Pith. sign in

REVIEW 3 major objections 4 minor 108 references

Structure-aware program embeddings, a learned mutation controller, and online language-model fine-tuning turn evolutionary search into a directed process that recovers closed-form oscillator equations.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-14 14:05 UTC pith:XRYCC5UN

load-bearing objection Strong oscillator numbers from a clean GNN+SAC+GRPO stack on OpenEvolve, but the abstract's parent-selection claim is false and the causal story is unablated. the 3 major comments →

arxiv 2607.10127 v1 pith:XRYCC5UN submitted 2026-07-11 cs.LG cs.AI

GAE: Graph-Augmented Evolution for Scientific Discovery via Reinforcement Optimization

Xuanzhou Chen , Taoli Cheng This is my paper
classification cs.LG cs.AI
keywords graph-augmented evolutionsymbolic regressionLLM program searchrelational graph neural networkonline reinforcement learningnonlinear oscillatorsscientific discovery
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Evolutionary program search that uses a language model as a mutation operator is limited by blind sampling of parents, sparse whole-program fitness signals, and a frozen mutation distribution. This paper proposes Graph-Augmented Evolution: each program is parsed into a typed computation graph and embedded by a relational graph network; a reinforcement-learning meta-controller then chooses mutation-type hints from those embeddings; and an online group-normalized policy update continuously adapts the language model to high-fitness structural edits. On symbolic regression for nonlinear damped oscillators, the coupled loop reaches lower in-distribution and out-of-distribution error than static language-model evolutionary baselines under a matched iteration budget, recovering interpretable closed-form physical expressions. A sympathetic reader cares because the same bottlenecks appear whenever evaluation is expensive and the search space of programs is open-ended: making mutation structure-aware and adaptive can convert largely stochastic exploration into a self-improving trajectory toward scientific models.

Core claim

The paper claims that three co-adaptive components—a relational graph neural network that embeds programs as typed computation graphs, a discrete Soft Actor-Critic controller that selects mutation types from those embeddings, and an online group-relative policy optimization loop that updates the language-model mutation operator—together resolve reward sparsity, uninformed selection, and static mutation, enabling efficient discovery of closed-form equations for nonlinear oscillators with state-of-the-art out-of-distribution accuracy under a fixed compute budget.

What carries the argument

Graph-Augmented Evolution: a shared relational GNN produces structural embeddings that condition a Discrete SAC mutation-type policy and supply context for online GRPO fine-tuning of the language model, so embeddings, mutation hints, and the generation distribution co-adapt during search while remaining gradient-decoupled.

Load-bearing premise

The performance gains are attributed to structure-aware mutation selection and online model updates even though parents remain uniformly sampled and the three components are not fully isolated under identical multi-run budgets.

What would settle it

Under the same 150-iteration multi-seed protocol, disable each pillar in turn (frozen language model, uniform mutation hints, ablated or stale graph embeddings) and test whether the reported in-distribution and out-of-distribution NMSE advantages over static evolutionary baselines vanish.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Whole-program scientific search can replace rank-only or uniform sampling with mutation policies conditioned on typed program graphs.
  • Test-time group-normalized rewards can keep sharpening the language-model edit distribution without stopping evolution.
  • Recovered closed-form models remain human-interpretable and can extrapolate better out of distribution than static baselines.
  • Any domain whose programs admit typed graphs and whose fitness arrives only at full evaluation can reuse the same three-pillar loop once node and relation vocabularies are redefined.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because parents stay uniform, a large share of the reported gain may come from full-weight language-model updates rather than the graph-conditioned selector; a clean ablation would settle how much structure is necessary.
  • Redefining the typed graph for other expensive scientific code spaces (numerical schemes, control laws, material models) would test whether the same directed trajectory generalizes beyond oscillator symbolic regression.
  • If the graph surrogate becomes accurate enough, it could reject low-promise candidates before full numerical evaluation and further amortize search cost.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes GAE, a three-component extension of OpenEvolve for LLM-guided evolutionary program search. A relational GNN encodes Python AST computation graphs into structure-aware embeddings that are fine-tuned online as a fitness surrogate; a Discrete SAC meta-controller conditions on those embeddings to select a mutation-type hint (not the parent); and an online GRPO loop updates the LLM mutation operator with group-normalized, parent-relative advantages. On the LLM-SR nonlinear-oscillator symbolic-regression task under a fixed 150-iteration budget, GAE (Qwen3.5-35B-A3B) reports mean/best ID log10 NMSE of −6.87/−7.24 and OOD −7.10/−7.45, outperforming single-run OpenEvolve and PACEvolve baselines, and recovers closed-form expressions with physical interpretations.

