Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

SpecXMaster claims an agentic RL system can interpret raw NMR FID data into chemical structures, achieving 70.2% top-1 accuracy on combined 1H/13C spectra.

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 ↗

An agentic reinforcement-learning system using raw FID NMR data, generate/search/repair/rerank tools, and a GRPO-trained policy reports improved hit@1 structure elucidation versus its own generation baseline and untrained LLM workflows on NMRexp.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection A real internal gain from RL, but the unvalidated in-house predictor and all-in-house evaluation keep the broad claims from being convincing yet. the 3 major comments →

arxiv 2603.23101 v3 pith:NOWG57HY submitted 2026-03-24 cs.LG

SpecXMaster Technical Report

Yutang Ge , Yaning Cui , Hanzheng Li , Jun-Jie Wang , Fanjie Xu , Jinhan Dong , Yongqi Jin , Dongxu Cui
show 10 more authors
Peng Jin Guojiang Zhao Hengxing Cai Tianci Yangfeng Xueqing Chen Hongshuai Wang Rong Zhu Linfeng Zhang Xiaohong Ji Zhifeng Gao
This is my paper
classification cs.LG
keywords NMR spectroscopyreinforcement learningLLM agentstructure elucidationFID processingmolecular generationspectral matchingautomated chemistry
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.

The reading

This paper attempts to show that NMR structure elucidation, traditionally an expert-dependent human task, can be fully automated from raw time-domain FID signals. The proposed system, SpecXMaster, first converts FID data into structured multiplicity text, then uses a reinforcement-trained LLM agent to generate, search, repair, and rerank candidate molecular structures. On a large public benchmark (NMRexp), the agent reports hit@1 of 0.702 for joint 1H and 13C input, exceeding the generation-only baseline (0.639) and two general-purpose LLM workflow baselines. The authors argue that the improvement comes from learned sequential decision-making over a tool environment, not from better tools alone. If true, this would remove a major human bottleneck in AI-driven chemistry laboratories and enable closed-loop experiment-interpretation workflows.

Core claim

SpecXMaster's central claim is that a language-model policy trained with group-relative policy optimization learns when to generate, search, repair, and rerank molecular candidates, and that this learned control is what lifts NMR interpretation performance. The paper reports hit@1 = 0.702 on joint 1H/13C input, 0.455 on 13C alone, and 0.450 on 1H alone, outperforming the generation tool and two workflow baselines. RL training also drives format validity to near-perfection (0.999) and sharply improves the agent's stop-versus-continue decisions ('case judgement' rises from 0.316 to 0.890 on joint data). The ablation study identifies the alignment reward—which scores whether candidate summaries

What carries the argument

The load-bearing component is the agent policy: a 7B-parameter LLM (Qwen2.5-7B) acting as a controller over four tools—Generate (autoregressive SMILES from spectra), Search (HNSW retrieval over a 106-million-molecule simulated-spectrum database), Repair (REINVENT4 optimizing a spectral-similarity reward), and Rerank (structured peak assignment with a hyperbolic hard-case scorer). Training is two-stage: supervised fine-tuning on 2,000 validation examples, then GRPO. The alignment reward, which encodes improvement in candidate-pool summary and reduction of redundant hypotheses, is the reward term whose ablation hurts hit@1 most.

Load-bearing premise

The whole reward loop for Repair and Rerank trusts that the in-house forward spectral predictor produces chemical shifts accurate enough that spectral similarity to it is a reliable signal of structural correctness—and that predictor's accuracy on independent experimental spectra is not demonstrated in the paper.

What would settle it

Run the trained SpecXMaster agent on a set of molecules deliberately chosen to be under-represented in the training distribution (e.g., complex natural products or exotic heterocycles) and compare its hit@1 against the generation-only baseline. If the gap shrinks or reverses, the RL gains likely come from database memorization rather than learned reasoning. A second check: replace the in-house shift predictor with an independently validated model and observe whether Repair and Rerank quality collapse.

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

If this is right

  • If the reported gains hold, automated NMR interpretation shifts from one-shot prediction to a closed-loop agent process that can refine its own output through iterative reasoning.
  • The pipeline's ability to read raw FID data means chemists could feed instrument output directly to the agent without manual pre-processing into peak tables or shift lists.
  • The near-perfect format validity after RL suggests a practical path to reliable multi-step tool-use for chemistry-specific LLM agents.
  • The alignment reward's importance implies that progress signals about candidate-pool quality are more valuable than per-step action rewards alone.
  • The hyperbolic hard-case reranker implies that near-tie candidates can be disambiguated by embedding geometry, a mechanism potentially transferable to other difficulty-ranked retrieval problems.

