Pith. sign in

REVIEW 3 major objections 5 minor 42 references

IR spectra are ambiguous, and test-time search that scores candidate molecules by spectral consistency can lift top-1 reconstruction accuracy by up to 26 percentage points across existing base models.

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

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

T0 review · deepseek-v4-flash

2026-08-01 01:21 UTC pith:SFQMITE3

load-bearing objection Useful test-time calibration method for IR-to-molecule, with real gains, but the evaluation has important caveats: SynSet results rest on 211 samples and the spectral-reward mechanism lacks a random-reward control. the 3 major comments →

arxiv 2607.27788 v1 pith:SFQMITE3 submitted 2026-07-30 cs.AI cs.CE

SpecCal: Ambiguity-Aware Candidate Calibration for Infrared Spectrum-Based Molecular Structure Reconstruction

Yixuan Chen , Bo Liu , Yusen Tan , Guokun Yang , Wenjie Du , Jun Xia This is my paper
classification cs.AI cs.CE
keywords infrared spectroscopyIR-to-molecule predictionmolecular structure reconstructioncandidate calibrationMonte Carlo tree searchspectral consistencytest-time scalingSMILES generation
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 paper argues that the ranked candidate sets produced by IR-to-molecule models are shaped more by learned generation preferences than by how well each candidate actually explains the observed spectrum. It claims that a training-free calibration stage, driven by a spectral-consistency score, can both re-rank existing candidates and discover new ones, improving top-k reconstruction. The method is shown to work across multiple base models and datasets, suggesting that test-time computation is a practical substitute for retraining. A sympathetic reader sees this as evidence that ambiguity-aware candidate calibration, not just better generation, is a viable path to improve molecular structure reconstruction from IR spectra.

Core claim

The paper establishes that applying Monte Carlo Tree Search over SMILES sequences, guided by a reward based on agreement between the observed IR spectrum and a spectrum predicted from the candidate molecule, consistently improves top-k hit rates at both exact SMILES and molecular scaffold levels. The strongest gains appear at top-1 and top-5, with reported increases such as from 37.1% to 54.4% on a real experimental benchmark and from 21.3% to 47.9% on a large synthetic benchmark. The method also increases structural diversity of the predicted set, measured by lower Tanimoto similarity and higher scaffold diversity. The framework requires no parameter updates and can be layered onto any base

What carries the argument

The central mechanism is a spectral-consistency reward S(y,s)=1/(1+SID(f_IR(y),s)), where f_IR is a pretrained forward IR spectrum predictor and SID is spectral information divergence. This reward is embedded in a Monte Carlo Tree Search that expands candidate SMILES sequences by resampling the most uncertain token position and completing the sequence autoregressively. A Maximal Marginal Relevance selection step balances reward and structural diversity when assembling the final candidate set.

Load-bearing premise

The search reward assumes that the pretrained forward IR predictor assigns a higher spectral-consistency score to the true molecule than to plausible decoys; the paper never directly verifies this ordering on the evaluated distributions.

What would settle it

Take a set of IR spectra with known ground-truth molecules, compute the spectral-consistency reward for the true molecule and for the top-ranked candidates produced by a base model, and test whether the true molecule is systematically among the highest-reward candidates. If it is not, the MCTS guidance is driven by a ranking signal that does not correlate with structural correctness, and the reported gains would not transfer to new distributions.

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

If this is right

  • If the central claim holds, existing IR-to-molecule models can be improved without retraining, merely by allocating more computation at inference time.
  • Calibration can rescue cases where the ground-truth molecule is absent from the original beam candidates, as the case study demonstrates.
  • The gains concentrate at top-1 and top-5, meaning the method sharpens the ranking of plausible structures rather than merely expanding a broad pool.
  • Increased diversity of the candidate set suggests that calibrated predictions better reflect the true ambiguity of IR spectra, where multiple structures may be consistent with one observation.
  • The ablation indicates that the quality of the forward spectrum predictor is the main driver of calibration performance.