Significance. If the performance gains are real and attributable to the claimed architecture, the work supplies a concrete, reusable pattern for turning sparse whole-program evaluation into denser structural signals and for adapting both selection policy and mutation operator online. The recovered oscillator equations and the explicit physical reading of each term are a useful demonstration for AI-for-science. The modular design (Appendix B) and the detailed hyper-parameter tables are strengths that would aid re-implementation. The significance is currently limited by the single-task evaluation and by the mismatch between the abstract framing and the implemented parent-selection policy.

major comments (3)
  1. [Abstract / §3.4 / Alg. 1] Abstract, Introduction, and contribution statement claim that the RL meta-controller 'replace[s] blind evolutionary sampling with a directed policy, dynamically selecting optimal parents and mutation directions' and thereby resolves 'structurally blind parent selection.' Section 3.4 and Algorithm 1 line 1 state the opposite: parents are drawn uniformly at random and 'the reinforcement-learning agent does not select parents'; SAC only emits a mutation-type hint. Appendix B confirms parent selection remains uniform in every ablation. This is a load-bearing framing error: the three-bottleneck resolution narrative is not what the method implements.
  2. [Table 2 / §4] Table 2 reports GAE (3 runs) against single-run OpenEvolve and PACEvolve baselines under a 150-iteration budget, while the LLM weights are frozen for the baselines and fully updated online for GAE. No ablation isolates the GNN encoder, the SAC mutation-type policy, or the GRPO loop under matched compute, run count, and LLM-update status. Without those controls it is impossible to attribute the reported ID/OOD gains to the three-pillar architecture rather than to full-weight LLM fine-tuning or the four-stage validation cascade alone.
  3. [§3.4 / Appendix C] The action space of the Discrete SAC controller is defined as |A|=5 imes30 imes18=2700 (mutation type imes target node id imes argument index), yet only the mutation-type component is injected into the LLM prompt; the full tuple is used only on an AST-level fallback path. The paper never reports how often the fallback is taken or whether the large discrete space is effectively utilized. This leaves open whether the learned policy is doing more than a 5-way type classifier.
minor comments (4)
  1. [§1 / §4] Typographical inconsistency: 'Graph-Augmeted Evolution' in the contribution paragraph and 'GAEvolve' in the experimental section versus the title 'GAE'.
  2. [§3.5 / Eq. (8)] Equation (8) states that the KL term 'contributes no gradient' and is 'for monitoring only,' yet β_kl=0.1 is listed as a hyper-parameter; clarify whether the term is ever active.
  3. [Figure 2] Figure 2 trajectory is informative but lacks error bars or multiple-run envelopes; given that Table 2 already reports three runs, a multi-seed plot would strengthen the claim of steady improvement.
  4. [Table 1] The high-budget Table 1 numbers are taken from prior work under different backbones and iteration counts; a short note on comparability would help readers avoid over-interpreting the ranking.

Circularity Check

0 steps flagged

No circularity: empirical evolutionary method with external NMSE fitness; no derivation reduces by construction to its inputs.

full rationale

This is a methods paper proposing GAE (GNN embeddings + Discrete SAC mutation hints + online GRPO LLM updates) on top of an OpenEvolve-style archive for symbolic regression. Fitness is the external evaluator score R = −log10(NMSE) after BFGS constant fitting on held-out trajectories (ID/OOD splits from LLM-SRBench); the GNN is only a surrogate regressor trained by MSE on observed rewards (Eq. 5) to produce embeddings, not a definition of the target physics. SAC reward (Eq. 6) and GRPO advantages (Eqs. 8–9) are likewise computed from those external scores. No equation, uniqueness claim, or “prediction” is equivalent by construction to a fitted parameter or self-defined quantity. Parent selection remains uniform (Alg. 1, §3.4); the abstract’s stronger wording is a claim mismatch, not circular derivation. Self-references to the OpenEvolve backbone and LLM-SR data are ordinary scaffolding and do not load-bear any result. The paper is self-contained against external benchmarks; score 0 is the correct finding.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