Where Pith is reading between the lines

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

  • Editor's inference: The in-house forward spectral predictor that powers Repair's reward (Eq. 9) and Rerank's assignment cost (Eqs. 13–14) is a quiet load-bearer. If this predictor is biased toward molecules similar to its training database, the entire RL loop may reward chemically infeasible fits while still ranking database-like molecules correctly.
Share X Bluesky LinkedIn Reddit HN

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 / 6 minor

Summary. This technical report describes SpecXMaster, an agentic reinforcement learning system for molecular structure elucidation from raw 1H/13C NMR FID data. The pipeline combines (i) conventional FID processing (apodization, FFT, phase/baseline correction), (ii) a neural multiplet classifier, (iii) a multi-tool environment with Generate, Search, Repair, Rerank, and Assign operations, and (iv) an LLM policy trained with SFT and GRPO. The headline empirical claim is in Table 2: SpecXMaster reaches hit@1 = 0.702 (Joint), 0.455 (13C), 0.450 (1H), outperforming the internal Generation Tool (0.639/0.424/0.438) and two workflow baselines (GPT-5.2, Qwen2.5-7B). Table 3 attributes most of the gain to RL over SFT, and Table 4 shows each reward component contributes. The paper also introduces a hyperbolic 'hard-case processor' inside the Rerank tool.

Significance. If the results hold, the paper would make a useful contribution to automated NMR interpretation: it is one of the few systems that goes from raw FID to a ranked structure list, and the RL policy is a principled way to combine generation, retrieval, repair, and reranking. The main evaluation is not circular: hit@1 is scored against ground-truth molecules from NMRexp, not against the in-house spectral predictor's own outputs. The ablations in Tables 3–4 are carefully designed (same backbone/tools, one component removed at a time) and are internally consistent. However, the significance is limited by the absence of external published baselines, the lack of validation of the in-house forward spectral predictor that drives the reward/reranking signal, and the unsupported claim for the hyperbolic hard-case processor. These are fixable, but they are load-bearing for the claim of 'superior performance.'

major comments (3)
  1. [§5.1/§5.2, Table 2] The claim of superior performance is made only against internal baselines. No published method (e.g., NMRPeak, NMR-Solver, DeepSPInN, or other public benchmarks) is compared. The abstract says 'multiple public NMR interpretation benchmarks,' but the experiments report only NMRexp-derived splits. This is load-bearing because the central conclusion is a comparative superiority claim. Please add published/independent baselines or restrict the claim to 'improvement over the internal Generation Tool and workflow variants.'
  2. [§3.3, Eqs. (9) and (13)–(14)] The Repair reward and Rerank assignment cost depend on an in-house forward spectral predictor: Rrepair(m) = Sim(ŝ(m), s*) and L(A|c,X) uses structure-derived chemical-shift and Karplus-type coupling estimates. No independent validation of this predictor against held-out experimental spectra is reported. If the predictor is biased toward molecules similar to the Search database or its training distribution, the RL policy may optimize a self-consistent but not chemically faithful objective, and the reported hit@1 gains could reflect this bias rather than genuine spectral interpretation. This is not circular for the final hit@1 metric, but it is load-bearing for the interpretation of the RL gains. Please add (a) validation of the predictor on the NMRexp test set (shift-error distributions), (b) an ablation that replaces the in-house predictor with an oracle or a published shift-prediction m
  3. [§3.3 (Hard-case processor), §5.2.2 and Table 4] The hyperbolic hard-case processor is described in detail and credited in the conclusion with 'handling hard-case candidates,' but no experiment isolates it. Table 4's ablation removes reward components, not the hard-case processor, and no metric reports performance on near-tie candidates. Please add an ablation with and without the hard-case processor and report hard-case-subset hit@1, or remove the claim from the conclusion and abstract.
minor comments (6)
  1. [Table 2] The table cells are run together (e.g., '0.5730.5920.820'), making it difficult to read. Please format the numbers with clear column separators.
  2. [Tables 2 and 3] SpecXMaster's 1H hit@1 is reported as 0.450 in Table 2 and 0.449 in Table 3. Please reconcile the inconsistency.
  3. [Abstract and §5] The abstract and conclusion claim 'multiple public NMR interpretation benchmarks,' but only the NMRexp benchmark is evaluated. This wording should be adjusted or additional benchmarks added.
  4. [§3.3, Eq. (11)] The Karplus-type coupling relation J(θ) = Acos²θ + Bcosθ + C is stated without specifying the coefficients A, B, C or a reference table. For reproducibility, provide the values or the source used for the estimation.
  5. [§6 and Appendix B] The case study is described as an 'internal case study' based on feedback from one group; it is not an external or quantitative evaluation. The text 'refined through iterative evaluations by professional chemical spectroscopists' is likewise not documented. Please clarify the role of these qualitative examples relative to the benchmark results.
  6. [§5.1] No error bars or statistical significance tests are reported for the hit@1 differences. While the test sets are large, the reader cannot assess stability across random seeds or data subsets; reporting confidence intervals or repeated runs would strengthen the main claim.