Where Pith is reading between the lines

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

  • Editorial extension: the same pattern—test-time search with a forward-model reward—could transfer to other spectroscopy tasks such as NMR or MS, provided a reliable forward predictor exists; the paper does not test this.
  • Editorial extension: if the reward signal were explicitly validated to rank the true molecule above plausible decoys, the approach could compound with stronger base models; otherwise its ceiling depends on the forward predictor's fidelity.
  • Editorial extension: the entropy-guided expansion mutates only the single most uncertain token; targeting multiple uncertain positions or mutating scaffold-level fragments might yield further gains on scaffold hit rates, a hypothesis the paper does not explore.
  • Editorial extension: evaluating against a single ground truth may understate the method's value in genuinely ambiguous cases, where several structures are chemically valid.

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 / 5 minor

Summary. The paper proposes SpecCal, a test-time, training-free candidate calibration framework for IR-to-molecule prediction. Given a base model's ranked candidate SMILES, SpecCal reranks and expands the candidate set using Monte Carlo Tree Search guided by a spectral-consistency reward computed with a pretrained Chemprop-IR forward predictor, then applies Maximal Marginal Relevance to maintain diversity. The authors evaluate on the NIST dataset and a 211-sample subset of the SynSet dataset with IR2Mol, Patch Transformer, and CoCa as base models, reporting top-k SMILES/scaffold hit rates and diversity metrics. The central claim is that SpecCal consistently improves top-k reconstruction at both SMILES and scaffold levels across base models and datasets.

Significance. If the central claim is substantiated, this is a useful contribution: the framework is model-agnostic, requires no retraining, and provides a practical way to improve candidate sets in a fundamentally ambiguous inverse problem. The released code supports reproducibility, and the idea of using a forward spectral predictor as an inference-time reward is natural and potentially extendable to other spectroscopy tasks. However, the current empirical validation leaves the proposed mechanism unproven, and the 'consistent improvement' claim is statistically fragile due to small sample sizes, missing error bars, and several reported performance decreases.

major comments (3)
  1. [§4.4, Eq. (13)–(15)] The spectral-consistency reward is the sole guidance signal for MCTS, but the paper never validates that this reward ranks the ground-truth molecule above plausible decoys on the evaluated distributions. The case study (Fig. 7) reports a reward of 0.29 for the ground truth, but no reward distribution is given. The ablation in Fig. 4 replaces Chemprop-IR with a formula-matching score, which is a different reward rather than a no-reward control; moreover, Appendix B's formula score uses the target ground-truth formula, which is not part of the spectrum-only input in §3. Without a random-reward or unguided-MCTS control, the observed gains could be attributable to additional sampling/broadening rather than to spectral-consistency calibration. This should be addressed with a reward-validity analysis (e.g., ground-truth vs. decoy reward comparisons) and a control reward.
  2. [§5.1–5.2, Table 1] The claim of 'consistently improves' is not supported by the reported numbers. Several cells decline: Patch Transformer NIST SMILES Top-10 is -0.343 pp and SynSet Top-5 is -0.474 pp; IR2Mol SynSet Top-5/Top-10 are +0.474 pp. The SynSet evaluation is described as a 211-sample subset, so 0.474 pp corresponds to exactly one molecule. No error bars, confidence intervals, or paired significance tests are provided, and it is unclear whether the SynSet rows use the full synthetic test set or only the 211 overlapping samples. Please report the exact sample sizes in the table caption, include uncertainty estimates, use paired tests, and either soften 'consistent' or establish it statistically.
  3. [Appendix B / §5.3] The formula-reward ablation is a misleading baseline: it is given the target molecular formula, which is not part of the input spectrum under the problem definition in §3. Thus the statement that SpecCal 'still yields noticeable improvements even without a dedicated IR predictor' conflates a privileged oracle signal with a reward-free search. A fairer ablation would use only information available at inference (e.g., predicted formula from the base model) or a constant/random reward.
minor comments (5)
  1. [§4.3 vs Algorithm 1] Algorithm 1 writes expansion as y_k ∼ p_theta(·|y(z), s), but Section 4.3 defines expansion by sampling N_c tokens at an uncertain position and then completing the sequence. The algorithm should reflect the actual expansion procedure, including the sampling temperature that is later used in §5.3.
  2. [§5.3 Computational Overhead] The section is qualitative; it would be useful to report actual wall-clock time or comparisons of inference cost for the different search budgets, especially since the paper motivates SpecCal as a practical test-time strategy.
  3. [Conclusion] The conclusion states that SpecCal 'achieves state-of-the-art accuracy', but the paper contains no comparison to state-of-the-art test-time scaling or other candidate-calibration methods; only base-model comparisons are reported. Please revise or add such a comparison.
  4. [Table 1] Some table cells appear merged or missing separators (e.g., Patch Transformer NIST base row). The table also does not indicate that SynSet results use the 211-sample subset; this should be stated in the caption.
  5. [General] The manuscript contains template placeholders (conference name, ACM reference format, 'Do Not Use This Code') and should be formatted for the target venue before resubmission.

