LLM-Guided Test-Time Discovery of Quantum-Chemical Approximation Algorithms

Masaki Adachi; Masaya Hagai; Shuhei Kurita; Tomoya Murata; Yuta Suzuki

arxiv: 2606.20729 · v1 · pith:GOYFWDADnew · submitted 2026-06-17 · ⚛️ physics.chem-ph · cond-mat.mtrl-sci· cs.AI· cs.LG

LLM-Guided Test-Time Discovery of Quantum-Chemical Approximation Algorithms

Masaya Hagai , Yuta Suzuki , Tomoya Murata , Shuhei Kurita , Masaki Adachi This is my paper

Pith reviewed 2026-06-26 19:16 UTC · model grok-4.3

classification ⚛️ physics.chem-ph cond-mat.mtrl-scics.AIcs.LG

keywords quantum chemistryapproximation algorithmslarge language modelsCCSDCISDtest-time discoveryelectronic structurecoupled cluster

0 comments

The pith

LADeQ uses an out-of-the-box LLM to discover and implement approximation algorithms that accelerate CCSD and CISD calculations with controlled errors.

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

The paper presents LADeQ as a workflow that directs a standard language model to generate candidate approximation schemes for quantum chemistry problems at the moment of computation. These schemes draw techniques from fields such as spatial statistics and circuit simulation, then get coded and tested inside existing electronic-structure programs. The resulting speedups for CCSD and CISD keep correlation-energy errors inside limits chosen by the user, and the generated code remains readable so the source of each error can be traced. This removes the need for large pretraining datasets or fixed libraries of tools, letting the same model adapt to new chemical systems on demand.

Core claim

LADeQ constructs candidate approximation schemes on demand by prompting an LLM, implements them directly in quantum chemistry codes, and benchmarks them at test time. It accelerates coupled cluster singles and doubles (CCSD) and configuration interaction singles and doubles (CISD) calculations while keeping correlation-energy errors within user-specified tolerances, producing transparent implementations whose approximation errors are explicitly traceable.

What carries the argument

The LADeQ workflow, which prompts an LLM to propose, code, and validate approximation schemes drawn from non-quantum-chemistry disciplines inside existing electronic-structure solvers.

If this is right

CCSD and CISD calculations complete faster while correlation-energy errors remain inside user-chosen bounds.
Generated approximations carry explicit, traceable error terms that permit direct accuracy-efficiency trade-offs.
No task-specific pretraining or curated datasets are required for the workflow to function.
Approximation code remains human-inspectable because it is produced as ordinary source rather than opaque parameters.
The same LLM can be reused across different molecules and methods without retraining.

Where Pith is reading between the lines

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

The same test-time prompting pattern could be applied to other costly electronic-structure methods such as higher-order coupled-cluster or multireference approaches.
Because the generated code is readable, domain experts could manually edit the approximations or extract reusable motifs for future manual design.
The approach may reduce dependence on large pretraining corpora in scientific domains where data are sparse or new systems appear frequently.
Test-time discovery of hybrid algorithms that mix ideas from statistics, circuits, and kernel methods could become a routine step in computational chemistry pipelines.

Load-bearing premise

An unmodified general-purpose LLM can produce correct, efficient, and inspectable approximation code from unrelated fields that actually reduces cost inside quantum chemistry solvers without uncontrolled errors.

What would settle it

Running LADeQ on a benchmark molecule such as water and finding that every generated approximation either exceeds the chosen error tolerance or fails to shorten runtime compared with standard CCSD would show the claim does not hold.

Figures

Figures reproduced from arXiv: 2606.20729 by Masaki Adachi, Masaya Hagai, Shuhei Kurita, Tomoya Murata, Yuta Suzuki.