Circularity Check

0 steps flagged

No significant circularity: hit@1 is measured against held-out ground-truth molecules, and no equation reduces the central claim to the in-house predictor's own outputs.

full rationale

The central empirical claim is hit@1 on the NMRexp test split, which is an external ground-truth label (the correct molecular structure). The Repair reward (Eq. 9) and Rerank cost (Eqs. 13-16) use an in-house forward spectral predictor, but the evaluated metric is not Sim(ŝ(m), s*) or S(c); it is whether the ground-truth molecule appears in the top-k. The paper states that agent-training data are drawn from the validation split and the test split is reserved exclusively for evaluation (Section 5.1), so the RL objective is not fitted to the test labels. The same-group citations (NMRexp, NMR-Solver, NMRPeak, Uni-Mol, the shift predictor) supply components and benchmarks, but no load-bearing step is justified solely by a self-citation; the evaluation is an experiment, not a derivation from the cited works. The in-house predictor's lack of independent validation is a correctness/generalization concern, not a circularity: if the predictor were biased, hit@1 against experimental spectra would suffer rather than being forced. I find no equation or passage where the predicted output is defined as the fitted input, nor any uniqueness theorem or ansatz smuggled in via self-citation. Hence no circular step can be exhibited.

Axiom & Free-Parameter Ledger

11 free parameters · 8 axioms · 0 invented entities

The system leans on a large inherited stack: NMRexp, NMRPeak/NMR-Solver, Uni-Mol, and the in-house forward predictor. The only quantities the paper itself fits are the RL/rerank/hard-case hyperparameters, none of which are reported with values or sensitivity. The most safety-critical assumption is the unvalidated forward predictor, because it is used in both the reward and the final rerank.

free parameters (11)
  • RL reward weights λ_fmt, λ_eff, λ_tool, λ_align
    Eq. (30); hand-tuned, values not reported, control trade-off between format, efficiency, tool success, and alignment.
  • Tool-usage coefficients α_succ, α_prog, α_fail
    Eq. (33); balance positive/negative tool outcomes; no values given.
  • Alignment reward weights β1, β2
    Eq. (34); no values given.
  • Rerank cost weights λ_match, λ_coverage, λ_observation
    Eq. (14); control assignment cost; no values given.
  • Coverage penalties α, β
    Eq. (15); penalize unexplained signals/sites; no values given.
  • Hyperbolic curvature κ and hard-case loss temperature η, margin m, weights λ_rank, λ_reg
    Eqs. (17), (23)–(26); all chosen without reported values or sensitivity analysis.
  • GRPO clipping ϵ_c and group size G
    Eqs. (38)–(40); standard hyperparameters; G and ϵ_c not specified (batch size 32 and 900 steps are given in §5.1).
  • Efficiency budget T0 and reward penalty
    Eq. (32); T0 not reported.
  • Multiplet classifier architecture and training hyperparameters
    §2.5; the NN for multiplet classification is described only qualitatively; training set and hyperparameters omitted.
  • Apodization window parameters
    §2.4; window type and parameters are not specified; they affect peak resolution and integrals.
  • Peak-detection noise threshold and boundary parameters
    §2.4; SNRs and derivative thresholds unstated.
axioms (8)
  • standard math FFT, apodization, phase/baseline correction operate as standard textbook/NMRglue implementations.
    §2.3–2.4; standard DSP, low risk.
  • domain assumption ACME entropy-minimization phase correction converges to correct phase on heterogeneous spectra.
    §2.4; no evaluation of phase-correction failures.
  • domain assumption Trained neural multiplet classifier generalizes to unseen/overlapping peaks.
    §2.5; training data not described.
  • domain assumption In-house forward spectral predictor gives accurate enough spectra to serve as RL reward and rerank signal.
    §3.3 Repair (Eq. 9) and Rerank (Eqs. 13–14); no independent validation.
  • domain assumption The structure-based chemical-shift model (cited [7]) is accurate on novel molecules outside its training distribution.
    §3.3 used for ϕδ and shift predictions; [7] is same-group prior work.
  • domain assumption Karplus-type relations J(θ)=A cos²θ+B cosθ+C are valid for coupling pathways in the benchmark molecules.
    Eq. (11); literature approximation, no error model.
  • domain assumption NMRexp database structures and labels are correct and spectral simulations for 106M PubChem compounds are reliable for retrieval.
    §5.1 and Search tool; both built by same group in prior papers [12,14].
  • domain assumption Hyperbolic representation learning improves hard-case candidate discrimination.
    §3.3; motivated by prior work but no ablation in this paper.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SpecXMaster Technical Report." pith.science (2026). https://pith.science/paper/NOWG57HY