Circularity Check

0 steps flagged

No circularity: SpecCal's performance gains are an empirical, externally benchmarked result, not a derivation equivalent to its inputs.

full rationale

The central claim of SpecCal is that a training-free, MCTS-based calibration using a spectral-consistency reward improves top-k reconstruction. The reward S(y,s)=1/(1+SID(f_IR(y),s)) is defined in Eqs. 13-15 using a pretrained Chemprop-IR forward predictor from prior external work [25] and the observed spectrum s. No parameter of SpecCal is fitted to the evaluation labels, and the evaluation metric (top-k SMILES/scaffold hit rate) is measured against ground-truth structures, not against the reward itself. The objective in Eq. 4 is a design choice, not a mathematical consequence of the evaluation metric; the reported gains are empirical results on external benchmarks (NIST and SynSet). There is no self-citation chain: Chemprop-IR, the base models (IR2Mol, Patch Transformer, CoCa), and the datasets are all external prior work. The ablation replacing Chemprop-IR with a formula-matching reward shows the reward contributes independently, further supporting that the mechanism is not tautological. The skeptical concern that the reward may not rank true molecules above plausible decoys is a validity/correctness risk, not a circularity: it questions whether the guiding signal is useful, not whether the output is equivalent to the input by construction. The only stated limitation is computational overhead (§7), which is orthogonal to circularity. No equation in the paper reduces to another by construction, and no fitted parameter is renamed as a prediction. Therefore no significant circularity is present.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The central claim rests on a learned reward model (Chemprop-IR) whose validity as a structural oracle is unexamined, plus several hand-set search hyperparameters with unspecified defaults. No new physical entities are introduced.

free parameters (6)
  • exploration constant C = 1.5
    Hand-chosen and used throughout; no sensitivity analysis for accuracy is reported.
  • search iterations T = 30-50 depending on base model
    Per-model choice (IR2Mol 50, Patch 30, CoCa 50) with no tuning protocol stated.
  • expansion size Ne = 1-2
    Per-model choice (IR2Mol 1, Patch 2, CoCa 2); not ablated.
  • MMR trade-off lambda = not stated for main results
    Sensitivity analysis covers 0.4-1.0, but the default used in Tables 1-2 is unspecified.
  • sampling temperature = not stated for main results
    Mentioned in sensitivity analysis; default for main experiments not given.
  • token candidates Nc = not specified
    Number of tokens sampled at the uncertain position in Eq. (11) is not reported.
axioms (4)
  • domain assumption Chemprop-IR predicted spectra are a sufficiently accurate proxy for observed IR spectra to rank molecular candidates.
    Section 4.4 uses f_IR (pretrained Chemprop-IR) to predict spectra for candidate molecules and computes SID with the observed spectrum; no validation is provided that this ranking correlates with structural correctness on NIST experimental spectra.
  • domain assumption SID-based reward correlates with structural identity.
    Eqs. (14)-(15): lower SID is assumed to mean the candidate is more likely to be the true structure, but the paper itself notes that distinct molecules can have highly similar IR spectra, creating spectral ambiguity.
  • domain assumption Entropy-based token selection identifies positions where the base model is uncertain and worth editing.
    Section 4.3: high Shannon entropy at a token position is taken as a signal of low-confidence regions; no direct evidence that editing these positions produces valid diverse candidates.
  • domain assumption Base-model conditional distributions remain meaningful when sampling replacements and completing partial sequences.
    Eqs. (6), (11)-(12): the model p_theta is used to propose edits; the paper does not analyze whether such local edits produce chemically valid, diverse SMILES.

pith-pipeline@v1.3.0-daily-deepseek · 13941 in / 11101 out tokens · 96906 ms · 2026-08-01T01:21:06.318368+00:00 · methodology