**Figure 2.** Figure 2: Runtime reduction achieved by the best implementation trial for each approximation idea: (a) CCSD and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Section-wise runtime breakdown for the baseline and the top three implementations for (a) CCSD and (b) [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: Best-so-far trajectories of runtime reduction over 10 independent implementation attempts for the top three [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

**Figure 5.** Figure 5: Molecules used in the benchmark: (a) linear hydrocarbon C [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

**Figure 6.** Figure 6: Runtime–accuracy trade-off for DLPNO calculations in ORCA and the top three LADeQ-generated approxi [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗

**Figure 7.** Figure 7: Implementation-level speed–accuracy distributions for the LLM-generated approximation candidates for (a) [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗

read the original abstract

Quantum chemistry simulations underpin modern materials discovery, yet their impact is limited by steep computational cost and dependence on fixed approximation schemes. Foundation models, such as machine-learned interatomic potentials, have accelerated parts of this workflow, but their reliance on large-scale pretraining restricts adaptability at the frontier of chemical space, where methodological innovation and sparse data are the norm. Agentic AI systems can automate existing simulation pipelines, yet they remain constrained by the predefined tools and algorithms they orchestrate. In response, we introduce LADeQ, an LLM-guided workflow that discovers, implements, and benchmarks candidate approximation algorithms at test-time within existing quantum chemistry codes. Rather than selecting from a predefined repertoire, LADeQ constructs candidate approximation schemes on demand, drawing on techniques from disciplines such as spatial statistics, circuit simulation, and kernel methods that have had little prior presence in electronic-structure theory. Because it builds on an out-of-the-box language model, LADeQ requires no task-specific pretraining or curated data, and the resulting implementations are transparent and inspectable, with explicitly traceable approximation errors that enable principled control of accuracy--efficiency trade-offs. We show that LADeQ accelerates coupled cluster singles and doubles (CCSD) and configuration interaction singles and doubles (CISD) calculations while keeping correlation-energy errors within user-specified tolerances, demonstrating autonomous, objective-driven discovery of approximation algorithms inside existing electronic-structure solvers.

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.

Desk Editor's Note private letter to a colleague

LADeQ uses an out-of-the-box LLM at test time to generate approximation code for CCSD/CISD from non-chemistry fields, but the abstract supplies no benchmarks to check whether the generated schemes actually work.

read the letter

The central point is that this paper describes LADeQ, a workflow that prompts an unmodified LLM to invent and implement approximation algorithms on demand inside existing quantum chemistry packages. It pulls techniques from spatial statistics, circuit simulation, and kernel methods, then claims the resulting schemes accelerate CCSD and CISD while holding correlation-energy errors inside user-set bounds.

What is actually new is the test-time construction step itself. Most prior agentic or ML work in this area either fine-tunes models on curated data or selects from a fixed menu of existing solvers. Here the LLM is asked to synthesize fresh code at runtime with no task-specific training, and the approximations are meant to be inspectable with explicit error traces. That framing addresses a real limitation: fixed approximations break down at the edge of chemical space where data is sparse.

The paper does a reasonable job laying out the motivation and the practical advantage of staying inside production solvers rather than replacing them. The emphasis on traceable errors and user-controlled tolerances is also a clear strength compared with black-box ML potentials.

The soft spot is straightforward: the abstract contains no numbers. There are no reported wall-time reductions, no error histograms, no examples of the generated code, and no comparison against standard approximations. Without those, it is impossible to tell whether the LLM produces correct, efficient implementations or whether the error control actually holds in practice. The weakest link remains the assumption that an unmodified model can reliably translate ideas from distant fields into working electronic-structure code.

This is for readers working on AI-assisted method development or adaptive electronic-structure tools. A computational chemist or someone building agentic systems for science would find the conceptual setup useful. It deserves peer review because the claims are concrete enough to be tested once the implementations and benchmarks appear in the full manuscript.

Referee Report

1 major / 0 minor

Summary. The manuscript introduces LADeQ, an LLM-guided workflow that discovers, implements, and benchmarks candidate approximation algorithms at test time inside existing quantum-chemistry codes. It claims to accelerate CCSD and CISD calculations by drawing techniques from spatial statistics, circuit simulation, and kernel methods, while keeping correlation-energy errors within user-specified tolerances, all without task-specific pretraining or curated data and with transparent, inspectable implementations.

