pith. sign in

arxiv: 2606.10587 · v1 · pith:TNEVT3YRnew · submitted 2026-06-09 · 💻 cs.LG · cs.AI

Towards Diverse Scientific Hypothesis Search with Large Language Models

Haorui Wang , Parshin Shojaee , Kazem Meidani , Kunyang Sun , Jos\'e Miguel Hern\'andez-Lobato , Teresa Head-Gordon , Jiajun He , Chandan K. Reddy
show 2 more authors
Chao Zhang Yuanqi Du
This is my paper

Pith reviewed 2026-06-27 14:07 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords scientific hypothesis generationlarge language modelsevolutionary searchparallel temperingdiversity in discoverymolecular discoveryequation discoveryalgorithm discovery
0
0 comments X

The pith

A multi-temperature evolutionary framework generates more diverse and higher-quality scientific hypotheses than standard search methods under fixed budgets.

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

The paper claims that typical evolutionary search with large language models for scientific hypotheses collapses diversity because of strong selection pressure favoring optimization. It reframes the problem as one of sampling a set of high-quality alternatives under a limited validation budget rather than hunting for a single optimum. The proposed solution runs hypothesis generation at several temperature levels simultaneously and exchanges information across those levels in a manner drawn from parallel tempering. Experiments across molecular design, equation discovery, and algorithm discovery show that the approach raises both quality and diversity metrics while the resulting candidates hold up when subjected to more costly downstream checks.

Core claim

By treating hypothesis search as a sampling task instead of pure optimization, the evolutionary framework performs searches at multiple temperature levels and enables information exchange between them. This produces sets of hypotheses that are simultaneously higher in quality and greater in diversity than those obtained from conventional single-temperature evolutionary search, and the improvement holds across molecular, equation, and algorithm discovery domains under an identical validation budget.

What carries the argument

The evolutionary framework that searches hypotheses at multiple temperature levels and performs principled information exchange across temperatures, modeled on parallel tempering.

If this is right

  • Higher-quality and more diverse hypothesis sets are obtained in molecular discovery under the same validation budget.
  • The same gains appear in equation discovery and algorithm discovery tasks.
  • Generated candidates remain effective when later subjected to more expensive computational validation procedures.
  • Diversity does not come at the expense of quality when the multi-temperature exchange is used.

Where Pith is reading between the lines

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

  • The temperature-exchange mechanism may transfer to other LLM-driven creative search tasks where mode collapse is a known issue.
  • One could test whether the number of temperature levels or the exchange frequency can be tuned per domain to further increase gains.
  • The sampling perspective suggests that similar multi-level strategies could be applied to non-evolutionary LLM generation pipelines.
  • If the robustness under downstream validation generalizes, the method could reduce wasted effort on hypotheses that fail expensive checks.

Load-bearing premise

That running hypothesis search at multiple temperature levels with information exchange improves exploration without harming convergence.

What would settle it

A controlled comparison in one of the tested domains where the multi-temperature method yields either lower average hypothesis quality or lower diversity than single-temperature evolutionary search when both are given the same number of validation evaluations.

Figures

Figures reproduced from arXiv: 2606.10587 by Chandan K. Reddy, Chao Zhang, Haorui Wang, Jiajun He, Jos\'e Miguel Hern\'andez-Lobato, Kazem Meidani, Kunyang Sun, Parshin Shojaee, Teresa Head-Gordon, Yuanqi Du.