0 comments
read the original abstract

Inferring molecular structures from infrared (IR) spectra is a fundamental yet challenging problem. A key difficulty is that an IR spectrum provides limited structural information: different molecules may share similar functional groups and local vibrational patterns, leading to highly similar spectral responses. Thus, even when an observed spectrum has a unique underlying structure, reconstructing it from the spectrum remains ambiguous. Existing IR-to-molecule models usually generate a ranked set of candidate molecules, but this set is largely determined by the model's learned generation preference and may not fully capture the structures that best satisfy the observed spectral constraints. To address this limitation, we propose SpecCal, a training-free candidate calibration framework for IR-to-molecule prediction. SpecCal operates on the candidate outputs of existing base models and improves the prediction set by re-ranking current candidates while introducing additional structurally plausible alternatives guided by spectral consistency. The framework is plug-and-play and model-agnostic, requiring no parameter updates for integration with diverse base models. Experiments on multiple benchmarks show that SpecCal consistently improves top-k reconstruction at both SMILES and scaffold levels across different base models. Further analyses demonstrate that calibrating candidate sets under spectral ambiguity provides a practical way to improve molecular reconstruction from IR spectra. The code is available at: https://anonymous.4open.science/r/SpecCal-B18A.

Figures

Figures reproduced from arXiv: 2607.27788 by Bo Liu, Guokun Yang, Jun Xia, Wenjie Du, Yixuan Chen, Yusen Tan.

Figure 1
Figure 1. Figure 1: Overview of SpecCal. An input IR spectrum is first translated into a set of candidate SMILES sequences by a pretrained [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (a) low diversity and low spectral consistency, (b) low diversity and high spectral consistency, (c) high diversity and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Bucket-wise analysis of Tanimoto similarity and [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Sensitivity analysis of SpecCal with respect to key [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study on the NIST dataset. Each plot shows [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Top-3 predictions from the base model. The ground [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Partial visualization of the SpecCal calibration pro [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

42 extracted references · 6 linked inside Pith

  1. [1]

    Marvin Alberts, Teodoro Laino, and Alain C Vaucher. 2024. Leveraging infrared spectroscopy for automated structure elucidation.Communications Chemistry7, 1 (2024), 268

    2024

  2. [2]

    Matthew J Baker, Júlio Trevisan, Paul Bassan, Rohit Bhargava, Holly J Butler, Konrad M Dorling, Peter R Fielden, Simon W Fogarty, Nigel J Fullwood, Kelly A Heys, et al. 2014. Using Fourier transform IR spectroscopy to analyze biological materials.Nature protocols9, 8 (2014), 1771–1791

    2014

  3. [3]

    Axel D Becke. 1993. Density-functional thermochemistry. III. The role of exact exchange.The Journal of chemical physics98, 7 (1993), 5648–5652

    1993

  4. [4]

    Guy W Bemis and Mark A Murcko. 1996. The properties of known drugs. 1. Molecular frameworks.Journal of medicinal chemistry39, 15 (1996), 2887–2893

    1996

  5. [5]

    Cameron B Browne, Edward Powley, Daniel Whitehouse, Simon M Lucas, Peter I Cowling, Philipp Rohlfshagen, Stephen Tavener, Diego Perez, Spyridon Samoth- rakis, and Simon Colton. 2012. A survey of monte carlo tree search methods. IEEE Transactions on Computational Intelligence and AI in games4, 1 (2012), 1–43

    2012

  6. [6]

    Jaime Carbonell and Jade Goldstein. 1998. The use of MMR, diversity-based reranking for reordering documents and producing summaries. InProceedings of the 21st annual international ACM SIGIR conference on Research and development in information retrieval. 335–336

    1998

  7. [7]

    Chein-I Chang. 1999. Spectral information divergence for hyperspectral image analysis. InIEEE 1999 International Geoscience and Remote Sensing Symposium. IGARSS’99 (Cat. No. 99CH36293), Vol. 1. IEEE, 509–511

    1999

  8. [8]

    Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al . 2024. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling.arXiv preprint arXiv:2412.05271(2024)

    Pith/arXiv arXiv 2024

  9. [9]