As an empirical ML systems paper, load-bearing content is mostly design choices and hyperparameters rather than physical axioms. The central performance claim rests on the assumption that typed AST graphs plus discrete mutation hints plus group-relative LLM updates improve search, on a large set of free hyperparameters, and on the LLM-SRBench oscillator data-generating process. No new physical entities are postulated; the 'invented' objects are the algorithmic modules themselves.

free parameters (5)
  • λ1 novelty weight / λ2 complexity weight = 0.1 / 0.05
    Reward coefficients in Eq. 6 (0.1 and 0.05) chosen by authors; directly shape SAC training signal.
  • GRPO learning rate and clip ε = 1e-6 / 0.2
    AdamW 1e-6 and ε_clip=0.2 control how aggressively the LLM mutator is updated online; not derived.
  • SAC action space size |A|=2700 and warm-up W=32 = 2700 / 32
    5 mutation types × 30 node ids × 18 args; discrete design choice that defines the policy's expressivity.
  • Archive size N=50, group size G=8, GNN interval K=3 = 50 / 8 / 3
    Population and update-frequency knobs that affect both sample efficiency and reported fitness trajectory.
  • Task-score clip bounds rmin,rmax and complexity onset 20 AST nodes = -5,15 / 20
    Hand-set reward shaping in Eq. 6; change the scale of r_task and r_complexity.
axioms (4)
  • domain assumption Local structural similarity in program space implies that small mutations produce gradual fitness changes, so graph embeddings are useful for directing search.
    Stated in Introduction bottleneck (2); without locality the GNN+SAC design has no inductive advantage.
  • domain assumption Whole-program NMSE after BFGS constant optimization is a valid fitness for discovering the true governing ODE of the synthetic oscillator.
    §4 evaluation protocol; assumes the data-generating process is recoverable as a closed-form expression tree over the allowed operator set.
  • ad hoc to paper Group-normalized parent-relative advantages (Eq. 9) plus PPO-clip yield stable online LLM updates without collapsing diversity.
    §3.5 design choice; KL term is monitoring-only, so stability rests on clip and group normalization.
  • ad hoc to paper Python AST typed graphs with the listed node/edge vocabularies adequately represent symbolic-regression programs for GNN message passing.
    §3.3 and Appendix A; Limitations section admits redesign is required for other languages/domains.
invented entities (2)
  • GAE three-pillar architecture (relational GNN encoder + Discrete SAC mutation meta-controller + online GRPO LLM loop) no independent evidence
    purpose: Jointly address reward sparsity, uninformed selection, and static mutation in LLM evolutionary search.
    The paper's primary contribution; modules are algorithmic constructs, not physical entities. Independent evidence is the empirical oscillator results only.
  • Discrete mutation-type action space of size 2700 over (type, node id, argument index) no independent evidence
    purpose: Give SAC a fixed discrete action set that can be injected as LLM prompt hints or applied as AST edits.
    Appendix C; invented for this controller. No external validation outside the paper's runs.

pith-pipeline@v1.1.0-grok45 · 18613 in / 3892 out tokens · 38592 ms · 2026-07-14T14:05:17.299432+00:00 · methodology

0 comments
read the original abstract

Evolutionary program search guided by Large Language Models (LLMs) has emerged as a powerful paradigm for automated scientific discovery. However, current approaches are fundamentally constrained by three bottlenecks: structurally blind parent selection, sparse whole-program evaluation rewards, and static mutation operators that fail to adapt during search. We present GAE (Graph-Augmented Evolution), a framework that resolves these limitations through a tightly coupled, three-pillar architecture. First, a relational graph neural network (GNN) parses programs into typed computation graphs, producing structure-aware embeddings. Second, an RL-optimized meta-controller leverages these embeddings to replace blind evolutionary sampling with a directed policy, dynamically selecting optimal parents and mutation directions based on reward history. Third, an online GRPO fine-tuning loop continuously updates the LLM mutation operator at test-time using group-normalized evaluation rewards, directly aligning the model's generation distribution with high-fitness structural edits. We evaluate GAE on a challenging scientific discovery task: symbolic regression for complex nonlinear oscillator systems. By transforming stochastic search into a directed, self-improving trajectory, GAE efficiently discovers closed-form physical equations, consistently matching or outperforming static LLM-driven baselines and achieving state-of-the-art out-of-distribution performance.