Figure 1
Figure 1. Figure 1: Illustration of our parallel tempered evolutionary al￾gorithm. High-temperature EA explores the hypothesis space faster, while low-temperature EA converges to local minima faster. The swap between the two pools encourages exploration at the lower temperature. Swapping with Metropolis–Hastings provides a cleaner communication mechanism than direct migration: it exchanges individuals only when the move is ap… view at source ↗
Figure 2
Figure 2. Figure 2: Optimization procedure visualization for molecular discovery. (a) Evolution of average JNK3 scores (lines) and the corresponding number of samples (bars) passing diversity-aware selection for MOLLEO and EvoDiverse. (b) Optimization trajectories of average GSK3β and JNK3 scores compared against baselines for molecules satisfying QED > 0.5 and SA < 5.5 from each iteration [PITH_FULL_IMAGE:figures/full_fig_p… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of EvoDiverse and baselines along best-score trajectories in equation discovery (normalized error; lower is better). averaged across datasets within each category and LLM backbones; detailed per-dataset and per-backbone results for more metrics are provided in the Appendix B.3. Results [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: shows that EvoDiverse maintains consistent improvement throughout the search horizon, while baselines exhibit earlier plateaus. The performance gap between EvoDiverse and Ensemble and Island baselines demonstrates that the principled communication is more effective than heuristic-based approach which can lead to either premature or late convergence. We also annotate the solution evolution found by EvoDiver… view at source ↗
Figure 5
Figure 5. Figure 5: Additional results for molecular discovery. (a) Evolution of average GSK3β scores (lines) and the corresponding number of samples (bars) passing diversity-aware selection for MOLLEO and EvoDiverse. (b) Optimization trajectories of average GSK3β and JNK3 scores compared against baselines. Average scores are calculated based on diversity-aware Top-10 molecules from each iteration [PITH_FULL_IMAGE:figures/fu… view at source ↗
Figure 6
Figure 6. Figure 6: Trajectory of explored chemical space for EvoDiverse during optimization for JNK3 and GSK3β scores. Gray dots represent a UMAP projection of Morgan fingerprints for all ChEMBL 36 data. Red dots indicate the centroids of molecules optimized by EvoDiverse from each iteration, illustrating an evolving trajectory through chemical space. For all colored points, increasing brightness denotes later iterations. Th… view at source ↗
Figure 7
Figure 7. Figure 7: Optimization trajectories of average GSK3β and JNK3 scores of EvoDiverse compared against GraphGA. Here, EvoDiverse is an adapted version of the GraphGA method, distinct from EvoDiverse in Main which uses LLM for optimization. (a) Average scores are calculated based on all molecules from each iteration satisfying QED > 0.5 and SA < 5.5. (b) Average scores are calculated based on diversity-aware Top-10 mole… view at source ↗
Figure 8
Figure 8. Figure 8: Diagnostic curves for EvoDiverse on molecular discovery. The adaptive powering factor ξ (top) and the swap acceptance rate (bottom). highest-scoring equation is selected as the discovered governing law. We evaluate this process on LLM-SRBench (Shojaee et al., 2025b), a diverse benchmark spanning physics, biology, chemistry, and materials science. Each task provides standardized train/test splits. All data … view at source ↗
Figure 9
Figure 9. Figure 9: Performance comparison of EvoDiverse and LLM-based baselines against PySR across LLM-SRBench datasets (results averaged over backbones for LLM-based methods). domains despite its strong performance as a non-LLM symbolic regressor, EvoDiverse shows substantial gains across all datasets. Appendix [PITH_FULL_IMAGE:figures/full_fig_p028_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The qualitative embedding visualization of discovered equation program embeddings across different baselines on LLM￾SRBench. Biology (left) and Chemistry (right) are shown as representative qualitative examples. EvoDiverse represents more diverse coverage of the embedding space [PITH_FULL_IMAGE:figures/full_fig_p029_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Diagnostic curves for EvoDiverse on equation discovery [PITH_FULL_IMAGE:figures/full_fig_p033_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Circle packing problem illustration showing the best solution discovered by EvoDiverse (n = 26, sum of radii = 2.5461). recomputed over all unique programs in the evolution trace (for consistency with the Random Swap variant), which yields slightly lower values than the MAP-Elites–archive–based diversity reported in [PITH_FULL_IMAGE:figures/full_fig_p034_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: PCA visualization of final populations using TF-IDF embeddings. Left: Points colored by method and pool: EvoDiverse shows differentiated Cold and Hot pools with partial overlap; Island shows homogenized pools; Ensemble shows disconnected pools. Right: Same projection colored by program scores [PITH_FULL_IMAGE:figures/full_fig_p035_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Diagnostic curves for EvoDiverse on circle packing. The adaptive powering factor ξ (left) and the swap acceptance rate (right) over swap iterations. 1 import numpy as np 2 3 def construct_packing(): 4 n = 26 5 centers = np.zeros((n, 2)) 6 idx = 0 7 8 # Place 4 circles at corners with optimized offset (from top performer Program 1) 9 offset = 0.065 10 corners = [(offset, offset), (1-offset, offset), (offse… view at source ↗
Figure 15
Figure 15. Figure 15: t-SNE visualization of final populations on circle packing using TF-IDF embeddings, complementing the PCA projection in the main paper. 52 steps = [0.04, 0.02, 0.01, 0.005] 53 for step in steps: 54 improved = True 55 while improved: 56 improved = False 57 order = np.random.permutation(n) 58 for i in order: 59 # Try moves in a 5x5 grid for larger steps, 3x3 for smaller 60 num_points = 5 if step >= 0.02 els… view at source ↗
read the original abstract

