consumption, why not choose a later and frontier model to sovle the task.

  2. The added contribution over prior work remains unclear. NADER already uses multiple agents for iterative architecture design. Although RevoNAS reports better results, the current ablations do not establish how much multi-expert consensus contributes. Appendix B.1 measures decomposition stability rather than improvements in architecture quality.

Questions for authors
  1. How does RevoNAS compare with an iterative single agent using the stronger LLM, the same information, and matched token and architecture-evaluation budgets?

  2. How much does multi-expert consensus improve the results over a single expert when exploration, selection, and total budgets are held fixed? What benefit does this mechanism add beyond NADER?

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

Related papers

Discussion (0)

Sign in to comment.