Pith. sign in

REVIEW 5 major objections 4 minor 35 references

QFoldAgent: An Autonomous Quantum Optimization Multi-Agent System for Protein Structure Prediction

T0 review · 5 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read A closed-loop multi-agent system that iteratively refines Hamiltonian penalties systematically improves quantum protein-folding optimization.

desk verdict New closed-loop penalty-refinement idea, honestly evaluated, but the E/P evidence conflates penalty scale with optimization quality; the structural validity results are the real contribution. read the letter →

arxiv 2607.22549 v1 pith:6OTDI7ZD submitted 2026-05-11 cs.AI

Winson Chen , Yuqi Zhang , Sixu Chen , Nuo Xu , Qiang Guan , Caiwen Ding This is my paper · ORCID
classification cs.AI
keywords proteinstructurepredictionquantumoptimizationHamiltonianpenaltytuningvariationaleigensolverLLMagentsclosed-looprefinementlatticefoldingfeedbackcontrol
verification ladder T0 review T1 audit T2 compute T3 formal

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 tries to establish that the bottleneck in hybrid quantum-classical protein structure prediction is not the quantum solver alone but the manual setting of Hamiltonian penalty weights, and that this bottleneck can be automated by a closed-loop multi-agent system. A design agent proposes sequence-conditioned penalties; a variational-quantum pipeline optimizes the resulting Hamiltonian under simulated noise; a feedback agent uses energy-landscape diagnostics and structural-validation signals to refine penalties over cycles. On 55 ground-truth fragments, the loop reduces median RMSD from 3.64 to 3.20 Å, with the largest gains on the hardest targets, and on 100 unseen sequences it raises structural validity from 87.5% to 98.7%, recovering 87% of initially invalid cases. The key behavioral evidence is that a normalized energy-to-penalty ratio improves across cycles in 5 of 6 agent configurations while never improving in the same number of LLM-only sampling runs, isolating the feedback loop as the cause. If right, this points to agentic control over the optimization objective itself as a general lever for quantum applications.

What carries the argument

The load-bearing object is the closed-loop feedback mechanism operating on the optimization objective itself. A Design Agent converts sequence features (composition, flexibility, motif and interaction statistics) into three penalty coefficients for a Hamiltonian whose interaction term is fixed; a variational quantum eigensolver with a parameterized circuit and a limited iteration/shots budget optimizes the resulting objective under a noisy simulator; a Feedback Agent, following an LLM-as-judge pattern, receives energy-landscape metrics (convergence quality score, energy range) and structural validation signals (backbone-geometry statistics) and proposes refined penalties. The quantitative wo

What would settle it

Re-run the closed loop with the feedback agent's penalty suggestions replaced by random perturbations drawn from the same range, keeping everything else fixed; if E/P still improves in the same fraction of configurations and structural validity still rises to ~99%, then the specific content of the feedback is not the cause. Alternatively, recompute the same experiments with 10,000 shots per energy evaluation; if the E/P improvement trends disappear, the original energy estimates were too noisy to support the directed-optimization claim.

Watch

Extended reading notes

Core claim

The central claim is that iterative, feedback-guided refinement of the Hamiltonian's three constraint-penalty coefficients—chirality, backbone-geometry, and local-overlap terms—produces directed improvement that independent sampling does not. The agent pipeline never sees ground-truth RMSD; it adjusts penalties based on VQE energy trajectories and structural-validation metrics, and over three to five cycles the normalized energy-to-penalty ratio E/P decreases significantly in most agent configurations while remaining flat for LLM-only controls. Structural validity on unseen sequences rises from 87.5% to 98.7%, and 87% of initially invalid candidates are recovered by best-of-three selection;

Load-bearing premise

The behavioral claims rest on the assumption that VQE energies from a small fixed budget—50 optimizer iterations and 200 shots per evaluation, with the six lowest-energy runs retained—are reliable enough to distinguish good from bad penalty settings; if shot noise or local minima dominate, the energy-based evidence for directed optimization weakens.