Figures

Figures reproduced from arXiv: 2607.10127 by Taoli Cheng, Xuanzhou Chen.

Figure 1
Figure 1. Figure 1: GAE framework overview. A relational GNN encoder parses each program’s abstract syntax tree into a typed computation graph and produces a structural embedding z, fine-tuned online against observed rewards to track the evolving program distribution. Parents are sampled from a fixed-size elite archive; the parent embedding zα then serves as the state of a Discrete Soft Actor–Critic meta-controller (policy πϕ… view at source ↗
Figure 2
Figure 2. Figure 2: Performance trajectory of the GAE method on the Non￾linear Harmonic Oscillator symbolic regression problem. The curve shows the running-best log10(NMSE) as a function of the number of evolution iterations. Grey dots mark iterations at which a new best solution was discovered. Lower values of log10(NMSE) indicate better fit. putational budget. As shown in [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

108 extracted references · 4 canonical work pages

  1. [1]

    arXiv preprint arXiv:2504.10415 , year=

    Llm-srbench: A new benchmark for scientific equation discovery with large language models , author=. arXiv preprint arXiv:2504.10415 , year=

    Pith/arXiv arXiv

  2. [2]

    2025 , publisher =

    OpenEvolve: an open-source evolutionary coding agent , author =. 2025 , publisher =

    2025

  3. [3]

    arXiv preprint arXiv:1504.04909 , year=

    Illuminating search spaces by mapping elites , author=. arXiv preprint arXiv:1504.04909 , year=

    Pith/arXiv arXiv

  4. [4]

    Uni-mol: A universal 3d molecular representation learning framework , author=

  5. [5]

    2021 , organization=

    Graph neural architecture search , booktitle=. 2021 , organization=

    2021

  6. [6]

    Advances in neural information processing systems , volume=

    Do transformers really perform badly for graph representation? , author=. Advances in neural information processing systems , volume=

  7. [7]

    arXiv preprint arXiv:2203.04810 , year=

    Benchmarking graphormer on large-scale molecular modeling datasets , author=. arXiv preprint arXiv:2203.04810 , year=

    Pith/arXiv arXiv

  8. [8]

    Journal of chemical information and modeling , volume=

    Analyzing learned molecular representations for property prediction , author=. Journal of chemical information and modeling , volume=. 2019 , publisher=

    2019

  9. [9]

    International conference on machine learning , pages=

    Neural message passing for quantum chemistry , author=. International conference on machine learning , pages=. 2017 , organization=

    2017

  10. [10]

    Once-for-All: Train One Network and Specialize it for Efficient Deployment , author=

  11. [11]

    DARTS: Differentiable Architecture Search , author=

  12. [12]

    International conference on machine learning , pages=

    Efficient neural architecture search via parameters sharing , author=. International conference on machine learning , pages=. 2018 , organization=

    2018

  13. [13]

    Proceedings of the aaai conference on artificial intelligence , volume=

    Regularized evolution for image classifier architecture search , author=. Proceedings of the aaai conference on artificial intelligence , volume=

  14. [14]

    arXiv preprint arXiv:1611.01578 , year=

    Neural architecture search with reinforcement learning , author=. arXiv preprint arXiv:1611.01578 , year=

    Pith/arXiv arXiv

  15. [15]

    IEEE Transactions on Evolutionary Computation , volume=

    Unsupervised behavior discovery with quality-diversity optimization , author=. IEEE Transactions on Evolutionary Computation , volume=. 2022 , publisher=

    2022

  16. [16]

    ACM Transactions on Evolutionary Learning , volume=

    Synergizing quality-diversity with descriptor-conditioned reinforcement learning , author=. ACM Transactions on Evolutionary Learning , volume=. 2025 , publisher=

    2025

  17. [17]

    Advances in neural information processing systems , volume=

    D-vae: A variational autoencoder for directed acyclic graphs , author=. Advances in neural information processing systems , volume=

  18. [18]