    Rémi Coulom. 2006. Efficient selectivity and backup operators in Monte-Carlo tree search. InInternational conference on computers and games. Springer, 72–83

    2006

  10. [10]

    Sriram Devata, Bhuvanesh Sridharan, Sarvesh Mehta, Yashaswi Pathak, Sid- dhartha Laghuvarapu, Girish Varma, and U Deva Priyakumar. 2024. DeepSPInN– deep reinforcement learning for molecular structure prediction from infrared and 13 C NMR spectra.Digital Discovery3, 4 (2024), 818–829

    2024

  11. [11]

    Shuo Feng, Meng Huang, Yanbo Li, Aoran Cai, Xiaoyu Yue, Song Wang, Linjiang Chen, Jun Jiang, and Yi Luo. 2025. Intelligent understanding of spectra: from structural elucidation to property design.Chemical Society Reviews54, 18 (2025), 8243–8286

    2025

  12. [12]

    Markus Freitag and Yaser Al-Onaizan. 2017. Beam search strategies for neural machine translation. InProceedings of the First Workshop on Neural Machine Translation. 56–60

    2017

  13. [13]

    Michael Gastegger, Jörg Behler, and Philipp Marquetand. 2017. Machine learning molecular dynamics for the simulation of infrared spectra.Chemical science8, 10 (2017), 6924–6935

    2017

  14. [14]

    Shibo Hao, Yi Gu, Haodi Ma, Joshua Hong, Zhen Wang, Daisy Wang, and Zhiting Hu. 2023. Reasoning with language model is planning with world model. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 8154–8173

    2023

  15. [15]

    Di He, Hanqing Lu, Yingce Xia, Tao Qin, Liwei Wang, and Tie-Yan Liu. 2017. Decoding with value networks for neural machine translation.Advances in Neural information processing systems30 (2017)

    2017

  16. [16]

    Henning Henschel, Alfred T Andersson, Willem Jespers, Mohammad Mehdi Ghahremanpour, and David Van der Spoel. 2020. Theoretical infrared spectra: quantitative similarity measures and force fields.Journal of chemical theory and computation16, 5 (2020), 3307–3315

    2020

  17. [17]

    Coleman Hooper, Sehoon Kim, Suhong Moon, Kerem Dilmen, Monishwaran Maheswaran, Nicholas Lee, Michael W Mahoney, Sophia Shao, Kurt Keutzer, and Amir Gholami. 2025. Ets: Efficient tree search for inference-time scaling.arXiv preprint arXiv:2502.13575(2025)

    Pith/arXiv arXiv 2025

  18. [18]

    Kas J Houthuijs, Giel Berden, Udo FH Engelke, Vasuk Gautam, David S Wishart, Ron A Wevers, Jonathan Martens, and Jos Oomens. 2023. An in silico infrared spectral library of molecular ions for metabolite identification.Analytical chem- istry95, 23 (2023), 8998–9005

    2023

  19. [19]

    Neena Jaggi and DR Vij. 2006. Fourier transform infrared spectroscopy. In Handbook of Applied Solid State Spectroscopy. Springer, 411–450

    2006

  20. [20]

    Ganesh Chandan Kanakala, Bhuvanesh Sridharan, and U Deva Priyakumar. 2024. Spectra to structure: contrastive learning framework for library ranking and generating molecular structures for infrared spectra.Digital Discovery3, 12 (2024), 2417–2423

    2024

  21. [21]

    Madanakrishna Katari, Edith Nicol, Vincent Steinmetz, Guillaume van der Rest, Duncan Carmichael, and Gilles Frison. 2017. Improved infrared spectra prediction by DFT from a new experimental database.Chemistry–A European Journal23, 35 (2017), 8414–8423

    2017

  22. [22]

    Tobias Kind and Oliver Fiehn. 2010. Advances in structure elucidation of small molecules using mass spectrometry.Bioanalytical reviews2, 1 (2010), 23–60

    2010

  23. [23]

    Levente Kocsis and Csaba Szepesvári. 2006. Bandit based monte-carlo planning. InEuropean conference on machine learning. Springer, 282–293

    2006

  24. [24]

    Peter J Linstrom and William G Mallard. 2001. The NIST Chemistry WebBook: A chemical data resource on the internet.Journal of Chemical & Engineering Data 46, 5 (2001), 1059–1063