Large language models (LLMs) are on the rise for accelerating scientific discovery, most recently in advanced tasks such as generating valid scientific hypotheses. Yet in many discovery settings, the goal is not to identify a single best hypothesis since validation can be noisy and expensive, and scientists benefit from a set of high-quality alternative hypotheses that hedge against downstream uncertainty for the best solutions. Nevertheless, commonly used evolutionary search recipes tend to prioritize optimization over exploration in hypothesis generation, and the resulting selection pressure during the search process leads to diversity collapse. Motivated by these limitations, we formulate hypothesis search as a sampling problem, where the objective is to efficiently produce diverse, high-quality hypotheses under a fixed validation budget. Building on this perspective, we propose \ours, an evolutionary framework inspired by the classical parallel tempering algorithm that searches hypotheses at multiple temperature levels and enables principled information exchange across temperatures to improve exploration without disrupting convergence. Across domains including molecular discovery, equation discovery, and algorithm discovery, our approach consistently improves both hypothesis quality and diversity under the same validation budget, and produces candidates that remain robust under more expensive downstream computational validations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The paper formulates hypothesis search with LLMs as a sampling problem rather than pure optimization, and introduces  ours, a parallel-tempering-inspired evolutionary framework that runs searches at multiple temperature levels with cross-temperature information exchange. It claims that this produces higher-quality and more diverse hypotheses than standard evolutionary recipes under a fixed validation budget, with results shown across molecular discovery, equation discovery, and algorithm discovery; the generated candidates are also reported to remain robust under more expensive downstream validations.

Significance. If the empirical improvements and robustness claims hold after detailed verification of the sampling procedure and controls, the work would offer a concrete, reusable technique for mitigating premature diversity collapse in LLM-driven scientific search, directly addressing a practical bottleneck in hypothesis generation pipelines.

major comments (1)
  1. [Abstract] Abstract (paragraph on the proposed framework): the central claim that multi-temperature search with principled information exchange improves exploration without disrupting convergence is presented as load-bearing for the method, yet the manuscript provides no derivation, pseudocode, or ablation isolating the exchange operator, temperature schedule, or acceptance rule, leaving the assumption unverified and preventing assessment of whether the reported gains are attributable to the framework.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive comment on the abstract and the need to verify the framework components. We address this directly below.

read point-by-point responses

  1. Referee: [Abstract] Abstract (paragraph on the proposed framework): the central claim that multi-temperature search with principled information exchange improves exploration without disrupting convergence is presented as load-bearing for the method, yet the manuscript provides no derivation, pseudocode, or ablation isolating the exchange operator, temperature schedule, or acceptance rule, leaving the assumption unverified and preventing assessment of whether the reported gains are attributable to the framework.

    Authors: We agree that the abstract presents the central claim at a high level and that the manuscript as described does not include a derivation, pseudocode, or isolating ablations for the exchange operator, temperature schedule, or acceptance rule. This leaves the attribution of gains unverified in the current version. We will revise the manuscript to add (1) a short derivation section explaining the parallel-tempering motivation and how the exchange and acceptance rules preserve convergence properties, (2) explicit pseudocode for the full procedure, and (3) targeted ablations that isolate each component while holding the validation budget fixed. These will appear in the methods and experimental sections, and the abstract will be updated to reference the new supporting material. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper formulates hypothesis search as a sampling problem and introduces a new evolutionary framework ( exttt{\\ours}) inspired by parallel tempering with multi-temperature search and cross-temperature exchange. No equations, fitted parameters, or self-citations are shown in the abstract or context that reduce the central claims (improved quality and diversity under fixed budget) to inputs by construction. The method is presented as an independent algorithmic proposal whose performance is evaluated empirically across domains, rendering the derivation chain self-contained against external benchmarks with no load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only; no explicit free parameters, axioms, or invented entities are stated. The method implicitly relies on the standard assumption that LLM outputs can be treated as samples from a temperature-controlled distribution and that exchange moves preserve the target distribution, but these are not enumerated.

pith-pipeline@v0.9.1-grok · 5762 in / 1164 out tokens · 14498 ms · 2026-06-27T14:07:39.454179+00:00 · methodology

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

70 extracted references · 2 canonical work pages

  1. [1]