@misc{pith2026260323101,
  author       = {Pith},
  title        = {Pith review of: SpecXMaster Technical Report},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NOWG57HY}},
  note         = {Machine review of arXiv:2603.23101}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Intelligent spectroscopy serves as a pivotal element in AI-driven closed-loop scientific discovery, functioning as the critical bridge between matter structure and artificial intelligence. However, conventional expert-dependent spectral interpretation encounters substantial hurdles, including susceptibility to human bias and error, dependence on limited specialized expertise, and variability across interpreters. To address these challenges, we propose SpecXMaster, an intelligent framework leveraging Agentic Reinforcement Learning (RL) for NMR molecular spectral interpretation. SpecXMaster enables automated extraction of multiplicity information from both 1H and 13C spectra directly from raw FID (free induction decay) data. This end-to-end pipeline enables fully automated interpretation of NMR spectra into chemical structures. It demonstrates superior performance across multiple public NMR interpretation benchmarks and has been refined through iterative evaluations by professional chemical spectroscopists. We believe that SpecXMaster, as a novel methodological paradigm for spectral interpretation, will have a profound impact on the organic chemistry community.

Figures

Figures reproduced from arXiv: 2603.23101 by Dongxu Cui, Fanjie Xu, Guojiang Zhao, Hanzheng Li, Hengxing Cai, Hongshuai Wang, Jinhan Dong, Jun-Jie Wang, Linfeng Zhang, Peng Jin, Rong Zhu, Tianci Yangfeng, Xiaohong Ji, Xueqing Chen, Yaning Cui, Yongqi Jin, Yutang Ge, Zhifeng Gao.

Figure 1
Figure 1. Figure 1: End-to-end pipeline of SpecXMaster 2 Full Processing Steps from FID to NMR Spectrum 2.1 Overview NMR instruments record experimental signals in the form of Free Induction Decay (FID) time￾domain data. Converting FID signals into interpretable spectra and structured chemical information is a critical step in automated NMR analysis. We developed a Python-based automated FID processing and spectral interpreta… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of Nuclear Magnetic Resonance (NMR) Data Processing [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of SpecXMaster, the proposed agentic framework for NMR-based molecular [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: A comprehensive case study of elucidating molecular structure from FID data: (a) trans [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. NMR Elucidation as an Agentic Search Problem, Not a Modeling Problem

    cs.LG 2026-07 conditional novelty 6.0

    An LLM-driven agent with NMR processing and validation tools identifies molecular structures from raw 1D NMR spectra at accuracy comparable to graduate students on the Alberts benchmark and above zero-shot deep learni...