Significance. If the performance claims hold with reproducible benchmarks, the approach could enable on-demand, field-agnostic approximation discovery in electronic-structure theory, reducing dependence on fixed schemes and supporting better accuracy-efficiency trade-offs at the frontier of chemical space. The emphasis on inspectable code and traceable errors is a constructive feature for principled control.

major comments (1)

[Abstract] Abstract: the central claim that LADeQ accelerates CCSD and CISD calculations while keeping correlation-energy errors within user-specified tolerances is presented without any quantitative benchmarks, error distributions, timing data, or implementation details. This absence prevents evaluation of whether the discovered approximations actually deliver the stated speedups inside existing solvers without uncontrolled errors.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their thoughtful review and the opportunity to clarify aspects of our manuscript. We respond to the major comment point by point below.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim that LADeQ accelerates CCSD and CISD calculations while keeping correlation-energy errors within user-specified tolerances is presented without any quantitative benchmarks, error distributions, timing data, or implementation details. This absence prevents evaluation of whether the discovered approximations actually deliver the stated speedups inside existing solvers without uncontrolled errors.

Authors: The abstract serves as a high-level summary of the manuscript's contributions and findings. Detailed quantitative benchmarks, including speedups for CCSD and CISD calculations, error distributions relative to user-specified tolerances, timing data, and implementation details of the discovered algorithms, are provided throughout the full manuscript. These are reported in the Results section with specific examples, error statistics, and performance metrics that support the claims. We acknowledge that including a few representative quantitative results in the abstract could enhance immediate evaluability. Accordingly, we will revise the abstract to incorporate key benchmark highlights. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper introduces LADeQ as an LLM-guided workflow that constructs approximation schemes at test time from an unmodified out-of-the-box language model, drawing on external disciplines and existing quantum-chemistry solvers without task-specific pretraining or curated data. No equations, fitted parameters, or self-citations are shown that reduce the claimed accelerations of CCSD/CISD or the error-control mechanism to quantities defined inside the paper itself. The central demonstration remains an empirical application of external LLM capabilities and solvers, rendering the reported results self-contained against external benchmarks rather than circular by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the untested premise that a general LLM can produce chemically valid and computationally advantageous approximation code on demand; no free parameters, axioms, or invented entities are explicitly quantified in the abstract.

axioms (1)

domain assumption An unmodified foundation-model LLM possesses sufficient cross-domain knowledge to generate correct and efficient approximation code for electronic-structure methods.
The workflow description assumes the LLM can translate techniques from spatial statistics, circuit simulation, and kernel methods into working quantum-chemistry approximations without task-specific fine-tuning.

pith-pipeline@v0.9.1-grok · 5804 in / 1331 out tokens · 17421 ms · 2026-06-26T19:16:50.657894+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

35 extracted references · 26 canonical work pages

[1]

Pharmacol.15, 1331062 (2024)

Gangwal, A.et al.Generative artificial intelligence in drug discovery: basic framework, recent advances, chal- lenges, and opportunities.Front. Pharmacol.15, 1331062 (2024). doi:10.3389/fphar.2024.1331062

work page doi:10.3389/fphar.2024.1331062 2024
[2]

doi:10.1038/ s41586-025-08628-5

Zeni, C.et al.A generative model for inorganic materials design.Nature639, 624–632 (2025). doi:10.1038/ s41586-025-08628-5

2025
[3]

Sci.16, 1417–1431 (2025).doi:10.1039/d4sc05894a

Lin, H.et al.DiffBP: generative diffusion of 3D molecules for target protein binding.Chem. Sci.16, 1417–1431 (2025).doi:10.1039/d4sc05894a

work page doi:10.1039/d4sc05894a 2025
[4]