    2025 , eprint=

    Reasoning with Sampling: Your Base Model is Smarter Than You Think , author=. 2025 , eprint=

    2025

  2. [2]

    Proceedings of the National Academy of Sciences , volume=

    Sampling can be faster than optimization , author=. Proceedings of the National Academy of Sciences , volume=. 2019 , publisher=

    2019

  3. [3]

    Conference on Learning Theory , pages=

    Non-convex learning via stochastic gradient langevin dynamics: a nonasymptotic analysis , author=. Conference on Learning Theory , pages=. 2017 , organization=

    2017

  4. [4]

    arXiv preprint arXiv:2512.15567 , year=

    Evaluating Large Language Models in Scientific Discovery , author=. arXiv preprint arXiv:2512.15567 , year=

    Pith/arXiv arXiv

  5. [5]

    Optimization: Methods and Applications, Possibilities and Limitations: Proceedings of an International Seminar Organized by Deutsche Forschungsanstalt f

    Evolution strategy: Nature’s way of optimization , author=. Optimization: Methods and Applications, Possibilities and Limitations: Proceedings of an International Seminar Organized by Deutsche Forschungsanstalt f. 1989 , organization=

    1989

  6. [6]

    1992 , publisher=

    Adaptation in natural and artificial systems: an introductory analysis with applications to biology, control, and artificial intelligence , author=. 1992 , publisher=

    1992

  7. [7]

    Physical Chemistry Chemical Physics , volume=

    Parallel tempering: Theory, applications, and new perspectives , author=. Physical Chemistry Chemical Physics , volume=. 2005 , publisher=

    2005

  8. [8]

    The Annals of Probability , pages=

    Laplace's method revisited: weak convergence of probability measures , author=. The Annals of Probability , pages=. 1980 , publisher=

    1980

  9. [9]

    Recursive stochastic algorithms for global optimization in R\^

    Gelfand, Saul B and Mitter, Sanjoy K , journal=. Recursive stochastic algorithms for global optimization in R\^. 1991 , publisher=

    1991

  10. [10]

    AI for Accelerated Materials Design-ICLR 2025 , year=

    Large language models are innate crystal structure generators , author=. AI for Accelerated Materials Design-ICLR 2025 , year=

    2025

  11. [11]

    Journal of the American Chemical Society , year=

    Generative Design of Functional Metal Complexes Utilizing the Internal Knowledge and Reasoning Capability of Large Language Models , author=. Journal of the American Chemical Society , year=

  12. [12]

    Forty-second International Conference on Machine Learning , year=

    LLM-Augmented Chemical Synthesis and Design Decision Programs , author=. Forty-second International Conference on Machine Learning , year=

  13. [13]

    arXiv preprint arXiv:2509.20988 , year=

    AOT*: Efficient Synthesis Planning via LLM-Empowered AND-OR Tree Search , author=. arXiv preprint arXiv:2509.20988 , year=

    arXiv

  14. [14]

    ACS Central Science , volume=

    SynLlama: generating synthesizable molecules and their analogs with large language models , author=. ACS Central Science , volume=. 2025 , publisher=

    2025

  15. [15]

    arXiv preprint arXiv:2102.09548 , year=

    Therapeutics data commons: Machine learning datasets and tasks for drug discovery and development , author=. arXiv preprint arXiv:2102.09548 , year=

    arXiv

  16. [16]

    Nature Computational Science , ISSN =

    Smileyllama: Modifying large language models for directed chemical space exploration , author=. Nature Computational Science , ISSN =. 2026 , type =. doi:10.1038/s43588-026-00986-y , url =

    work page doi:10.1038/s43588-026-00986-y 2026

  17. [17]

    Journal of Experimental & Theoretical Artificial Intelligence , volume=

    Genitor II: A distributed genetic algorithm , author=. Journal of Experimental & Theoretical Artificial Intelligence , volume=. 1990 , publisher=

    1990

  18. [18]

    Foundations of genetic algorithms , volume=

    Evolution in time and space--the parallel genetic algorithm , author=. Foundations of genetic algorithms , volume=. 1991 , publisher=

    1991

  19. [19]

    Markov chain

    Geyer, Charles J , year=. Markov chain

  20. [20]

    Exchange

    Hukushima, Koji and Nemoto, Koji , journal=. Exchange. 1996 , publisher=

    1996

  21. [21]

    and Wang, Jian-Sheng , journal =

    Swendsen, Robert H. and Wang, Jian-Sheng , journal =. Replica

  22. [22]

    arXiv preprint arXiv:2509.23265 , year=

    CREPE: Controlling Diffusion with Replica Exchange , author=. arXiv preprint arXiv:2509.23265 , year=

    arXiv

  23. [23]