Editorial extensions

If this is right

  • If correct, Hamiltonian penalty tuning in quantum protein workflows no longer requires a human expert in the loop; an autonomous agent can adapt the objective per sequence.
  • The same closed-loop architecture, with only the VQE execution layer swapped, is expected to extend to longer fragments and real quantum hardware.
  • Structural-validity gains (98.7% vs 87.5%) and 87% recovery of invalid predictions suggest agentic feedback materially reduces failure cases, which could matter for reliability in drug-discovery pipelines.
  • The contrast between agent and LLM-only trajectories implies that structured feedback, not more sampling, is the active ingredient—an encouraging sign for agent-guided optimization beyond proteins.
  • The finding that noise can increase improvement rates suggests that noisy intermediate-scale quantum settings may actually benefit more from adaptive penalty control than noiseless ones.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If penalty magnitude is as dominant as the ablations indicate (change in penalty explains R²=0.72 of energy improvement for the strongest controller), one testable extension is whether a much simpler scalar scheduler—e.g., a proportional controller on validity signals—matches the LLM feedback agent, isolating what the language model adds.
  • The paper's U-shaped structural-validity trajectory across five cycles suggests an exploration–exploitation tradeoff built into the feedback prompts; a natural follow-up is to make the exploration schedule explicit and tunable rather than left to the LLM.
  • Because the framework only exposes three penalty coefficients, a natural next experiment is to let the agent also modify the Hamiltonian's operator structure (which constraints are included), testing whether the benefit extends beyond coefficient tuning.
  • The result that weaker controllers can be 'rescued' by noise hints at an interaction between optimizer landscape roughness and feedback effectiveness; this could be probed directly by varying noise levels in finer steps and measuring the controller improvement rate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. QFoldAgent proposes a closed-loop, LLM-driven framework for tuning Hamiltonian penalty weights in a 5-residue tetrahedral-lattice protein-folding problem solved by VQE under Qiskit Aer noise. A Designer Agent proposes sequence-conditioned penalties, a VQE pipeline optimizes and decodes structures, and a Feedback Agent refines penalties across cycles using energy-landscape metrics and MolProbity validation. On 55 QDockBank fragments, the paper reports median RMSD improvement from 3.64 Å to 3.20 Å, while acknowledging per-target Wilcoxon tests are non-significant. On 100 unseen sequences, it reports structural validity gains (87.5% → 98.7%), 87% recovery of initially invalid cases, and E/P trajectory improvements in 5/6 agent configurations versus 0/6 for LLM-only controls. The central claim is that iterative agent control, not merely repeated LLM sampling, drives directed optimization and reduces failure cases.

Significance. If established, the framework would be a useful demonstration of LLM agents as controllers for quantum optimization workflows, with a clean separation between training/evaluation signals and a same-LLM no-feedback control. The authors contribute a reusable evaluation suite and a fixed VQE budget to isolate penalty adaptation. However, the core behavioral evidence is currently undermined by the definition and scaling of the E/P metric, and the endpoint RMSD gains are not statistically significant. The structural-validity and recovery results are real but are local-geometry outcomes rather than global-folding accuracy. The manuscript is therefore a promising proof-of-concept whose main quantitative claims require reanalysis before publication.