Reference graph

Works this paper leans on

53 extracted references · 6 linked inside Pith · cited by 1 Pith paper

  1. [1]

    gold standard

    Abdul-Hamid Emwas, Kacper Szczepski, Benjamin Gabriel Poulson, Kousik Chandra, Ryan T McKay, Manel Dhahri, Fatimah Alahmari, Lukasz Jaremko, Joanna Izabela Lachowicz, and Mariusz Jaremko. Nmr as a “gold standard” method in drug design and discovery, 2020

    2020

  2. [2]

    Schmid, Sterling G

    Gary Tom, Stefan P. Schmid, Sterling G. Baird, Yang Cao, Kourosh Darvish, Han Hao, Stanley Lo, Sergio Pablo-García, Ella M. Rajaonson, Marta Skreta, Naruki Yoshikawa, Samantha Corapi, Gun Deniz Akkoc, Felix Strieth-Kalthoff, Martin Seifrid, and Alán Aspuru-Guzik. Self-driving laboratories for chemistry and materials science. Chemical Reviews , 124(16):963...

    2024

  3. [3]

    Granda, Liva Donina, Vincenza Dragone, De-Liang Long, and Leroy Cronin

    Jarosław M. Granda, Liva Donina, Vincenza Dragone, De-Liang Long, and Leroy Cronin. Controlling an organic synthesis robot with machine learning to search for new reactivity. Nature , 559(7714):377–381, Jul 2018

    2018

  4. [4]

    Merz, Kenneth M

    Susanta Das and Jr. Merz, Kenneth M. Exploring the frontiers of computational nmr: Methods, applications, and challenges. Chemical Reviews , 125(19):9256–9295, Sep 2025

    2025

  5. [5]

    Artificial intelligence in spectroscopy: Advancing chemistry from prediction to generation and beyond

    Kehan Guo, Yili Shen, Gisela Abigail Gonzalez-Montiel, Yue Huang, Yujun Zhou, Mihir Surve, Zhichun Guo, Prayel Das, Nitesh V Chawla, Olaf Wiest, and Xiangliang Zhang. Artificial intelligence in spectroscopy: Advancing chemistry from prediction to generation and beyond. 2025

    2025

  6. [6]

    Toward a unified benchmark and framework for deep learning-based prediction of nuclear magnetic resonance chemical shifts

    Fanjie Xu, Wentao Guo, Feng Wang, Lin Yao, Hongshuai Wang, Fujie Tang, Zhifeng Gao, Linfeng Zhang, Weinan E, Zhong-Qun Tian, et al. Toward a unified benchmark and framework for deep learning-based prediction of nuclear magnetic resonance chemical shifts. Nature Computational Science , 5(4):292--300, 2025

    2025

  7. [7]

    From human labels to literature: Semi-supervised learning of nmr chemical shifts at scale

    Yongqi Jin, Yecheng Wang, Jun-jie Wang, Rong Zhu, Guolin Ke, et al. From human labels to literature: Semi-supervised learning of nmr chemical shifts at scale. arXiv preprint arXiv:2601.18524 , 2026

    arXiv 2026

  8. [8]

    Accurate and efficient structure elucidation from routine one-dimensional nmr spectra using multitask machine learning

    Frank Hu, Michael S Chen, Grant M Rotskoff, Matthew W Kanan, and Thomas E Markland. Accurate and efficient structure elucidation from routine one-dimensional nmr spectra using multitask machine learning. ACS Central Science , 10(11):2162--2170, 2024

    2024

  9. [9]

    Language models can understand spectra: A multimodal model for molecular structure elucidation

    Yunyue Su, Jiahui Chen, Zao Jiang, Zhenyi Zhong, Liang Wang, and Qiang Liu. Language models can understand spectra: A multimodal model for molecular structure elucidation. arXiv preprint arXiv:2508.08441 , 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Molspectllm: A molecular foundation model bridging spectroscopy, molecule elucidation, and 3d structure generation

    Shuaike Shen, Jiaqing Xie, Zhuo Yang, Antong Zhang, Shuzhou Sun, Ben Gao, Tianfan Fu, Biqing Qi, and Yuqiang Li. Molspectllm: A molecular foundation model bridging spectroscopy, molecule elucidation, and 3d structure generation. arXiv preprint arXiv:2509.21861 , 2025

    arXiv 2025

  11. [11]

    Spectro: A multi-modal approach for molecule elucidation using ir and nmr data

    Edwin Chacko, Rudra Sondhi, Arnav Praveen, Kylie L Luska, and Rodrigo Alejandro Vargas Hernandez. Spectro: A multi-modal approach for molecule elucidation using ir and nmr data. 2024