C., Collison, C

Ramos, M. C., Collison, C. J. & White, A. D. A review of large language models and autonomous agents in chemistry.Chem. Sci.16, 2514–2572 (2025).doi:10.1039/d4sc03921a

work page doi:10.1039/d4sc03921a 2025
[5]

arXiv:2402.04379

Gruver, N.et al.Fine-tuned language models generate stable inorganic materials as text.arXiv [cs.LG](2024). arXiv:2402.04379

arXiv 2024
[6]

doi: 10.1016/j.matt.2025.102263

Zou, Y .et al.El agente: An autonomous agent for quantum chemistry.Matter8, 102263 (2025). doi: 10.1016/j.matt.2025.102263

work page doi:10.1016/j.matt.2025.102263 2025
[7]

& Parrinello, M

Behler, J. & Parrinello, M. Generalized neural-network representation of high-dimensional potential-energy surfaces.Phys. Rev. Lett.98, 146401 (2007).doi:10.1103/PhysRevLett.98.146401

work page doi:10.1103/physrevlett.98.146401 2007
[8]

Kocer, E., Ko, T. W. & Behler, J. Neural network potentials: A concise overview of methods.Annu. Rev. Phys. Chem.73, 163–186 (2022).doi:10.1146/annurev-physchem-082720-034254

work page doi:10.1146/annurev-physchem-082720-034254 2022
[9]

T.et al.Biomolecular dynamics with machine-learned quantum-mechanical force fields trained on diverse chemical fragments.Sci

Unke, O. T.et al.Biomolecular dynamics with machine-learned quantum-mechanical force fields trained on diverse chemical fragments.Sci. Adv.10, eadn4397 (2024).doi:10.1126/sciadv.adn4397

work page doi:10.1126/sciadv.adn4397 2024
[10]

Chem.16, 727–734 (2024).doi:10.1038/s41557-023-01427-3

Zhang, S.et al.Exploring the frontiers of condensed-phase chemistry with a general reactive machine learning potential.Nat. Chem.16, 727–734 (2024).doi:10.1038/s41557-023-01427-3

work page doi:10.1038/s41557-023-01427-3 2024
[11]

Energy Chem.106, 911–929 (2025).doi:10.1016/j.jechem.2025.02.051

Zhang, J.et al.Atomistic simulation of batteries via machine learning force fields: From bulk to interface.J. Energy Chem.106, 911–929 (2025).doi:10.1016/j.jechem.2025.02.051

work page doi:10.1016/j.jechem.2025.02.051 2025
[12]

T.et al.Machine learning force fields.Chem

Unke, O. T.et al.Machine learning force fields.Chem. Rev.121, 10142–10186 (2021). doi:10.1021/acs. chemrev.0c01111

work page doi:10.1021/acs 2021
[13]

Kohn, L.J

Kohn, W. & Sham, L. J. Self-consistent equations including exchange and correlation effects.Phys. Rev.140, A1133–A1138 (1965).doi:10.1103/physrev.140.a1133

work page doi:10.1103/physrev.140.a1133 1965
[14]

Impossibility of collective intelligence.arXiv [cs.LG](2022).arXiv:2206.02786

Muandet, K. Impossibility of collective intelligence.arXiv [cs.LG](2022).arXiv:2206.02786

arXiv 2022
[15]

M Bran, A.et al.Augmenting large language models with chemistry tools.Nat. Mach. Intell.6, 525–535 (2024). doi:10.1038/s42256-024-00832-8

work page doi:10.1038/s42256-024-00832-8 2024
[16]

Boiko, Robert MacKnight, Ben Kline, and Gabe Gomes

Boiko, D. A., MacKnight, R., Kline, B. & Gomes, G. Autonomous chemical research with large language models. Nature624, 570–578 (2023).doi:10.1038/s41586-023-06792-0

work page doi:10.1038/s41586-023-06792-0 2023
[17]

& Fung, V

Jia, S., Zhang, C. & Fung, V . LLMatDesign: Autonomous materials discovery with large language models.arXiv [cond-mat.mtrl-sci](2024).arXiv:2406.13163

arXiv 2024
[18]