major comments (5)
  1. [§V-B2, Eq. (2), Table IV] The E/P metric does not isolate optimization quality. Since H_t = H_MJ + λ_chiral H_chiral + λ_back H_back + λ_1 H_overlap, E/P = <H_t>/Σλ. For a fixed or constraint-satisfying conformation, the penalty expectations vanish and E/P ≈ <H_MJ>/Σλ, so simply reducing total penalty improves E/P even if the decoded structure never changes. Conversely, increasing penalties with positive constraint expectations can improve E/P while worsening the raw energy. Table IV shows Qwen3.5 under large noise: mean final energy rises from 1.385 to 1.617 while total penalty rises from 3.44 to 4.51, so E/P falls from 0.403 to 0.359. The paper's 5/6 vs. 0/6 E/P contrast is therefore compatible with controllers that shrink or inflate penalties, not with improved optimization. Report raw energies and normalized metrics separately, and test whether E/P changes survive when the decoded structure is held fixed.
  2. [§V-D3] The claim that the change-based R² = 0.72 'cannot be explained by the Hamiltonian definition alone' is not demonstrated. If, over the explored region, E ≈ A + λ_chiral c_chiral + λ_back c_back + λ_1 c_1 with positive constraint expectations c_i, then regressing ΔE on Δ(total penalty) will produce a high R² whenever penalty changes are the dominant varying input, without any structural improvement. The same concern applies to the R² > 0.99 correlation between penalty_back and VQE energy, which the paper concedes is partly definitional. The directional claim—that the agent decreases penalties only when constraints are satisfied—needs a direct test: e.g., compare energy changes against decoded-structure quality or constraint-violation counts, not against penalty changes alone.
  3. [§IV-B(1), §V-C] The behavioral claims (E/P trajectories, per-cycle energy improvement, convergence quality) rely on VQE runs with random initialization, COBYLA limited to 50 iterations, and only 200 shots per energy evaluation. No repeated-seed error bars, confidence intervals, or per-target variance are reported. The Wilcoxon tests compare each target once; they do not address shot noise or local-minima variability. Without replicated runs, a paired trajectory could reflect stochastic VQE outcomes rather than directed penalty refinement. Report seed-averaged trajectories or bootstrap intervals for at least the main results.
  4. [Abstract, §V-B1] The abstract states the framework 'reduces median RMSD from 3.64 Å to 3.20 Å,' but the paper's own text reports that per-protein Wilcoxon tests are non-significant (p > 0.05) for every model–noise configuration, and that the comparison is between best-of-3 agent cycles and a single-pass fixed-penalty baseline. Moreover, QDockBank fragments are 5–14 residues while the agent uses fixed 5-residue sub-fragments. The endpoint claim should be rephrased as a descriptive, workload-level observation, not as a statistically supported improvement. This matters because the abstract and conclusion currently overstate the endpoint result.
  5. [§V-C1] The structural-validity comparison is one of the strongest results, but it needs clarification of the selection procedure. The agent is described as 'best-of-3 cycles,' and the LLM-only baseline runs 'the same number of independent cycles without feedback.' If the LLM-only baseline does not also select the best of 3 cycles, then the validity gain could partly reflect a best-of-k selection effect rather than feedback-driven improvement. State explicitly whether both conditions use identical best-of-k selection; if not, the 87.5% vs. 98.7% comparison is confounded.
minor comments (4)
  1. [§V-A] The definition of E/P as 'isolates optimization quality from penalty scale' is asserted without justification; given the major concern, this sentence should be removed or replaced with a caveat.
  2. [§I] Typo in the introduction: 'a closed-loop multi-agent framework to to enable' — 'to' is duplicated.
  3. [§IV-B] The text references 'Section III-C' when describing low-energy candidate retention, but the relevant material appears in §IV-B(1). Please fix the cross-reference.
  4. [§V-B1] The RMSD comparison with QDockBank is described as 'not a matched-length benchmark' in the text, but the abstract and several later statements omit this caveat. Add the caveat wherever endpoint gains are reported.

Circularity Check

2 steps flagged · score 6.0 of 10

E/P 'directed optimization' evidence is partly built into the Hamiltonian-penalty definition; validity results remain independent.

  1. self definitional [Section V-A (Evaluation metrics) and V-B2 (Directed optimization evidence); relation to Eq. (2) in Section IV Stage 1]
    "The normalized energy-to-penalty ratio E/P=E_final/Σλ isolates optimization quality from penalty scale; a monotonic decrease across cycles indicates directed refinement. [...] Across all three models under large noise, penalty_back is strongly associated with VQE final energy (R2 >0.99), although part of this relationship is definitional because penalty_back enters the Hamiltonian directly as the coefficient of Hback (Eq. 1)."

    In Eq. (2), H_t = H_MJ + λ_chiral H_chiral + λ_back H_back + λ_1 H_overlap, so any VQE state with positive constraint expectation has E_final increasing in each λ. E/P therefore responds mechanically to the agent's own penalty choices: GPT-5.2 shrinks Σλ, lowering both numerator and E/P; Qwen3.5 grows Σλ (Table IV: 3.44→4.51 under large noise) while its raw final energy worsens (1.385→1.617), yet E/P still falls (0.403→0.359) from denominator arithmetic alone. The paper's 5/6-vs-0/6 'directed refinement' contrast is thus not independent of the definition of the metric; the admitted definitional R2>0.99 relationship shows the energy scale is dominated by penalty variation, not structural quality.

  2. other [Section V-D3 (Penalty Dynamics)]
    "The more informative test is whether changes in penalty track changes in energy across cycles, which cannot be explained by the Hamiltonian definition alone."