    2001

  25. [25]

    Charles McGill, Michael Forsuelo, Yanfei Guan, and William H Green. 2021. Predicting infrared spectra with message passing neural networks.Journal of chemical information and modeling61, 6 (2021), 2594–2609

    2021

  26. [26]

    David Rogers and Mathew Hahn. 2010. Extended-connectivity fingerprints. Journal of chemical information and modeling50, 5 (2010), 742–754

    2010

  27. [27]

    Naseem Saquer, Razib Iqbal, Joshua D Ellis, and Keiichi Yoshimatsu. 2024. Infrared spectra prediction using attention-based graph neural networks.Digital Discovery 3, 3 (2024), 602–609

    2024

  28. [28]

    Claude Elwood Shannon. 1948. A mathematical theory of communication.The Bell system technical journal27, 3 (1948), 379–423

    1948

  29. [29]

    Shuaike Shen, Jiaqing Xie, Zhuo Yang, Antong Zhang, Shuzhou Sun, Ben Gao, Tianfan Fu, Biqing Qi, and Yuqiang Li. 2025. MolSpectLLM: A Molecular Foun- dation Model Bridging Spectroscopy, Molecule Elucidation, and 3D Structure Generation.arXiv preprint arXiv:2509.21861(2025)

    arXiv 2025

  30. [30]

    David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershel- vam, Marc Lanctot, et al . 2016. Mastering the game of Go with deep neural networks and tree search.nature529, 7587 (2016), 484–489

    2016

  31. [31]

    2011.Fundamentals of Fourier transform infrared spectroscopy

    Brian C Smith. 2011.Fundamentals of Fourier transform infrared spectroscopy. CRC press

    2011

  32. [32]

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2024. Scaling llm test- time compute optimally can be more effective than scaling model parameters. arXiv preprint arXiv:2408.03314(2024)

    Pith/arXiv arXiv 2024

  33. [33]

    Stephen E Stein and Donald R Scott. 1994. Optimization and testing of mass spectral library search algorithms for compound identification.Journal of the American Society for Mass Spectrometry5, 9 (1994), 859–866

    1994

  34. [34]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need.Advances in neural information processing systems30 (2017)

    2017

  35. [35]

    Ashwin K Vijayakumar, Michael Cogswell, Ramprasath R Selvaraju, Qing Sun, Ste- fan Lee, David Crandall, and Dhruv Batra. 2016. Diverse beam search: Decoding diverse solutions from neural sequence models.arXiv preprint arXiv:1610.02424 (2016)

    Pith/arXiv arXiv 2016

  36. [36]

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171(2022)

    Pith/arXiv arXiv 2022

  37. [37]

    David Weininger. 1988. SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules.Journal of chemical information and computer sciences28, 1 (1988), 31–36

    1988

  38. [38]

    Wenjin Wu, Ales Leonardis, Jianbo Jiao, Jun Jiang, and Linjiang Chen. 2025. Transformer-based models for predicting molecular structures from infrared spectra using patch-based self-attention.The Journal of Physical Chemistry A 129, 8 (2025), 2077–2085

    2025

  39. [39]

    Yilin Yang, Liang Huang, and Mingbo Ma. 2018. Breaking the beam search curse: A study of (re-) scoring methods and stopping criteria for neural machine translation. InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 3054–3059

    2018

  40. [40]

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of thoughts: Deliberate problem solving with large language models.Advances in neural information processing systems36 (2023), 11809–11822

    2023

  41. [41]

    Colin Zhang and Yang Ha. 2025. Toward Complete Molecular Structure Prediction from Infrared Spectroscopy Using Deep Learning.Journal of Chemical Information and Modeling66, 1 (2025), 100–109

    2025

  42. [42]

    Qiyuan Zhang, Fuyuan Lyu, Zexu Sun, Lei Wang, Weixu Zhang, Wenyue Hua, Haolun Wu, Zhihan Guo, Yufei Wang, Niklas Muennighoff, et al. 2025. A survey on test-time scaling in large language models: What, how, where, and how well? arXiv preprint arXiv:2503.24235(2025). A Evaluation Metrics To evaluate the structural diversity of the predicted molecular sets, ...

    Pith/arXiv arXiv 2025