In Globerson, A.et al.(eds.)Advances in Neural Information Processing Systems 37, vol

Breazeal, C.et al.MDAgents: An adaptive collaboration of LLMs for medical decision-making. In Globerson, A.et al.(eds.)Advances in Neural Information Processing Systems 37, vol. 37, 79410–79452 (Neural Information Processing Systems Foundation, Inc. (NeurIPS), San Diego, California, USA, 2024). doi:10.52202/079017-2522

work page doi:10.52202/079017-2522 2024
[19]

A.et al.DynaMate: leveraging AI-agents for customized research workflows.Mol

Mendible-Barreto, O. A.et al.DynaMate: leveraging AI-agents for customized research workflows.Mol. Syst. Des. Eng.10, 585–598 (2025).doi:10.1039/d5me00062a

work page doi:10.1039/d5me00062a 2025
[20]

& White, A

Campbell, Q., Cox, S., Medina, J., Watterson, B. & White, A. D. MDCrow: Automating molecular dynamics workflows with large language models.arXiv [cs.AI](2025).arXiv:2502.09565. 12 LLM-Guided Test-Time Discovery of Quantum-Chemical Approximation Algorithms

arXiv 2025
[21]

Sun, Q.et al.PySCF: the python-based simulations of chemistry framework: The PySCF program.Wiley Interdiscip. Rev. Comput. Mol. Sci.8, e1340 (2018).doi:10.1002/wcms.1340

work page doi:10.1002/wcms.1340 2018
[22]

& Neese, F

Riplinger, C. & Neese, F. An efficient and near linear scaling pair natural orbital based local coupled cluster method.J. Chem. Phys.138, 034106 (2013).doi:10.1063/1.4773581

work page doi:10.1063/1.4773581 2013
[23]

The orca program system.WIRES Comput

Neese, F. The orca program system.WIRES Comput. Molec. Sci.2, 73–78 (2012). doi:10.1002/wcms.81

work page doi:10.1002/wcms.81 2012
[24]

Chemical Physics Letters , mendeley-groups =

Pulay, P. Convergence acceleration of iterative sequences. the case of scf iteration.Chem. Phys. Lett.73, 393–398 (1980).doi:10.1016/0009-2614(80)80396-4

work page doi:10.1016/0009-2614(80)80396-4 1980
[25]

& Ahlrichs, R

Häser, M. & Ahlrichs, R. Improvements on the direct SCF method: Improved direct SCF method.J. Comput. Chem.10, 104–111 (1989).doi:10.1002/jcc.540100111

work page doi:10.1002/jcc.540100111 1989
[26]

E., Sodt, A

Subotnik, J. E., Sodt, A. & Head-Gordon, M. A near linear-scaling smooth local coupled cluster algorithm for electronic structure.J. Chem. Phys.125, 074116 (2006).doi:10.1063/1.2336426

work page doi:10.1063/1.2336426 2006
[27]

Anderson, D. G. Iterative procedures for nonlinear integral equations.J. ACM12, 547–560 (1965). doi: 10.1145/321296.321305

work page doi:10.1145/321296.321305 1965
[28]

Oosterlee, C. W. & Washio, T. Krylov subspace acceleration of nonlinear multigrid with application to recirculating flows.SIAM Journal on Scientific Computing(2006).doi:10.1137/S1064827598338093

work page doi:10.1137/s1064827598338093 2006
[29]

& Bock, N

Challacombe, M. & Bock, N. Fast multiplication of matrices with decay.arXiv [cs.DS](2010). arXiv: 1011.3534

Pith/arXiv arXiv 2010
[30]

& Seeger, M

Williams, C. & Seeger, M. Using the nyström method to speed up kernel machines.Advances in Neural Information Processing Systems13(2000)

2000
[31]

& Mahoney, M

Drineas, P. & Mahoney, M. W. On the nyström method for approximating a gram matrix for improved kernel-based learning.J. Mach. Learn. Res.6, 2153–2175 (2005)

2005
[32]