    Under Eq. (2), ΔE_final = ΔH_MJ + Σ(Δλ_i <H_i> + λ_i Δ<H_i>), so when λ changes dominate, a change-based regression of energy on penalty is also expected from the same definitional coupling. The paper provides no counterfactual (e.g., fixing decoded structures) showing that the R2=0.72 is non-definitional, so this defense does not remove the circularity; it restates the energy-penalty coupling in first-difference form.

full rationale

The structural-validity result (87.5%→98.7%, 87% recovery) and the RMSD recoveries on ground-truth targets are externally grounded and not circular: RMSD is never fed to the agents, validity is MolProbity-derived, and the QDockBank comparison is an external baseline. However, the paper's central behavioral evidence for 'directed optimization' is the E/P trajectory, and that metric is partly defined by the same Hamiltonian it is supposed to evaluate: total energy includes the penalty terms whose coefficients the agent adjusts, and the denominator multiplies that confound. The authors admit the definitional link for penalty_back (R2>0.99). The change-based R2=0.72 defense does not break the coupling because first-differencing preserves the affine λ dependence whenever constraint expectations are positive. Thus the 5/6-vs-0/6 E/P contrast can be produced by penalty-scale drift rather than improved structure search, making part of the 'iterative control improves optimization' claim circular. The independent validity/recovery findings keep the paper from being fully circular, so the score is 6.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central quantitative claims depend on LLM-chosen penalty coefficients and several modeling proxies (lattice encoding, VQE budget, MODELLER/MolProbity validity). No new physical entities are introduced.

free parameters (5)
  • penalty_chiral (λ_chiral) = sequence-dependent; not reported per-sequence (total penalties in Table IV)
    LLM-proposed and feedback-refined coefficient of the chirality constraint term in Eq. (2); energy improvements depend on its trajectory.
  • penalty_back (λ_back) = sequence-dependent; decreased in 78-89% of cases for GPT-5.2
    Coefficient of H_back; ablation shows strong correlation with VQE energy, partly definitional.
  • penalty_1 (λ_1) = sequence-dependent
    Coefficient of the overlap/steric-conflict term in Eq. (2).
  • convergence quality score weights = 0.4, 0.35, 0.25
    Hand-chosen weights for q = 0.4 q_imp + 0.35 q_stab + 0.25 q_mono used in the feedback prompt; affects agent decisions.
  • VQE budget = 50 iterations, 200 shots/evaluation, 6 lowest-energy states, 100,000 final shots
    Fixed control budget; if insufficient, energy metrics may be noisy and the behavioral conclusions weakened.
assumptions (5)
  • domain assumption Tetrahedral lattice with 4 residue directions and ~109.4° bond angles captures enough stereochemistry for fragment folding.
    Adopted from QDockBank/ref [26]; central encoding on which all decoding and RMSD rests.
  • domain assumption Low-energy states of the Hamiltonian correspond to plausible fragment structures.
    Core premise of quantum folding; penalties enforce constraints, but the physical correctness of the mapping is assumed from the lattice model literature.
  • domain assumption VQE with EfficientSU2 + COBYLA, 50 iterations, 200 shots, is a sufficient optimizer proxy.
    Section IV-B(1); controls effort but also limits how much the energy metric reflects true optimization quality.
  • domain assumption MODELLER reconstruction and MolProbity metrics are reliable proxies for structural validity.
    Stages 2-3; used to evaluate unseen sequences where no ground-truth RMSD exists.
  • ad hoc to paper The LLM agents can translate metrics into useful penalty updates.
    Central mechanism; only indirectly evidenced by the empirical comparisons, not by independent validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of QFoldAgent: An Autonomous Quantum Optimization Multi-Agent System for Protein Structure Prediction." pith.science (2026). https://pith.science/paper/6OTDI7ZD

@misc{pith2026260722549,
  author       = {Pith},
  title        = {Pith review of: QFoldAgent: An Autonomous Quantum Optimization Multi-Agent System for Protein Structure Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6OTDI7ZD}},
  note         = {Machine review of arXiv:2607.22549}
}
read the original abstract

