acceptodds
Withdrawn from ICLR 2027

equiNAS: A Typed DSL for Certified Evolution of Equivariant Architectures

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Neural architecture design has evolved from manual engineering to algorithmic search over predefined spaces, and more recently to LLM-assisted generation of architectural modifications. While LLMs make architecture search substantially more flexible, scientific neural networks often obey hard structural constraints that cannot be reliably enforced by executable code alone. In equivariant models, for example, representation types and geometric transformation laws must remain consistent throughout the computation graph; a candidate may compile and train successfully while silently violating the intended symmetry. We introduce equiNAS, a typed domain-specific language that turns architecture evolution into certified program transformation. Architectural values expose their representation-theoretic content as semantic types, and every mutation—whether proposed by an LLM or a programmatic search policy—is expressed as a typed patch whose preconditions, dimensional consistency, and equivariance contracts are checked before training. This constrains generation to physically valid transformations without requiring the proposer itself to reason perfectly about representation theory. On EquiformerV3, the typed interface verifies 300/300 proposals against 192/300 (64%) for untyped structured patches under 300 matched architectural intentions per arm, with only 103/600 unrestricted code edits passing the finite-sample equivariance probe in an open-ended stress test; the same explicit intermediate representation enables ahead-of-time compilation and a improvement in steady-state training throughput over the production static backend. Searching the certified space also produces parameter-matched architectures with lower mean validation and test error on QM9 across three seeds and improved long-horizon optimization, with additional exploratory transfer gains across molecular properties. Together, these results indicate that typed architectural representations provide a practical interface between flexible LLM-based search and the hard structural requirements of scientific neural networks.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Equivariant Neural Architectures, Neural Architecture Search, Type Systems, Typed Domain-Specific Language, Architecture Evolution, Equivariance Certification

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_bHgMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces equiNAS, a typed DSL for architecture search over equivariant neural networks. Architectures are represented as computation graphs with semantic types encoding representation-theoretic information, and proposed modifications are expressed as typed patches that undergo static verification before training. The framework supports both deterministic and LLM-based mutation proposers.

The authors evaluate equiNAS primarily on EquiformerV3. They report improved proposal validity over untyped editing, a 1.87× training throughput improvement from graph-based compiler optimizations, and lower mean validation and test MAE on QM9. Additional experiments examine long-horizon optimization and transfer to other molecular properties.

Critical strengths
  1. The typed DSL provides a useful interface for enforcing representation and dimension constraints during architecture search. The matched-intention experiment shows a clear improvement in proposal validity.

  2. The implementation is carefully documented, including compilation checks, architecture and protocol hashes, and mutation histories. The reported 1.87× steady-state training speedup also demonstrates a practical benefit of the explicit graph representation.

Critical weaknesses
  1. Limited methodological novelty and generality. The NAS procedure uses standard evolutionary sampling over nine predefined mutation families designed for EquiformerV3. The main architecture results come from the deterministic chooser, while the LLM selects from an existing catalogue without a demonstrated advantage over conventional search. Type-based NAS and equivariance-aware NAS have both been studied previously (e.g., NeuralArTS; Maile et al., 2023). It is unclear what new search capability the typed DSL provides beyond verifying predefined mutations. The proposal-validity comparison also does not isolate type checking from restricting the action space, as acknowledged in Appendix J.

  2. Insufficient evidence for architecture improvements. The main comparison does not cleanly separate architectural changes from implementation differences. At 20k steps on seed 201, the unmodified DSL-compiled EquiformerV3 achieves 0.186 validation MAE, outperforming TypeEvo at 0.189 (Table 5). However, Table 3 compares TypeEvo against the upstream implementation at 0.191. The other seeds and long-horizon experiments lack corresponding DSL-compiled baseline controls. Moreover, the long-horizon improvement is based on a single seed, and the three-seed test improvement is small relative to the reported variance. These results do not yet convincingly establish an advantage from the discovered architecture.

Questions for authors
  1. Can the authors provide comparisons against the unmodified DSL-compiled EquiformerV3 under the same protocol across all seeds and at longer training horizons? The existing comparison in Table 5 reverses the improvement reported for seed 201 in Table 3.

  2. How are the primitive type contracts and mutation families constructed and verified? Can the type system certify architecture modifications outside the nine predefined families, and what would be required to extend it to another equivariant architecture?

  3. What evidence supports the benefit of the LLM proposer over conventional evolutionary search? In particular, how would GLM-5.2 compare against deterministic or random action selection under the same 88-action catalogue and evaluation budget?