    Frontiers in Robotics and AI , volume=

    Quality diversity: A new frontier for evolutionary computation , author=. Frontiers in Robotics and AI , volume=. 2016 , publisher=

    2016

  24. [24]

    Forty-first International Conference on Machine Learning , year=

    Evolution of Heuristics: Towards Efficient Automatic Algorithm Design Using Large Language Model , author=. Forty-first International Conference on Machine Learning , year=

  25. [25]

    Nguyen and Nakul Rampal and Ali H

    Zhiling Zheng and Oufan Zhang and Ha L. Nguyen and Nakul Rampal and Ali H. Alawadhi and Zichao Rong and Teresa Head-Gordon and Christian Borgs and Jennifer T. Chayes and Omar M. Yaghi , doi =. ChatGPT Research Group for Optimizing the Crystallinity of MOFs and COFs , volume =. ACS Cent. Sci. , month =

  26. [26]

    arXiv preprint arXiv:2512.21782 , year=

    Accelerating Scientific Discovery with Autonomous Goal-evolving Agents , author=. arXiv preprint arXiv:2512.21782 , year=

    arXiv

  27. [27]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    Non-reversible parallel tempering: a scalable highly parallel MCMC scheme , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2022 , publisher=

    2022

  28. [28]

    and Head-Gordon, T

    Brown, S. and Head-Gordon, T. , title =. J Comput Chem , volume =. doi:10.1002/jcc.10181 , year =

    work page doi:10.1002/jcc.10181

  29. [29]

    Drug Discovery Today: Technologies , volume=

    On failure modes in molecule generation and optimization , author=. Drug Discovery Today: Technologies , volume=. 2019 , publisher=

    2019

  30. [30]

    Science , number=

    The method of multiple working hypotheses , author=. Science , number=. 1890 , publisher=

  31. [31]

    Proceedings of the 13th annual conference on Genetic and evolutionary computation , pages=

    Evolving a diversity of virtual creatures through novelty search and local competition , author=. Proceedings of the 13th annual conference on Genetic and evolutionary computation , pages=

  32. [32]

    1970 , publisher=

    The structure of scientific revolutions , author=. 1970 , publisher=

    1970

  33. [33]

    The Structure of Scientific Revolutions , author=

  34. [34]

    arXiv preprint arXiv:2110.14168 , year=

    Training verifiers to solve math word problems , author=. arXiv preprint arXiv:2110.14168 , year=

    Pith/arXiv arXiv

  35. [35]

    Forty-first International Conference on Machine Learning , year=

    Improving factuality and reasoning in language models through multiagent debate , author=. Forty-first International Conference on Machine Learning , year=

  36. [36]

    arXiv preprint arXiv:2508.15260 , year=

    Deep think with confidence , author=. arXiv preprint arXiv:2508.15260 , year=

    Pith/arXiv arXiv

  37. [37]

    International Conference on Machine Learning , pages=

    Fast inference from transformers via speculative decoding , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  38. [38]

    arXiv preprint arXiv:2203.11171 , year=

    Self-consistency improves chain of thought reasoning in language models , author=. arXiv preprint arXiv:2203.11171 , year=

    Pith/arXiv arXiv

  39. [39]

    Advances in Neural Information Processing Systems , volume=

    Aligning large language models with representation editing: A control perspective , author=. Advances in Neural Information Processing Systems , volume=

  40. [40]

    Advances in Neural Information Processing Systems , volume=

    Inference-time intervention: Eliciting truthful answers from a language model , author=. Advances in Neural Information Processing Systems , volume=

  41. [41]

    arXiv preprint arXiv:2309.08600 , year=

    Sparse autoencoders find highly interpretable features in language models , author=. arXiv preprint arXiv:2309.08600 , year=

    Pith/arXiv arXiv

  42. [42]

    Advances in neural information processing systems , volume=

    Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in neural information processing systems , volume=

  43. [43]

    Advances in neural information processing systems , volume=

    Language models are few-shot learners , author=. Advances in neural information processing systems , volume=

