arxiv: 2605.10348 · v1 · submitted 2026-05-11 · ⚛️ physics.chem-ph

Recognition: 2 theorem links

· Lean Theorem

Learning to Rank for Selected Configuration Interaction

and Jun Yang, Songwei Liu, Wan Nie, Yingying Yu, Zhiwen Wang

Authors on Pith no claims yet

Pith reviewed 2026-05-12 05:21 UTC · model grok-4.3

classification ⚛️ physics.chem-ph

keywords selected configuration interactionlearning to rankmachine learningelectron correlationTransformerchemical accuracyiron-sulfur

0 comments

The pith

Reframing determinant selection as a pairwise ranking task lets a Transformer model reach chemical accuracy in selected configuration interaction with substantially fewer terms than classification or regression baselines.

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

The paper establishes that selected configuration interaction can be made more efficient by treating the choice of which Slater determinants to keep as a ranking problem instead of trying to classify or regress their individual energies. A Transformer learns orbital dependencies across pairs of determinants and is trained to produce a partial ordering that better matches their true cumulative contribution to the correlation energy. On molecules including N2, CO, H2O, NH3, and C2 this yields convergence with 23 to over 50 percent less runtime and only 55 percent as many determinants as prior machine-learning SCI methods. The same model still reaches chemical accuracy for the challenging iron-sulfur system while using just 12 percent of the full configuration-interaction space and improves on regression-based approaches by either 15 percent higher accuracy at fixed size or 15 percent greater compactness at fixed accuracy.

Core claim

Ranking configuration interaction reframes determinant selection as a pairwise ranking problem and trains a Transformer-based model to optimize the partial ordering of determinants according to their contribution to electron correlation. Benchmarks across plane-wave and Gaussian basis sets show that this alignment of training objective with the intrinsic ranking nature of SCI produces faster convergence and more compact wavefunctions than classification or regression alternatives.

What carries the argument

The pairwise ranking loss inside a Transformer encoder that processes orbital information to score relative importance between pairs of determinants.

Load-bearing premise

That a model trained to rank determinants by pairwise comparisons on small molecules will correctly identify the highest-contributing determinants when applied to new systems or larger active spaces.

What would settle it

A test on an additional molecule or active space where the ranking model requires more determinants than a simple energy-threshold selection to reach the same target accuracy.

Figures

Figures reproduced from arXiv: 2605.10348 by and Jun Yang, Songwei Liu, Wan Nie, Yingying Yu, Zhiwen Wang.

**Figure 2.** Figure 2: Wall time proportion of each computational stage across RCI iterations for N [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

**Figure 3.** Figure 3: Comparison of RCI and NNCI on molecular systems with plane-wave basis. The [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗

**Figure 4.** Figure 4: Comparison of RCI and NNCI on Gaussian-basis benchmark systems: N [PITH_FULL_IMAGE:figures/full_fig_p022_4.png] view at source ↗

**Figure 5.** Figure 5: Potential energy curves of N2 along the bond dissociation coordinate. RCI is compared with NNCI22 calculations using different active orbital spaces, as well as with the same reference curves considered in the NNCI study, including RHF, reduced-space FCI results from Herzog et al. 14 (18 MOs) and Coe 12 (28 MOs), and the experimental (Exp.) fit reported by Le Roy et al. 39 All curves are aligned at their m… view at source ↗

**Figure 6.** Figure 6: Ablation study on the effects of model architecture and training objective. The [PITH_FULL_IMAGE:figures/full_fig_p028_6.png] view at source ↗

**Figure 7.** Figure 7: Analysis of the active pair sampling strategy during pairwise ranking training. [PITH_FULL_IMAGE:figures/full_fig_p029_7.png] view at source ↗

read the original abstract

The accurate description of electron correlation is a central challenge in computational chemistry, with selected configuration interaction (SCI) emerging as a powerful tool to approach the full CI limit. While recent machine learning (ML) integrations have accelerated determinant selection, existing regression and classification approaches suffer from a fundamental objective-loss mismatch: they evaluate the importance of determinants in isolation without explicitly accounting for their relative importance ranking. Here, we introduce ranking configuration interaction (RCI), a novel ML-supported SCI framework that reframes determinant selection as a pairwise ranking problem. Building upon a Transformer-based architecture to capture complex, non-local orbital dependencies, RCI progressively optimizes the partial ordering of determinants. By doing so, RCI aligns the training objective more closely with the intrinsic ranking nature of SCI. Extensive benchmarks across both plane-wave and Gaussian basis sets, including the molecules N$_2$, CO, H$_2$O, NH$_3$, and C$_2$, demonstrate the substantial efficiency of RCI. Compared to previously reported classification baselines, RCI consistently accelerates convergence-reducing overall computational time by 23% to over 50% depending on the system, and requiring only 55% of the determinant count in representative cases such as N$_2$ and CO. Furthermore, RCI exhibits robust performance and reaches chemical accuracy on the highly challenging iron-sulfur using only 12% of the full CI space. Notably, RCI outperforms recent regression-based SCI methods by delivering either further 15% improvement in accuracy at comparable determinant counts, or 15% gain in compactness at similar accuracy. This pairwise learning-to-rank model provides a lightweight and modular plugin that can be seamlessly incorporated into other supervised-learning frameworks.