Reviewer_5mBUOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors address NAS for equivariant neural networks, which requires additional constraints to preserve mathematical symmetries, which existing NAS approaches do not adhere to. To address this, the authors propose equiNAS, a typed DSL that allows an LLM or search algorithm to select predefined architectural mutations, which are verified before training. This prevents invalid architectures from consuming GPU resources.

Critical strengths

S1. The proposed typed DSL reduces the generation of structurally invalid architectures, preventing unnecessary GPU resource consumption.

S2. The authors provide the experimental code in the supplementary material, including the prompts used in the experiments, thereby improving reproducibility.

Critical weaknesses

W1. TypeEvo achieves only a modest improvement in mean test MAE over EquiformverV3 (0.207 vs 0.215 MAE), with the gain being small relative to the reported standard deviations (0.16 and 0.012). Furthermore, TypeEvo exhibits greater variability and outperforms the baseline only on two or three seeds, with most of the improvement driven by a single seed.

W2. The paper is difficult to follow due to its dense technical language and limited explanations of key concepts. Section 2 introduces numerous technical terms without sufficient elaboration or concrete examples. Illustrative examples would substantially improve the paper's readability and accessibility.

Questions for authors

Q1: If the LLM's hypotheses are only used as contextual information for subsequent proposals, does generating explicit falsifiable hypotheses actually improve the search compared with simply providing previous mutations and their validation scores?

Q2. How much of the improvement in proposal validity comes from the DSL's type-checking capabilities, and how much simply comes from restricting the LLM to a predefined catalog of already-valid modifications?

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

This paper proposes equiNAS, a Neural Architecture Search (NAS) framework built on EquiformerV3. To modify architecture with rotation equivariance, the authors introduce a typed Domain-Specific Language (DSL) that represents neural architectures as computational graphs. Each primitive is associated with representation-type information and equivariance contracts, allowing the framework to verify architectural modifications correctly. The NAS process operates on predefined mutation families for the seed architecture. Mutations are selected by an LLM, which uses feedback from previous evaluations to guide subsequent mutation choices. The authors demonstrate that the typed interface substantially improves the validity rate of generated architectures compared to free Python editing and untyped structured patches. Additionally, the explicit graph representation enables ahead-of-time compilation and computational optimizations, achieving a 1.87× improvement in training throughput. The framework also discovers architectures that outperform the EquiformerV3 baseline on QM9.

Critical strengths

The paper attempts to address an important and relatively underexplored problem: applying Neural Architecture Search (NAS) to equivariant neural networks.

Critical weaknesses
  1. Lack of clarity and consistency throughout the paper. The paper suffers from substantial issues in presentation, terminology, and experimental consistency, making it difficult to assess its main contributions and conclusions. For example, the term TypeEvo\textbf{TypeEvo} appears in the experimental section without a clear introduction or explanation of its relationship to the proposed equiNAS framework. Furthermore, the paper introduces nine mutation families (S1–S9), but their actual architectural modifications are not sufficiently explained in the main text. Several operators, such as FFN scalar gate orientation, channel normalization, and residual bypass, are described only by their names, making it difficult to understand how they modify the underlying architecture. Since these mutations define the search space of the proposed framework, the lack of detailed explanations makes it difficult to assess the methodology and its contributions.

  2. Insufficient validation of baseline performance. The reported QM9 results are worse than those of earlier models, such as ComENet (NeurIPS 2022), raising concerns about whether EquiformerV3 was properly trained. The authors should compare against established QM9 benchmarks or evaluate their framework on datasets where EquiformerV3 has demonstrated strong performance, such as OC20 or OMat24.

  3. Insufficient analysis of NAS and LLM contributions. The paper does not clearly distinguish the contributions of NAS and LLM-guided mutation selection. Since the LLM only selects from predefined mutations, comparisons with random search, rule-based methods, or conventional evolutionary NAS under the same computational budget are needed to justify its effectiveness.

Questions for authors
  1. Clarity of the proposed framework: Could the authors clarify the relationship between equiNAS and TypeEvo and provide more detailed explanations of the nine mutation families (S1–S9), including how they modify the EquiformerV3 architecture?

  2. Baseline performance: Why are the reported QM9 errors considerably higher than those of earlier models, such as ComENet? Could the authors clarify the training setup and provide evidence that EquiformerV3 is adequately trained under the proposed protocol?

  3. NAS and LLM contributions: What is the specific benefit of using an LLM to select predefined mutations? Could the authors compare LLM-guided search against random search or rule-based selection under the same search space and computational budget?

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

Related papers

Discussion (0)

Sign in to comment.