Kerns, K. J. & Yang, A. T. Preservation of passivity during RLC network reduction via split congruence transformations. InProceedings of the 34th annual conference on Design automation conference - DAC ’97 (ACM Press, New York, New York, USA, 1997).doi:10.1145/266021.266031

work page doi:10.1145/266021.266031 1997
[33]

& Pileggi, L

Odabasioglu, A., Celik, M. & Pileggi, L. T. PRIMA: passive reduced-order interconnect macromodeling algorithm. IEEE Trans. Comput.-aided Des. Integr. Circuits Syst.17, 645–654 (1998).doi:10.1109/43.712097

work page doi:10.1109/43.712097 1998
[34]

Wood, A. T. A. & Chan, G. Simulation of stationary gaussian processes in [0, 1]d.J. Comput. Graph. Stat.3, 409–432 (1994).doi:10.1080/10618600.1994.10474655

work page doi:10.1080/10618600.1994.10474655 1994
[35]

Dietrich, C. R. & Newsam, G. N. Fast and exact simulation of stationary gaussian processes through circu- lant embedding of the covariance matrix.SIAM Journal on Scientific Computing(2006). doi:10.1137/ S1064827592240555. 13 LLM-Guided Test-Time Discovery of Quantum-Chemical Approximation Algorithms Appendix A Prompts used in LADeQ Prompt for identifying ...

2006

[1] [1]

Pharmacol.15, 1331062 (2024)

Gangwal, A.et al.Generative artificial intelligence in drug discovery: basic framework, recent advances, chal- lenges, and opportunities.Front. Pharmacol.15, 1331062 (2024). doi:10.3389/fphar.2024.1331062

work page doi:10.3389/fphar.2024.1331062 2024

[2] [2]

doi:10.1038/ s41586-025-08628-5

Zeni, C.et al.A generative model for inorganic materials design.Nature639, 624–632 (2025). doi:10.1038/ s41586-025-08628-5

2025

[3] [3]

Sci.16, 1417–1431 (2025).doi:10.1039/d4sc05894a

Lin, H.et al.DiffBP: generative diffusion of 3D molecules for target protein binding.Chem. Sci.16, 1417–1431 (2025).doi:10.1039/d4sc05894a

work page doi:10.1039/d4sc05894a 2025

[4] [4]

C., Collison, C

Ramos, M. C., Collison, C. J. & White, A. D. A review of large language models and autonomous agents in chemistry.Chem. Sci.16, 2514–2572 (2025).doi:10.1039/d4sc03921a

work page doi:10.1039/d4sc03921a 2025

[5] [5]

arXiv:2402.04379

Gruver, N.et al.Fine-tuned language models generate stable inorganic materials as text.arXiv [cs.LG](2024). arXiv:2402.04379

arXiv 2024

[6] [6]

doi: 10.1016/j.matt.2025.102263

Zou, Y .et al.El agente: An autonomous agent for quantum chemistry.Matter8, 102263 (2025). doi: 10.1016/j.matt.2025.102263

work page doi:10.1016/j.matt.2025.102263 2025

[7] [7]

& Parrinello, M

Behler, J. & Parrinello, M. Generalized neural-network representation of high-dimensional potential-energy surfaces.Phys. Rev. Lett.98, 146401 (2007).doi:10.1103/PhysRevLett.98.146401

work page doi:10.1103/physrevlett.98.146401 2007

[8] [8]

Kocer, E., Ko, T. W. & Behler, J. Neural network potentials: A concise overview of methods.Annu. Rev. Phys. Chem.73, 163–186 (2022).doi:10.1146/annurev-physchem-082720-034254

work page doi:10.1146/annurev-physchem-082720-034254 2022

[9] [9]

T.et al.Biomolecular dynamics with machine-learned quantum-mechanical force fields trained on diverse chemical fragments.Sci

Unke, O. T.et al.Biomolecular dynamics with machine-learned quantum-mechanical force fields trained on diverse chemical fragments.Sci. Adv.10, eadn4397 (2024).doi:10.1126/sciadv.adn4397

work page doi:10.1126/sciadv.adn4397 2024

[10] [10]