  44. [44]

    arXiv preprint arXiv:2310.17022 , year=

    Controlled decoding from language models , author=. arXiv preprint arXiv:2310.17022 , year=

    arXiv

  45. [45]

    arXiv preprint arXiv:2305.18090 , year=

    Chatgpt-powered conversational drug editing using retrieval and domain feedback , author=. arXiv preprint arXiv:2305.18090 , year=

    arXiv

  46. [46]

    Nature , volume=

    Optimizing generative AI by backpropagating language model feedback , author=. Nature , volume=. 2025 , publisher=

    2025

  47. [47]

    arXiv preprint arXiv:2501.09274 , year=

    Large Language Model is Secretly a Protein Sequence Optimizer , author=. arXiv preprint arXiv:2501.09274 , year=

    arXiv

  48. [48]

    NeurIPS 2024 Workshop on AI for New Drug Modalities , year=

    LLMs are highly-constrained biophysical sequence optimizers , author=. NeurIPS 2024 Workshop on AI for New Drug Modalities , year=

    2024

  49. [49]

    Nature , volume=

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

    2024

  50. [50]

    Nature , volume=

    Autonomous chemical research with large language models , author=. Nature , volume=. 2023 , publisher=

    2023

  51. [51]

    Nature Machine Intelligence , volume=

    Augmenting large language models with chemistry tools , author=. Nature Machine Intelligence , volume=. 2024 , publisher=

    2024

  52. [52]

    arXiv preprint arXiv:2511.02824 , year=

    Kosmos: An ai scientist for autonomous discovery , author=. arXiv preprint arXiv:2511.02824 , year=

    Pith/arXiv arXiv

  53. [53]

    The Thirteenth International Conference on Learning Representations , year=

    LLM-SR: Scientific Equation Discovery via Programming with Large Language Models , author=. The Thirteenth International Conference on Learning Representations , year=

  54. [54]

    The Thirteenth International Conference on Learning Representations , year=

    Efficient Evolutionary Search Over Chemical Space with Large Language Models , author=. The Thirteenth International Conference on Learning Representations , year=

  55. [55]

    arXiv preprint arXiv:2107.03374 , year=

    Evaluating large language models trained on code , author=. arXiv preprint arXiv:2107.03374 , year=

    Pith/arXiv arXiv

  56. [56]

    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

  57. [57]

    2025 , publisher =

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

    2025

  58. [58]

    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

  59. [59]

    arXiv preprint arXiv:2511.08522 , year=

    Alpharesearch: Accelerating new algorithm discovery with language models , author=. arXiv preprint arXiv:2511.08522 , year=

    arXiv

  60. [60]

    Advances in Neural Information Processing Systems , volume=

    Symbolic regression with a learned concept library , author=. Advances in Neural Information Processing Systems , volume=

  61. [61]

    arXiv preprint arXiv:2408.06292 , year=

    The ai scientist: Towards fully automated open-ended scientific discovery , author=. arXiv preprint arXiv:2408.06292 , year=

    Pith/arXiv arXiv

  62. [62]

    biorxiv , year=

    Biomni: A general-purpose biomedical ai agent , author=. biorxiv , year=

  63. [63]

    arXiv preprint arXiv:2002.08155 , year=

    Codebert: A pre-trained model for program-ming and natural languages , author=. arXiv preprint arXiv:2002.08155 , year=

    Pith/arXiv arXiv 2002

  64. [64]

    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=

    arXiv

  65. [65]

    Journal of chemical information and modeling , volume=

    ZINC- a free database of commercially available compounds for virtual screening , author=. Journal of chemical information and modeling , volume=. 2005 , publisher=

    2005

  66. [66]

    Advances in neural information processing systems , volume=

    Sample efficiency matters: a benchmark for practical molecular optimization , author=. Advances in neural information processing systems , volume=

  67. [67]

    2: Pushing the frontier of open large language models , author=

    Deepseek-v3. 2: Pushing the frontier of open large language models , author=. arXiv preprint arXiv:2512.02556 , year=

    Pith/arXiv arXiv

  68. [68]

    jl , author=

    Interpretable machine learning for science with PySR and SymbolicRegression. jl , author=. arXiv preprint arXiv:2305.01582 , year=

    Pith/arXiv arXiv

  69. [69]

    Mendez, David and Gaulton, Anna and Bento, A. Patr. Nucleic Acids Research , volume =. 2019 , doi =

    2019

  70. [70]

    Chemical science , volume=

    A graph-based genetic algorithm and generative model/Monte Carlo tree search for the exploration of chemical space , author=. Chemical science , volume=. 2019 , publisher=

    2019