    Proceedings of the IEEE , volume=

    Taking the human out of the loop: A review of Bayesian optimization , author=. Proceedings of the IEEE , volume=. 2015 , publisher=

    2015

  19. [19]

    International conference on machine learning , pages=

    BOHB: Robust and efficient hyperparameter optimization at scale , author=. International conference on machine learning , pages=. 2018 , organization=

    2018

  20. [20]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Bananas: Bayesian optimization with neural architectures for neural architecture search , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  21. [21]

    arXiv preprint arXiv:2511.23473 , year=

    Thetaevolve: Test-time learning on open problems , author=. arXiv preprint arXiv:2511.23473 , year=

    Pith/arXiv arXiv

  22. [22]

    arXiv preprint arXiv:2504.05108 , year=

    Algorithm discovery with llms: Evolutionary search meets reinforcement learning , author=. arXiv preprint arXiv:2504.05108 , year=

    Pith/arXiv arXiv

  23. [23]

    Proceedings of the Genetic and Evolutionary Computation Conference , pages=

    Policy gradient assisted map-elites , author=. Proceedings of the Genetic and Evolutionary Computation Conference , pages=

  24. [24]

    Proceedings of the Genetic and Evolutionary Computation Conference , pages=

    Map-elites with descriptor-conditioned gradients and archive distillation into a single policy , author=. Proceedings of the Genetic and Evolutionary Computation Conference , pages=

  25. [25]

    arXiv preprint arXiv:2305.13795 , year=

    Proximal policy gradient arborescence for quality diversity reinforcement learning , author=. arXiv preprint arXiv:2305.13795 , year=

    Pith/arXiv arXiv

  26. [26]

    Hugging Face Blog , year =

    OpenEvolve: Open-Source Evolutionary Code Optimization with Real-World GPU Kernel Discovery , author =. Hugging Face Blog , year =

  27. [27]

    , author=

    Lora: Low-rank adaptation of large language models. , author=. Iclr , volume=

  28. [28]

    Chemical science , volume=

    MoleculeNet: a benchmark for molecular machine learning , author=. Chemical science , volume=. 2018 , publisher=

    2018

  29. [29]

    International conference on machine learning , pages=

    Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor , author=. International conference on machine learning , pages=. 2018 , organization=

    2018

  30. [30]

    Icml , volume=

    Policy invariance under reward transformations: Theory and application to reward shaping , author=. Icml , volume=. 1999 , organization=

    1999

  31. [31]

    arXiv preprint arXiv:2402.03300 , year=

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

    Pith/arXiv arXiv

  32. [32]

    arXiv preprint arXiv:2601.10657 , year=

    Pacevolve: Enabling long-horizon progress-aware consistent evolution , author=. arXiv preprint arXiv:2601.10657 , year=

    arXiv

  33. [33]

    arXiv preprint arXiv:2404.18400 , year=

    Llm-sr: Scientific equation discovery via programming with large language models , author=. arXiv preprint arXiv:2404.18400 , year=

    Pith/arXiv arXiv

  34. [34]

    Advances in Neural Information Processing Systems , volume=

    A unified framework for deep symbolic regression , author=. Advances in Neural Information Processing Systems , volume=

  35. [35]

    arXiv preprint arXiv:2506.13131 , year=

    Alphaevolve: A coding agent for scientific and algorithmic discovery , author=. arXiv preprint arXiv:2506.13131 , year=

    Pith/arXiv arXiv

  36. [36]

    arXiv preprint arXiv:2510.14150 , year=

    Codeevolve: An open source evolutionary coding agent for algorithm discovery and optimization , author=. arXiv preprint arXiv:2510.14150 , year=

    Pith/arXiv arXiv

  37. [37]

    Digital Discovery , volume=

    Uncertainty quantification for molecular property predictions with graph neural architecture search , author=. Digital Discovery , volume=. 2024 , publisher=

    2024

  38. [38]

    arXiv preprint arXiv:2509.19349 , year=

    Shinkaevolve: Towards open-ended and sample-efficient program evolution , author=. arXiv preprint arXiv:2509.19349 , year=

    Pith/arXiv arXiv

  39. [39]

    2019 , eprint=

    GraphNAS: Graph Neural Architecture Search with Reinforcement Learning , author=. 2019 , eprint=