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

The paper reframes SCI determinant selection as pairwise ranking with a Transformer and reports efficiency gains over regression baselines, but the gains rest on thin methodological detail.

read the letter

The main point is that this work treats determinant selection in selected configuration interaction as a learning-to-rank problem instead of regression or classification. They use a Transformer to capture orbital dependencies and train it to produce better partial orderings, then plug the ranked list into the SCI procedure. Benchmarks on N2, CO, H2O, NH3, C2 and the iron-sulfur cluster show 23-50% faster convergence and roughly half the determinants in representative cases, plus chemical accuracy on the hard system with only 12% of the full CI space. It also claims a 15% edge over recent regression methods in accuracy or compactness at fixed cost. Those numbers come from both plane-wave and Gaussian bases, which is a reasonable spread for the claim that the method is practical and modular.

Referee Report

3 major / 1 minor

Summary. The paper introduces Ranking Configuration Interaction (RCI), a machine-learning framework for selected configuration interaction (SCI) that reframes determinant selection as a pairwise ranking task solved by a Transformer model capturing non-local orbital dependencies. It claims this yields faster convergence than prior classification or regression ML-SCI approaches, with 23-50% reductions in computational time, 55% fewer determinants for N2 and CO, chemical accuracy for iron-sulfur clusters using only 12% of the full CI space, and a 15% edge in accuracy or compactness over recent regression baselines.

Significance. If the central claim holds after verification, RCI would supply a training objective more intrinsically matched to the ranking character of determinant importance in SCI, potentially enabling more compact and rapidly converging wavefunctions for strongly correlated systems. The modular plugin design and benchmarks across plane-wave and Gaussian bases on multiple molecules would make it a practical addition to existing supervised SCI pipelines.

major comments (3)

[Abstract and Methods] Abstract and Methods: The pairwise ranking loss and the procedure for generating training pair labels (i.e., how 'true' relative importance is derived from reference SCI runs) are not specified. Without these definitions it is impossible to determine whether the reported 15% gains over regression baselines arise from the ranking objective itself or from differences in data selection, architecture, or post-hoc tuning.
[Results] Results section: No ablation is presented that isolates the ranking formulation from the Transformer architecture or from the choice of training data splits. Consequently the claim that 'RCI aligns the training objective more closely with the intrinsic ranking nature of SCI' cannot be separated from possible confounding factors such as overfitting to the benchmark molecules or selection bias in the determinant pool.
[Benchmarks] Benchmarks: The reported time reductions (23-50%) and determinant-count savings (55% for N2/CO, 12% for iron-sulfur) are given without error bars, number of independent runs, or statistical tests against the classification and regression baselines. This leaves open the possibility that the efficiency gains are not robust or are driven by implementation details rather than the ranking approach.

minor comments (1)

[Abstract] The abstract states that RCI 'progressively optimizes the partial ordering'; the precise schedule (e.g., how many ranking iterations per SCI step) should be clarified for reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below. Where clarifications or additions are needed, we will revise the manuscript accordingly to improve reproducibility and strengthen the claims.

read point-by-point responses

Referee: [Abstract and Methods] Abstract and Methods: The pairwise ranking loss and the procedure for generating training pair labels (i.e., how 'true' relative importance is derived from reference SCI runs) are not specified. Without these definitions it is impossible to determine whether the reported 15% gains over regression baselines arise from the ranking objective itself or from differences in data selection, architecture, or post-hoc tuning.

