EvoTreeNAD: Genealogy-Guided Evolution for LLM-Driven Neural Architecture Discovery
Lishan Yu, Derek Jiu, Qizhen Lan, Xiaoqian Jiang
cs.NE · cs.LG
Quick verdict
Claims genealogy-guided LLM evolution beats NAS/NAD baselines on CIFAR/MedMNIST; results strong but recipe fairness and reproducibility need scrutiny.
Summary
EvoTreeNAD is a genealogy-guided evolutionary algorithm in which LLM agents discover neural architectures from an empty root, without a supplied seed or hand-specified search space. Each iteration routes through the genealogy using top-percentile family values—the mean of the highest τ-fraction of rewards of a node and its descendants—selects a root-to-node lineage, and an Idea Agent proposes while a Code Agent implements the next architecture. Theoretically, the paper proves stationary variation regimes exist and, under lineage-conditioned variation assumptions, that limiting top-percentile family values equal a regime's high-reward production probability; maximum-family routing is proven pathwise-equivalent to best-of-N greedy continuation. Empirically it reports 2.05±0.06% and 15.09±0.22% test error on CIFAR-10/100, exceeding listed NAS/NAD baselines, and surpasses the strongest listed baseline on all six MedMNIST-v2 tasks. Controlled 100-iteration CIFAR-10 runs favor it over direct generation, greedy continuation, and mean routing. Main limitations: baseline numbers are quoted from original publications under different training recipes, while discovered architectures receive a modern 500-epoch recipe with MixUp/CutMix/DropPath; strategy comparisons use three runs without significance tests; and code is promised but not yet released.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 2
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Starting from an empty root, EvoTreeNAD discovers complete trainable architectures without a supplied seed or hand-specified search space, guided by top-percentile family values over a persistent genealogy.
- Discovered architectures achieve 2.05±0.06% (CIFAR-10) and 15.09±0.22% (CIFAR-100) test error, outperforming the compared NAS and NAD baselines in Table 1.
- On all six MedMNIST-v2 tasks the discovered models surpass the strongest listed baseline (e.g., Vessel3D AUC 0.975±0.008 vs 0.940).
- Theorem 2 establishes that, under lineage-conditioned variation assumptions, sustained top-percentile family values exactly characterize the probability of generating high-reward architectures, while mean routing penalizes weak descendants and maximum routing is irreversibly greedy.
- Controlled CIFAR-10 experiments show EvoTreeNAD beats repeated direct generation, best-of-N greedy continuation, and full-family-mean routing (Table 2).
Strengths
Controlled within-pipeline ablations isolate the routing rule
Table 2 compares four discovery strategies that share the same lineage-conditioned generation and evaluation pipeline (three 100-iteration CIFAR-10 runs each): EvoTreeNAD mean 97.25% vs best-of-N greedy 96.77%, direct generation 96.71%, full-family-mean 96.32%. Because only the family-value rule changes, this is the paper's most convincing evidence.
Serious, fully proved theory tied to the mechanism
Appendix A proves Proposition 1 (maximum-family routing is pathwise equivalent to best-of-N greedy continuation), Proposition 2 (top-percentile value needs a nonvanishing fraction of high-reward descendants; single-reward influence bounded by |δ|/k), and Theorems 1–2 connecting limiting family values to ΛQ([q,b]) via Uτ(ΛQ) = r0 + (μ+−r0)/τ·ΛQ([q,b]), with explicit variation assumptions in A.4.
Unusual cost and process transparency
Broad evaluation and agent-configuration study
Weaknesses
Baseline comparison is recipe-mismatched
No statistical testing for the headline ablation
'No search space' claim is softened by heavy prompt scaffolding
Proxy-validity correlation computed on post-hoc selected nodes
Theory characterizes bookkeeping, not search success
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
EvoTreeNAD (arXiv 2609.29016, cs.NE/cs.LG) is a method paper combining LLM agents with neural architecture discovery. Its claims are fourfold: (1) seed-free, search-space-free architecture construction from an empty root via a persistent genealogy (Sec. 3); (2) a theory of 'genealogy-guided evolution' — stationary variation regimes (Theorem 1) and an exact relation between sustained top-percentile family values and high-reward production (Theorem 2, Appendix A); (3) empirical superiority over listed NAS/NAD baselines: 2.05±0.06% CIFAR-10 and 15.09±0.22% CIFAR-100 test error (Table 1) plus wins on all six MedMNIST-v2 tasks (Tables 3–4); (4) controlled evidence that top-percentile routing beats direct generation, best-of-N greedy continuation, and full-family-mean routing (Table 2).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is EvoTreeNAD?
It is an evolutionary algorithm for neural architecture discovery in which LLM agents (an Idea Agent and a Code Agent) repeatedly propose and implement complete neural networks. All evaluated architectures are stored in a 'genealogy' tree grown from an empty root, and a top-percentile statistic over each node's family rewards decides which branch of the tree to develop next — no seed architecture or manually defined search space is supplied.
How accurate are the architectures EvoTreeNAD discovers?
The paper reports 2.05±0.06% test error on CIFAR-10 and 15.09±0.22% on CIFAR-100, ahead of the compared NAS/LLM-NAD methods listed in its table, and wins over the strongest listed baseline on all six MedMNIST-v2 medical imaging tasks (e.g., 0.962 accuracy on PathMNIST, 0.975 AUC on VesselMNIST3D). Note the comparison baselines were trained under their original, different recipes.
Is code or data available for EvoTreeNAD?
Not at the time of this preprint: the paper only states that code 'will be released on GitHub'. The prompts, hyperparameter tables, and training recipes are described in appendices, but the exact discovered architectures are only verbally characterized, so results cannot yet be retrained from the paper alone.
How much does an EvoTreeNAD discovery run cost?
Per the paper's cost table: a 100-iteration CIFAR-10 run uses roughly 0.27–0.37 GPU-days of evaluation compute plus API costs of under $1 with a GPT-4.1 Idea Agent and a locally hosted gpt-oss-20b Code Agent (about $3–3.74 more if GPT-5 handles code generation). The best CIFAR-10 run's discovery cost is listed as 0.63 GPU-days.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–5 are not analyzable from the extracted text (only captions and axis labels), so visual evidence on genealogy growth and lineage trajectories cannot be checked.
- Appendix F.2 (Code Agent prompt) appears truncated at the end of the extraction; a few prompt details may be missing.
- Baseline numbers in Tables 1/3/4 are quoted from prior publications and could not be verified against sources within this review.
- The main body tables were flattened by text extraction; some cell alignments (e.g., Table 1 GPU-days per row) were reconstructed from surrounding text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.