Chem.16, 727–734 (2024).doi:10.1038/s41557-023-01427-3

Zhang, S.et al.Exploring the frontiers of condensed-phase chemistry with a general reactive machine learning potential.Nat. Chem.16, 727–734 (2024).doi:10.1038/s41557-023-01427-3

work page doi:10.1038/s41557-023-01427-3 2024

[11] [11]

Energy Chem.106, 911–929 (2025).doi:10.1016/j.jechem.2025.02.051

Zhang, J.et al.Atomistic simulation of batteries via machine learning force fields: From bulk to interface.J. Energy Chem.106, 911–929 (2025).doi:10.1016/j.jechem.2025.02.051

work page doi:10.1016/j.jechem.2025.02.051 2025

[12] [12]

T.et al.Machine learning force fields.Chem

Unke, O. T.et al.Machine learning force fields.Chem. Rev.121, 10142–10186 (2021). doi:10.1021/acs. chemrev.0c01111

work page doi:10.1021/acs 2021

[13] [13]

Kohn, L.J

Kohn, W. & Sham, L. J. Self-consistent equations including exchange and correlation effects.Phys. Rev.140, A1133–A1138 (1965).doi:10.1103/physrev.140.a1133

work page doi:10.1103/physrev.140.a1133 1965

[14] [14]

Impossibility of collective intelligence.arXiv [cs.LG](2022).arXiv:2206.02786

Muandet, K. Impossibility of collective intelligence.arXiv [cs.LG](2022).arXiv:2206.02786

arXiv 2022

[15] [15]

M Bran, A.et al.Augmenting large language models with chemistry tools.Nat. Mach. Intell.6, 525–535 (2024). doi:10.1038/s42256-024-00832-8

work page doi:10.1038/s42256-024-00832-8 2024

[16] [16]

Boiko, Robert MacKnight, Ben Kline, and Gabe Gomes

Boiko, D. A., MacKnight, R., Kline, B. & Gomes, G. Autonomous chemical research with large language models. Nature624, 570–578 (2023).doi:10.1038/s41586-023-06792-0

work page doi:10.1038/s41586-023-06792-0 2023

[17] [17]

& Fung, V

Jia, S., Zhang, C. & Fung, V . LLMatDesign: Autonomous materials discovery with large language models.arXiv [cond-mat.mtrl-sci](2024).arXiv:2406.13163

arXiv 2024

[18] [18]

In Globerson, A.et al.(eds.)Advances in Neural Information Processing Systems 37, vol

Breazeal, C.et al.MDAgents: An adaptive collaboration of LLMs for medical decision-making. In Globerson, A.et al.(eds.)Advances in Neural Information Processing Systems 37, vol. 37, 79410–79452 (Neural Information Processing Systems Foundation, Inc. (NeurIPS), San Diego, California, USA, 2024). doi:10.52202/079017-2522

work page doi:10.52202/079017-2522 2024

[19] [19]

A.et al.DynaMate: leveraging AI-agents for customized research workflows.Mol

Mendible-Barreto, O. A.et al.DynaMate: leveraging AI-agents for customized research workflows.Mol. Syst. Des. Eng.10, 585–598 (2025).doi:10.1039/d5me00062a

work page doi:10.1039/d5me00062a 2025

[20] [20]

& White, A

Campbell, Q., Cox, S., Medina, J., Watterson, B. & White, A. D. MDCrow: Automating molecular dynamics workflows with large language models.arXiv [cs.AI](2025).arXiv:2502.09565. 12 LLM-Guided Test-Time Discovery of Quantum-Chemical Approximation Algorithms

arXiv 2025

[21] [21]

Sun, Q.et al.PySCF: the python-based simulations of chemistry framework: The PySCF program.Wiley Interdiscip. Rev. Comput. Mol. Sci.8, e1340 (2018).doi:10.1002/wcms.1340

work page doi:10.1002/wcms.1340 2018

[22] [22]