Authors: We agree that explicit definitions of the pairwise ranking loss and label generation procedure are essential for reproducibility and for isolating the source of the reported gains. In the revised manuscript we will add a dedicated subsection in Methods that (i) specifies the exact pairwise loss (e.g., logistic or hinge loss on ordered pairs), (ii) describes how reference SCI runs are used to assign ground-truth labels (by comparing each determinant’s contribution to the total energy or to the squared norm of the wavefunction), and (iii) states the sampling strategy used to form training pairs. These additions will allow readers to verify that performance differences arise from the ranking objective rather than ancillary implementation choices. revision: yes
Referee: [Results] Results section: No ablation is presented that isolates the ranking formulation from the Transformer architecture or from the choice of training data splits. Consequently the claim that 'RCI aligns the training objective more closely with the intrinsic ranking nature of SCI' cannot be separated from possible confounding factors such as overfitting to the benchmark molecules or selection bias in the determinant pool.

Authors: We acknowledge the value of controlled ablations. The revised manuscript will include a new ablation table that fixes the Transformer architecture and training-data splits while varying only the training objective (ranking loss versus classification and regression losses). We will also report performance on held-out molecules and on alternative determinant-pool splits to address potential overfitting or selection bias. These experiments will provide direct evidence that the ranking formulation, rather than architecture or data choices, drives the observed improvements. revision: yes
Referee: [Benchmarks] Benchmarks: The reported time reductions (23-50%) and determinant-count savings (55% for N2/CO, 12% for iron-sulfur) are given without error bars, number of independent runs, or statistical tests against the classification and regression baselines. This leaves open the possibility that the efficiency gains are not robust or are driven by implementation details rather than the ranking approach.

Authors: We agree that statistical characterization is necessary to substantiate robustness. In the revised manuscript we will (i) report means and standard deviations over a stated number of independent runs (with different random seeds for both training and determinant selection), (ii) include error bars on all timing and determinant-count figures, and (iii) apply paired statistical tests (e.g., t-tests) against the classification and regression baselines. These additions will demonstrate that the efficiency gains are statistically significant and not attributable to implementation variability. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical ML ranking framework validated on external benchmarks

full rationale

The paper introduces RCI as an empirical supervised learning method that trains a Transformer on pairwise ranking labels derived from SCI determinant contributions. Reported gains (23-50% time reduction, 55% fewer determinants for N2/CO, 12% CI space for iron-sulfur, 15% edge over regression baselines) are obtained from direct numerical benchmarks on held-out molecular systems rather than from any fitted parameter, self-referential definition, or self-citation chain that reduces the claimed result to its own inputs by construction. No equations, uniqueness theorems, or ansatzes are presented that loop back; the derivation chain consists of standard ML training plus SCI energy evaluation, which remains externally falsifiable.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Review performed on abstract only; no explicit free parameters, axioms, or invented entities are detailed in the provided text.

pith-pipeline@v0.9.0 · 5606 in / 1174 out tokens · 39330 ms · 2026-05-12T05:21:18.383680+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
reframes determinant selection as a pairwise ranking problem... Pairwise Logistic Loss... L = 1/|P| Σ log(1 + exp(-(s_i - s_j)))
IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear
Transformer-based architecture to capture complex, non-local orbital dependencies

Reference graph

Works this paper leans on

49 extracted references · 49 canonical work pages

[1]

Learning to rank for information retrieval , ISBN =

Liu, Tie-Yan , year =. Learning to rank for information retrieval , ISBN =. doi:10.1007/978-3-642-14267-3 , publisher =

work page doi:10.1007/978-3-642-14267-3
[2]

Journal of Chemical Theory and Computation , volume=

Neural-network-based selective configuration interaction approach to molecular electronic structure , author=. Journal of Chemical Theory and Computation , volume=. 2025 , publisher=

work page 2025
[3]

SciPost Physics Codebases , pages=

SOLAX: A Python solver for fermionic quantum systems with neural network support , author=. SciPost Physics Codebases , pages=

work page
[4]

Advances in Neural Information Processing Systems , volume=

Pytorch: An imperative style, high-performance deep learning library , author=. Advances in Neural Information Processing Systems , volume=

work page
[5]

Proceedings of the 22nd International Conference on Machine Learning , pages=

Learning to rank using gradient descent , author=. Proceedings of the 22nd International Conference on Machine Learning , pages=

work page
[6]

Advances in Neural Information Processing Systems , volume=

Mcrank: Learning to rank using multiple classification and gradient boosting , author=. Advances in Neural Information Processing Systems , volume=

work page
[7]

Proceedings of the 24th International Conference on Machine Learning , pages=

Learning to rank: from pairwise approach to listwise approach , author=. Proceedings of the 24th International Conference on Machine Learning , pages=

work page
[8]

Advances in Neural Information Processing Systems , volume=

Attention is all you need , author=. Advances in Neural Information Processing Systems , volume=

work page
[9]