    2024

  12. [12]

    NMR-Solver : Automated structure elucidation via large-scale spectral matching and physics-guided fragment optimization

    Yongqi Jin, Jun-Jie Wang, Fanjie Xu, Xiaohong Ji, Zhifeng Gao, Linfeng Zhang, Guolin Ke, Rong Zhu, et al. NMR-Solver : Automated structure elucidation via large-scale spectral matching and physics-guided fragment optimization. arXiv preprint arXiv:2509.00640 , 2025

    Pith/arXiv arXiv 2025

  13. [13]

    Deepspinn--deep reinforcement learning for molecular structure prediction from infrared and 13 c nmr spectra

    Sriram Devata, Bhuvanesh Sridharan, Sarvesh Mehta, Yashaswi Pathak, Siddhartha Laghuvarapu, Girish Varma, and U Deva Priyakumar. Deepspinn--deep reinforcement learning for molecular structure prediction from infrared and 13 c nmr spectra. Digital Discovery , 3(4):818--829, 2024

    2024

  14. [14]

    Nmrexp: A database of 3.3 million experimental nmr spectra

    Jun-Jie Wang, Yongqi Jin, Chen-Yu Zhi, Yu-Jie Liu, Xu-Hao Huang, Fanjie Xu, Xiaohong Ji, Xi Fang, Haoyi Tao, Weinan E, Linfeng Zhang, Guolin Ke, and Rong Zhu. Nmrexp: A database of 3.3 million experimental nmr spectra. Scientific Data , 12(1), Dec 2025

    2025

  15. [15]

    Helmus and Christopher P

    Jonathan J. Helmus and Christopher P. Jaroniec. Nmrglue: an open source Python package for the analysis of multidimensional NMR data. Journal of Biomolecular NMR , 55(4):355--367, 2013

    2013

  16. [16]

    Ernst, Geoffrey Bodenhausen, and Alexander Wokaun

    Richard R. Ernst, Geoffrey Bodenhausen, and Alexander Wokaun. Principles of Nuclear Magnetic Resonance in One and Two Dimensions . Oxford University Press, Oxford, 1987

    1987

  17. [17]

    Bartholdi and R

    E. Bartholdi and R. R. Ernst. Fourier spectroscopy and the causality principle. Journal of Magnetic Resonance , 11(1):9--19, 1973

    1973

  18. [18]

    Cooley and John W

    James W. Cooley and John W. Tukey. An algorithm for the machine calculation of complex Fourier series. Mathematics of Computation , 19(90):297--301, 1965

    1965

  19. [19]

    An efficient algorithm for automatic phase correction of NMR spectra based on entropy minimization

    Li Chen, Zhiqiang Weng, LaiYoong Goh, and Marc Garland. An efficient algorithm for automatic phase correction of NMR spectra based on entropy minimization. Journal of Magnetic Resonance , 158(1--2):164--168, 2002

    2002

  20. [20]

    pybaselines : A Python library of algorithms for the baseline correction of experimental data

    Donald Erb. pybaselines : A Python library of algorithms for the baseline correction of experimental data

  21. [21]

    u ntert, and Kurt W \

    Reto Koradi, Martin Billeter, Max Engeli, Peter G \"u ntert, and Kurt W \"u thrich. Automated peak picking and peak integration in macromolecular NMR spectra using AUTOPSY . Journal of Magnetic Resonance , 135(2):288--297, 1998

    1998

  22. [22]

    NMR signal processing, prediction, and structure verification with machine learning techniques

    Carlos Cobas. NMR signal processing, prediction, and structure verification with machine learning techniques. Magnetic Resonance in Chemistry , 58(6):512--519, 2020

    2020

  23. [23]

    Rapid prediction of NMR spectral properties with quantified uncertainty

    Eric Jonas and Stefan Kuhn. Rapid prediction of NMR spectral properties with quantified uncertainty. Journal of Cheminformatics , 11(1):50, 2019

    2019

  24. [24]

    Computer-assisted structure elucidation ( CASE ): Current and future perspectives

    Mikhail Elyashberg and Dimitris Argyropoulos. Computer-assisted structure elucidation ( CASE ): Current and future perspectives. Magnetic Resonance in Chemistry , 59(7):669--690, 2021

    2021

  25. [25]