    2019

  40. [40]

    2019 , eprint=

    DARTS: Differentiable Architecture Search , author=. 2019 , eprint=

    2019

  41. [41]

    Molecular contrastive learning of representations via graph neural networks , volume=

    Wang, Yuyang and Wang, Jianren and Cao, Zhonglin and Barati Farimani, Amir , year=. Molecular contrastive learning of representations via graph neural networks , volume=. Nature Machine Intelligence , publisher=. doi:10.1038/s42256-022-00447-x , number=

    doi:10.1038/s42256-022-00447-x

  42. [42]

    2024 , eprint=

    MolTailor: Tailoring Chemical Molecular Representation to Specific Tasks via Text Prompts , author=. 2024 , eprint=

    2024

  43. [43]

    Junwei Yang and Kangjie Zheng and Siyu Long and Zaiqing Nie and Ming Zhang and Xinyu Dai and Wei-Ying Ma and Hao Zhou , booktitle=. Mol-. 2024 , url=

    2024

  44. [44]

    arXiv preprint arXiv:2405.10343 , year=

    Unicorn: A unified contrastive learning approach for multi-view molecular representation learning , author=. arXiv preprint arXiv:2405.10343 , year=

    Pith/arXiv arXiv

  45. [45]

    arXiv preprint arXiv:2003.03123 , year=

    Directional message passing for molecular graphs , author=. arXiv preprint arXiv:2003.03123 , year=

    Pith/arXiv arXiv 2003

  46. [46]

    The Journal of chemical physics , volume=

    Schnet--a deep learning architecture for molecules and materials , author=. The Journal of chemical physics , volume=. 2018 , publisher=

    2018

  47. [47]

    Nature Machine Intelligence , volume=

    Geometry-enhanced molecular representation learning for property prediction , author=. Nature Machine Intelligence , volume=. 2022 , publisher=

    2022

  48. [48]

    Journal of medicinal chemistry , volume=

    Pushing the boundaries of molecular representation for drug discovery with the graph attention mechanism , author=. Journal of medicinal chemistry , volume=. 2019 , publisher=

    2019

  49. [49]

    Journal of chemical information and modeling , volume=

    Chemprop: a machine learning package for chemical property prediction , author=. Journal of chemical information and modeling , volume=. 2023 , publisher=

    2023

  50. [50]

    Advances in neural information processing systems , volume=

    Self-supervised graph transformer on large-scale molecular data , author=. Advances in neural information processing systems , volume=

  51. [51]

    Machine Learning: Science and Technology , volume=

    Chemformer: a pre-trained transformer for computational chemistry , author=. Machine Learning: Science and Technology , volume=. 2022 , publisher=

    2022

  52. [52]

    M ol TRES : Improving Chemical Language Representation Learning for Molecular Property Prediction

    Park, Jun-Hyung and Kim, Yeachan and Lee, Mingyu and Park, Hyuntae and Lee, SangKeun. M ol TRES : Improving Chemical Language Representation Learning for Molecular Property Prediction. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024. doi:10.18653/v1/2024.emnlp-main.788

    work page doi:10.18653/v1/2024.emnlp-main.788 2024

  53. [53]

    2023 , eprint=

    SELFormer: Molecular Representation Learning via SELFIES Language Models , author=. 2023 , eprint=

    2023

  54. [54]

    Self-referencing embedded strings (SELFIES): A 100 volume=

    Krenn, Mario and Häse, Florian and Nigam, AkshatKumar and Friederich, Pascal and Aspuru-Guzik, Alan , year=. Self-referencing embedded strings (SELFIES): A 100 volume=. Machine Learning: Science and Technology , publisher=. doi:10.1088/2632-2153/aba947 , number=

    doi:10.1088/2632-2153/aba947

  55. [55]

    2022 , eprint=

    Evolution through Large Models , author=. 2022 , eprint=

    2022

  56. [56]

    2022 , eprint=

    Large-Scale Chemical Language Representations Capture Molecular Structure and Properties , author=. 2022 , eprint=

    2022

  57. [57]

    2020 , eprint=

    Molecular representation learning with language models and domain-relevant auxiliary tasks , author=. 2020 , eprint=

    2020

  58. [58]