Advances in Neural Information Processing Systems , volume=

Learning to rank with nonsmooth cost functions , author=. Advances in Neural Information Processing Systems , volume=

work page
[10]

Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages=

Structured learning for non-smooth ranking losses , author=. Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages=

work page
[11]

Chemical Physics Letters , volume=

A full-configuration benchmark for the N2 molecule , author=. Chemical Physics Letters , volume=. 1999 , publisher=

work page 1999
[12]

Journal of Chemical Theory and Computation , volume=

Distributed implementation of full configuration interaction for one trillion determinants , author=. Journal of Chemical Theory and Computation , volume=. 2024 , publisher=

work page 2024
[13]

Journal of Chemical Theory and Computation , volume=

Solving the Schr\"odinger equation in the configuration space with generative machine learning , author=. Journal of Chemical Theory and Computation , volume=. 2023 , publisher=

work page 2023
[14]

Journal of Chemical Theory and Computation , volume=

Accelerating many-body quantum chemistry via generative Transformer-enhanced configuration interaction , author=. Journal of Chemical Theory and Computation , volume=. 2025 , publisher=

work page 2025
[15]

The Journal of Chemical Physics , volume=

An accurate analytic potential function for ground-state N2 from a direct-potential-fit analysis of spectroscopic data , author=. The Journal of Chemical Physics , volume=. 2006 , publisher=

work page 2006
[16]

The Journal of Chemical Physics , volume=

Iterative perturbation calculations of ground and excited state energies from multiconfigurational zeroth-order wavefunctions , author=. The Journal of Chemical Physics , volume=. 1973 , publisher=

work page 1973
[17]

The Journal of Chemical Physics , volume=

A deterministic alternative to the full configuration interaction quantum Monte Carlo method , author=. The Journal of Chemical Physics , volume=. 2016 , publisher=

work page 2016
[18]

Journal of Chemical Theory and Computation , volume=

Modern approaches to exact diagonalization and selected configuration interaction with the adaptive sampling CI method , author=. Journal of Chemical Theory and Computation , volume=. 2020 , publisher=

work page 2020
[19]

Journal of Chemical Theory and Computation , volume=

Efficient heat-bath sampling in Fock space , author=. Journal of Chemical Theory and Computation , volume=. 2016 , publisher=

work page 2016
[20]

Journal of Chemical Theory and Computation , volume=

Heat-bath configuration interaction: An efficient selected configuration interaction algorithm inspired by heat-bath sampling , author=. Journal of Chemical Theory and Computation , volume=. 2016 , publisher=

work page 2016
[21]

Journal of Chemical Theory and Computation , volume=

Semistochastic heat-bath configuration interaction method: Selected configuration interaction with semistochastic perturbation theory , author=. Journal of Chemical Theory and Computation , volume=. 2017 , publisher=

work page 2017
[22]

International Journal of Quantum Chemistry , volume=

An application of perturbation theory ideas in configuration interaction calculations , author=. International Journal of Quantum Chemistry , volume=. 1968 , publisher=

work page 1968
[23]

Chemical Physics Letters , volume=

Direct selected configuration interaction using a hole-particle formalism , author=. Chemical Physics Letters , volume=. 1992 , publisher=

work page 1992
[24]

The Journal of Chemical Physics , volume=

Adaptive multiconfigurational wave functions , author=. The Journal of Chemical Physics , volume=. 2014 , publisher=

work page 2014
[25]

The Journal of Physical Chemistry Letters , volume=

Downfolded configuration interaction for chemically accurate electron correlation , author=. The Journal of Physical Chemistry Letters , volume=. 2022 , publisher=

work page 2022
[26]

Journal of Chemical Theory and Computation , volume=

Machine learning configuration interaction , author=. Journal of Chemical Theory and Computation , volume=. 2018 , publisher=

work page 2018
[27]

Journal of Chemical Theory and Computation , volume=

Machine learning configuration interaction for ab initio potential energy curves , author=. Journal of Chemical Theory and Computation , volume=. 2019 , publisher=

work page 2019
[28]

Physical Review C , volume=

Machine learning approach to pattern recognition in nuclear dynamics from the ab initio symmetry-adapted no-core shell model , author=. Physical Review C , volume=. 2022 , publisher=

work page 2022
[29]

Journal of Chemical Theory and Computation , volume=

Hamiltonian-guided autoregressive selected-configuration interaction achieves chemical accuracy in strongly correlated systems , author=. Journal of Chemical Theory and Computation , volume=. 2025 , publisher=

work page 2025
[30]