    Computationally-assisted discovery and structure elucidation of natural products

    Alfarius Eko Nugroho and Hiroshi Morita. Computationally-assisted discovery and structure elucidation of natural products. Journal of Natural Medicines , 73(4):687--695, 2019

    2019

  26. [26]

    Toolformer: Language models can teach themselves to use tools

    Timo Schick, Jane Dwivedi-Yu, Roberto Dessi, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. In Advances in Neural Information Processing Systems (NeurIPS) , 2023

    2023

  27. [27]

    Mrkl systems: A modular, neuro-symbolic architecture that combines large language models, external knowledge sources and discrete reasoning

    Ehud Karpas, Omri Abend, Yonatan Belinkov, Barak Lenz, Opher Lieber, Nir Ratner, Yoav Shoham, Hofit Bata, Yoav Levine, Kevin Leyton-Brown, Dor Muhlgay, Noam Rozen, Erez Schwartz, Gal Shachaf, Shai Shalev-Shwartz, Amnon Shashua, and Moshe Tenenholtz. Mrkl systems: A modular, neuro-symbolic architecture that combines large language models, external knowledg...

    Pith/arXiv arXiv 2022

  28. [28]

    React: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR) , 2023

    2023

  29. [29]

    Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face

    Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face. In Advances in Neural Information Processing Systems , 2023

    2023

  30. [30]

    Synergistic cross-modal learning for experimental nmr-based structure elucidation

    Fanjie Xu, Jinyuan Hu, Jingxiang Zou, Junjie Wang, Boying Huang, Zhifeng Gao, Xiaohong Ji, Zhong-Qun Tian, Fujie Tang, Jun Cheng, et al. Synergistic cross-modal learning for experimental nmr-based structure elucidation. arXiv preprint arXiv:2602.08752 , 2026

    arXiv 2026

  31. [31]

    Pubchem 2025 update

    Sunghwan Kim, Jie Chen, Tiejun Cheng, Asta Gindulyte, Jia He, Siqian He, Qingliang Li, Benjamin A Shoemaker, Paul A Thiessen, Bo Yu, et al. Pubchem 2025 update. Nucleic acids research , 53(D1):D1516--D1525, 2025

    2025

  32. [32]

    Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs

    Yu A Malkov and Dmitry A Yashunin. Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs. IEEE transactions on pattern analysis and machine intelligence , 42(4):824--836, 2018

    2018

  33. [33]

    Loeffler, Jiazhen He, Alessandro Tibo, Jon Paul Janet, Alexey Voronov, Lewis H

    Hannes H. Loeffler, Jiazhen He, Alessandro Tibo, Jon Paul Janet, Alexey Voronov, Lewis H. Mervin, and Ola Engkvist. Reinvent 4: Modern ai-driven generative molecule design. Journal of Cheminformatics , 16(1):20, 2024

    2024

  34. [34]

    Zare, and Patrick Riley

    Zhenpeng Zhou, Steven Kearnes, Li Li, Richard N. Zare, and Patrick Riley. Optimization of molecules via deep reinforcement learning. Scientific Reports , 9(1):10752, 2019

    2019

  35. [35]

    A review of reinforcement learning in chemistry

    Samuel Gow, Viktor Kagi, Jovana Stamenkovic, Iryna Korshunova, Pedro Neves, and Pietro Lio. A review of reinforcement learning in chemistry. Digital Discovery , 1(5):551--573, 2022

    2022

  36. [36]

    Evaluation of reinforcement learning in transformer-based molecular design

    Jiazhen He, Alessandro Tibo, Jon Paul Janet, Eva Nittinger, Christian Tyrchan, Werngard Czechtizky, and Ola Engkvist. Evaluation of reinforcement learning in transformer-based molecular design. Journal of Cheminformatics , 16(1):95, 2024

    2024

  37. [37]

    Contact electron-spin coupling of nuclear magnetic moments

    Martin Karplus. Contact electron-spin coupling of nuclear magnetic moments. The Journal of chemical physics , 30(1):11--15, 1959

    1959

  38. [38]

    Poincar \'e embeddings for learning hierarchical representations

    Maximilian Nickel and Douwe Kiela. Poincar \'e embeddings for learning hierarchical representations. In Advances in Neural Information Processing Systems , 2017

    2017

  39. [39]