    2022 , eprint=

    ChemBERTa-2: Towards Chemical Foundation Models , author=. 2022 , eprint=

    2022

  59. [59]

    SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules , author=. Journal of chemical information and computer sciences , volume=. 1988 , publisher=

    1988

  60. [60]

    2023 , eprint=

    EvoPrompting: Language Models for Code-Level Neural Architecture Search , author=. 2023 , eprint=

    2023

  61. [61]

    Nature , volume=

    Mathematical discoveries from program search with large language models , author=. Nature , volume=. 2024 , publisher=

    2024

  62. [62]

    BMC bioinformatics , volume=

    Convolutional neural network based on SMILES representation of compounds for detecting chemical motif , author=. BMC bioinformatics , volume=. 2018 , publisher=

    2018

  63. [63]

    2020 , eprint=

    ChemBERTa: Large-Scale Self-Supervised Pretraining for Molecular Property Prediction , author=. 2020 , eprint=

    2020

  64. [64]

    2017 , eprint=

    Neural Message Passing for Quantum Chemistry , author=. 2017 , eprint=

    2017

  65. [65]

    Proceedings of the genetic and evolutionary computation conference , pages=

    Evolutionary neural automl for deep learning , author=. Proceedings of the genetic and evolutionary computation conference , pages=

  66. [66]

    International conference on machine learning , pages=

    Large-scale evolution of image classifiers , author=. International conference on machine learning , pages=. 2017 , organization=

    2017

  67. [67]

    arXiv preprint arXiv:1611.02167 , year=

    Designing neural network architectures using reinforcement learning , author=. arXiv preprint arXiv:1611.02167 , year=

    Pith/arXiv arXiv

  68. [68]

    arXiv preprint arXiv:1806.09055 , year=

    Darts: Differentiable architecture search , author=. arXiv preprint arXiv:1806.09055 , year=

    Pith/arXiv arXiv

  69. [69]

    Journal of chemical information and modeling , volume=

    Extended-connectivity fingerprints , author=. Journal of chemical information and modeling , volume=. 2010 , publisher=

    2010

  70. [70]

    Nature Mental Health , volume=

    Brain mechanisms underlying the emotion processing bias in treatment-resistant depression , author=. Nature Mental Health , volume=. 2024 , publisher=

    2024

  71. [71]

    Journal of anxiety disorders , volume=

    Facial emotion recognition in trait anxiety , author=. Journal of anxiety disorders , volume=. 2006 , publisher=

    2006

  72. [72]

    Proceedings of the National Academy of Sciences , volume=

    Early experience is associated with the development of categorical representations for facial expressions of emotion , author=. Proceedings of the National Academy of Sciences , volume=. 2002 , publisher=

    2002

  73. [73]

    Behavior research methods , volume=

    The Chicago face database: A free stimulus set of faces and norming data , author=. Behavior research methods , volume=. 2015 , publisher=

    2015

  74. [74]

    Journal of vision , volume=

    Using psychophysics to ask if the brain samples or maximizes , author=. Journal of vision , volume=. 2015 , publisher=

    2015

  75. [75]

    CoRR , year=

    ID-Booth: Identity-consistent Face Generation with Diffusion Models , author=. CoRR , year=

  76. [76]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

    Face identity-aware disentanglement in stylegan , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

  77. [77]

    2020 , eprint=

    Disentangled and Controllable Face Image Generation via 3D Imitative-Contrastive Learning , author=. 2020 , eprint=

    2020

  78. [78]

    Communications in Statistics - Simulation and Computation , volume =

    Peiliang Xu and Seiichi Shimada , title =. Communications in Statistics - Simulation and Computation , volume =. 2000 , publisher =. doi:10.1080/03610910008813603 , URL =

    work page doi:10.1080/03610910008813603 2000

  79. [79]

    Proceedings of Thirty Sixth Conference on Learning Theory , pages =

    Near Optimal Heteroscedastic Regression with Symbiotic Learning , author =. Proceedings of Thirty Sixth Conference on Learning Theory , pages =. 2023 , editor =

    2023

  80. [80]

    2023 , eprint=

    Perceptual adjustment queries and an inverted measurement paradigm for low-rank metric learning , author=. 2023 , eprint=

    2023

Showing first 80 references.