Journal of Chemical Theory and Computation , volume=

Chembot: A machine learning approach to selective configuration interaction , author=. Journal of Chemical Theory and Computation , volume=. 2021 , publisher=

work page 2021
[31]

Journal of Chemical Theory and Computation , volume=

Active learning configuration interaction for excited-state calculations of polycyclic aromatic hydrocarbons , author=. Journal of Chemical Theory and Computation , volume=. 2021 , publisher=

work page 2021
[32]

Physical Review Letters , volume=

Deep-learning approach for the atomic configuration interaction problem on large basis sets , author=. Physical Review Letters , volume=. 2023 , publisher=

work page 2023
[33]

Physical Review A , volume=

Neural-network approach to running high-precision atomic computations , author=. Physical Review A , volume=. 2024 , publisher=

work page 2024
[34]

Physical Review B , volume=

Neural-network-supported basis optimizer for the configuration interaction problem in quantum many-body clusters: Feasibility study and numerical proof , author=. Physical Review B , volume=. 2025 , publisher=

work page 2025
[35]

arXiv preprint arXiv:2510.27665 , year=

Natural-orbital-based neural network configuration interaction , author=. arXiv preprint arXiv:2510.27665 , year=

work page arXiv
[36]

Machine Learning Assisted Selective Configuration Interaction for Accurate Ground and Excited State Calculations , volume =

Casier, Bastien and El Hamdi, Maissa and Herzog, Basile , year =. Machine Learning Assisted Selective Configuration Interaction for Accurate Ground and Excited State Calculations , volume =. Journal of Chemical Theory and Computation , publisher =. doi:10.1021/acs.jctc.5c01652 , number =

work page doi:10.1021/acs.jctc.5c01652
[37]

Journal of Chemical Theory and Computation , volume=

Reinforcement learning configuration interaction , author=. Journal of Chemical Theory and Computation , volume=. 2021 , publisher=

work page 2021
[38]

Synthetic analogs of the active sites of iron-sulfur proteins. XI. Synthesis and properties of complexes containing the iron sulfide (Fe2S2) core and the structures of bis [o-xylyl-. alpha.,. alpha.'-dithiolato-. mu.-sulfido-ferrate (III)] and bis [p-tolylthiolato-. mu.-sulfido-ferrate (III)] dianions , author=. Journal of the American Chemical Society , ...

work page 1975
[39]

Nature Chemistry , volume=

Low-energy spectrum of iron--sulfur clusters directly from many-particle quantum mechanics , author=. Nature Chemistry , volume=. 2014 , publisher=

work page 2014
[40]

The Journal of Chemical Physics , volume=

Communication: A flexible multi-reference perturbation theory by minimizing the Hylleraas functional with matrix product states , author=. The Journal of Chemical Physics , volume=. 2014 , publisher=

work page 2014
[41]

Journal of Chemical Theory and Computation , volume=

Spin-projected matrix product states: Versatile tool for strongly correlated systems , author=. Journal of Chemical Theory and Computation , volume=. 2017 , publisher=

work page 2017
[42]

Wiley Interdisciplinary Reviews: Computational Molecular Science , volume=

PySCF: the Python-based simulations of chemistry framework , author=. Wiley Interdisciplinary Reviews: Computational Molecular Science , volume=. 2018 , publisher=

work page 2018
[43]

The Journal of Chemical Physics , volume=

Recent developments in the PySCF program package , author=. The Journal of Chemical Physics , volume=. 2020 , publisher=

work page 2020
[44]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

Pads: Policy-adapted sampling for visual similarity learning , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

work page
[45]

The Journal of Chemical Physics , volume=

Fast semistochastic heat-bath configuration interaction , author=. The Journal of Chemical Physics , volume=. 2018 , publisher=

work page 2018
[46]

Advances in Neural Information Processing Systems , volume=

PSL: Rethinking and improving softmax loss from pairwise perspective for recommendation , author=. Advances in Neural Information Processing Systems , volume=

work page
[47]

ACM SIGIR Forum , volume=

IR evaluation methods for retrieving highly relevant documents , author=. ACM SIGIR Forum , volume=. 2017 , organization=

work page 2017
[48]

Physical Review B , volume=

Projector augmented-wave method , author=. Physical Review B , volume=. 1994 , publisher=

work page 1994
[49]

Bulletin of Materials Science , volume=

Projector augmented wave method: ab initio molecular dynamics with full wave functions , author=. Bulletin of Materials Science , volume=. 2003 , publisher=

work page 2003