& Neese, F

Riplinger, C. & Neese, F. An efficient and near linear scaling pair natural orbital based local coupled cluster method.J. Chem. Phys.138, 034106 (2013).doi:10.1063/1.4773581

work page doi:10.1063/1.4773581 2013

[23] [23]

The orca program system.WIRES Comput

Neese, F. The orca program system.WIRES Comput. Molec. Sci.2, 73–78 (2012). doi:10.1002/wcms.81

work page doi:10.1002/wcms.81 2012

[24] [24]

Chemical Physics Letters , mendeley-groups =

Pulay, P. Convergence acceleration of iterative sequences. the case of scf iteration.Chem. Phys. Lett.73, 393–398 (1980).doi:10.1016/0009-2614(80)80396-4

work page doi:10.1016/0009-2614(80)80396-4 1980

[25] [25]

& Ahlrichs, R

Häser, M. & Ahlrichs, R. Improvements on the direct SCF method: Improved direct SCF method.J. Comput. Chem.10, 104–111 (1989).doi:10.1002/jcc.540100111

work page doi:10.1002/jcc.540100111 1989

[26] [26]

E., Sodt, A

Subotnik, J. E., Sodt, A. & Head-Gordon, M. A near linear-scaling smooth local coupled cluster algorithm for electronic structure.J. Chem. Phys.125, 074116 (2006).doi:10.1063/1.2336426

work page doi:10.1063/1.2336426 2006

[27] [27]

Anderson, D. G. Iterative procedures for nonlinear integral equations.J. ACM12, 547–560 (1965). doi: 10.1145/321296.321305

work page doi:10.1145/321296.321305 1965

[28] [28]

Oosterlee, C. W. & Washio, T. Krylov subspace acceleration of nonlinear multigrid with application to recirculating flows.SIAM Journal on Scientific Computing(2006).doi:10.1137/S1064827598338093

work page doi:10.1137/s1064827598338093 2006

[29] [29]

& Bock, N

Challacombe, M. & Bock, N. Fast multiplication of matrices with decay.arXiv [cs.DS](2010). arXiv: 1011.3534

Pith/arXiv arXiv 2010

[30] [30]

& Seeger, M

Williams, C. & Seeger, M. Using the nyström method to speed up kernel machines.Advances in Neural Information Processing Systems13(2000)

2000

[31] [31]

& Mahoney, M

Drineas, P. & Mahoney, M. W. On the nyström method for approximating a gram matrix for improved kernel-based learning.J. Mach. Learn. Res.6, 2153–2175 (2005)

2005

[32] [32]

Kerns, K. J. & Yang, A. T. Preservation of passivity during RLC network reduction via split congruence transformations. InProceedings of the 34th annual conference on Design automation conference - DAC ’97 (ACM Press, New York, New York, USA, 1997).doi:10.1145/266021.266031

work page doi:10.1145/266021.266031 1997

[33] [33]

& Pileggi, L

Odabasioglu, A., Celik, M. & Pileggi, L. T. PRIMA: passive reduced-order interconnect macromodeling algorithm. IEEE Trans. Comput.-aided Des. Integr. Circuits Syst.17, 645–654 (1998).doi:10.1109/43.712097

work page doi:10.1109/43.712097 1998

[34] [34]

Wood, A. T. A. & Chan, G. Simulation of stationary gaussian processes in [0, 1]d.J. Comput. Graph. Stat.3, 409–432 (1994).doi:10.1080/10618600.1994.10474655

work page doi:10.1080/10618600.1994.10474655 1994

[35] [35]

Dietrich, C. R. & Newsam, G. N. Fast and exact simulation of stationary gaussian processes through circu- lant embedding of the covariance matrix.SIAM Journal on Scientific Computing(2006). doi:10.1137/ S1064827592240555. 13 LLM-Guided Test-Time Discovery of Quantum-Chemical Approximation Algorithms Appendix A Prompts used in LADeQ Prompt for identifying ...

2006