    Learning continuous hierarchies in the lorentz model of hyperbolic geometry

    Maximilian Nickel and Douwe Kiela. Learning continuous hierarchies in the lorentz model of hyperbolic geometry. In Proceedings of the 35th International Conference on Machine Learning , pages 3779--3788, 2018

    2018

  40. [40]

    Hyperbolic neural networks

    Octavian-Eugen Ganea, Gary B \'e cigneul, and Thomas Hofmann. Hyperbolic neural networks. In Advances in Neural Information Processing Systems , 2018

    2018

  41. [41]

    Beyond lipschitz: Ranking binding affinity in hyperbolic space

    Kelin Wu, Xin Hong, Wenyu Zhu, Bowen Gao, Wei-Ying Ma, and Yanyan Lan. Beyond lipschitz: Ranking binding affinity in hyperbolic space. bioRxiv , 2026

    2026

  42. [42]

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

    Gengmo Zhou, Zhifeng Gao, Qiankun Ding, Hang Zheng, Hongteng Xu, Zhewei Wei, Linfeng Zhang, and Guolin Ke. Uni-mol: A universal 3d molecular representation learning framework. In The eleventh international conference on learning representations , 2023

    2023

  43. [43]

    Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension

    Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th annual meeting of the association for computational linguistics , pages 7871--7880, 2020

    2020

  44. [44]

    Sutton and Andrew G

    Richard S. Sutton and Andrew G. Barto. Reinforcement Learning: An Introduction . MIT Press, 2 edition, 2018

    2018

  45. [45]

    Api-bank: A comprehensive benchmark for tool-augmented llms

    Minghao Li, Yingxiu Zhao, Bowen Yu, Feifan Song, Hangyu Li, Haiyang Yu, Zhoujun Li, Fei Huang, and Yongbin Li. Api-bank: A comprehensive benchmark for tool-augmented llms. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages 3102--3120, 2023

    2023

  46. [46]

    Patil, Tianjun Zhang, Xin Wang, and Joseph E

    Shishir G. Patil, Tianjun Zhang, Xin Wang, and Joseph E. Gonzalez. Gorilla: Large language model connected with massive apis. In Advances in Neural Information Processing Systems , 2024

    2024

  47. [47]

    Wu Gao, Yu Qiao, and Ping Luo

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Yang, Junxiao Xu, Miao Zheng, Ce Zhang, Ming Li, Y. Wu Gao, Yu Qiao, and Ping Luo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 , 2024

    Pith/arXiv arXiv 2024

  48. [48]

    Molreasoner: Toward effective and interpretable reasoning for molecular llms

    Guojiang Zhao, Zixiang Lu, Yutang Ge, Sihang Li, Zheng Cheng, Haitao Lin, Lirong Wu, Hanchen Xia, Hengxing Cai, Wentao Guo, Hongshuai Wang, Mingjun Xu, Siyu Zhu, Guolin Ke, Linfeng Zhang, and Zhifeng Gao. Molreasoner: Toward effective and interpretable reasoning for molecular llms. arXiv preprint arXiv:2508.02066 , 2025

    arXiv 2025

  49. [49]

    Encouraging good processes without the need for good answers: Reinforcement learning for llm agent planning

    Zhiwei Li, Yong Hu, and Wenqing Wang. Encouraging good processes without the need for good answers: Reinforcement learning for llm agent planning. arXiv preprint arXiv:2508.19598 , 2025

    Pith/arXiv arXiv 2025

  50. [50]

    Introducing gpt-5.2

    OpenAI . Introducing gpt-5.2. https://openai.com/index/introducing-gpt-5-2/, 2025. Accessed: 2026-03-17

    2025

  51. [51]

    Qwen2.5 technical report

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Huang, Fei Li, Haoran Liu, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115 , 2024

    Pith/arXiv arXiv 2024

  52. [52]

    Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback....

    2022

  53. [53]

    Unilabos: An ai-native operating system for autonomous laboratories

    Jing Gao, Junhan Chang, Haohui Que, Yanfei Xiong, Shixiang Zhang, Xianwei Qi, Zhen Liu, Jun-Jie Wang, Qianjun Ding, Xinyu Li, Ziwei Pan, Qiming Xie, Zhuang Yan, Junchi Yan, and Linfeng Zhang. Unilabos: An ai-native operating system for autonomous laboratories. 2025

    2025

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.