Hybrid quantum-classical protein structure prediction depends strongly on Hamiltonian penalty weights, yet existing lattice-based workflows typically fix these coefficients by hand and evaluate only very short fragments in simulation. We present QFoldAgent, a closed-loop multi-agent framework for 5-residue tetrahedral-lattice folding in which a design agent proposes sequence-conditioned penalties, a VQE-based quantum-classical pipeline optimizes the resulting Hamiltonian under Qiskit Aer noise, and a feedback agent uses energy-landscape diagnostics and MolProbity validation signals to refine penalties across cycles. Ground-truth metrics such as RMSD are never exposed to the agents and are used only for evaluation. We study the framework on two complementary datasets: 55 QDockBank-derived fragments with known structures and 100 coverage-optimized unseen sequences. On the QDockBank benchmark, QFoldAgent reduces median RMSD from 3.64 \AA{} to 3.20 \AA{}, with the largest gains on the hardest targets. On unseen sequences, the closed loop raises structural validity from 87.5% to 98.7%, recovers 87% of initially invalid cases, and the strongest controller improves cycle-3 energy on 87% of sequences while maintaining 96% Ramachandran-favored geometry. These results show that iterative agent control can systematically improve optimization behavior and reduce failure cases in a 5-residue quantum setting.

Figures

Figures reproduced from arXiv: 2607.22549 by the authors.

Figure 1
Figure 1. LLM-based mean RMSD vs. non-agent QDockBank baseline (dashed [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Traditional human-in-the-loop workflow for Hamiltonian refinement. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of the closed-loop quantum-classical pipeline for lattice-based protein structure prediction. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Quantum folding pipeline. The amino acid sequence is encoded into [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Best RMSD (A) on the five QDockBank targets with the largest ˚ improvement under small Aer noise. The agent drives several difficult cases below 2 A where the fixed-parameter baseline exceeds 4 ˚ A. ˚ validity—Ramachandran favored ≥95%, zero Cβ outliers, zero twisted p…
Figure 6
Figure 6. Figure 6: E/P trajectory on the 55 ground-truth targets. Agent configurations [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Per-sequence energy comparison between cycle 1 and cycle 3 for [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Mean total penalty (λchiral + λback + λ1) across 5 cycles (log scale). GPT-5.2 consistently reduces penalties, GPT-4o plateaus, and Qwen3.5 tends to increase. hide distinct controller behaviors: GPT-5.2’s improvements are broadly distributed across the dataset rather t…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 8 canonical work pages

  1. [1]

    Accurate structure prediction of biomolecular interactions with AlphaFold 3,

    J. Abramson, J. Adler, J. Dungeret al., “Accurate structure prediction of biomolecular interactions with AlphaFold 3,”Nature, vol. 630, pp. 493–500, 2024, https://doi.org/10.1038/s41586-024-07487-w

    doi:10.1038/s41586-024-07487-w 2024

  2. [2]

    How good are AlphaFold models for docking-based virtual screening?

    V . Scardino, J. I. Di Filippo, and C. N. Cavasotto, “How good are AlphaFold models for docking-based virtual screening?”iScience, vol. 26, no. 1, p. 105920, Jan. 2023, https://doi.org/10.1016/j.isci.2022.105920

    arXiv 2023

  3. [3]

    Challenging AlphaFold in predicting proteins with large- scale allosteric transitions,

    B. H. Perkins-Jechow, J. P. Iglesias Ahualli, H. T. Nhu, A. Omidi, C. Li, J. A. Holguin-Cruz, D. Cho, D. Na, N. Malhis, J. M. Bui, and J. Gsponer, “Challenging AlphaFold in predicting proteins with large- scale allosteric transitions,”Commun. Chem., vol. 8, no. 1, p. 378, Nov. 2025, https://doi.org/10.1038/s42004-025-01763-0

    work page doi:10.1038/s42004-025-01763-0 2025

  4. [4]

    Assessment of protein complex predictions in casp16: Are we making progress?

    J. Zhang, Q. Conget al., “Assessment of protein complex predictions in casp16: Are we making progress?”Proteins: Structure, Function, and Bioinformatics, 2025, https://doi.org/10.1002/prot.70068

    work page doi:10.1002/prot.70068 2025

  5. [5]

    Resource- efficient quantum algorithm for protein folding,

    A. Robert, P. K. Barkoutsos, S. Woerner, and I. Tavernelli, “Resource- efficient quantum algorithm for protein folding,”npj Quantum Informa- tion, vol. 7, p. 38, 2021, https://doi.org/10.1038/s41534-021-00368-4

    doi:10.1038/s41534-021-00368-4 2021

  6. [6]

    Large language model agents for biological intelligence across genomics, proteomics, spatial biology, and biomedicine,

    S. A. Dip, D. Mallick, U. Acharjee Shuvo, S. Barua Soumma, F. Rafsani, B. Kumar Paul, N. Ahmed Moumi, S. Ahmed, and L. Zhang, “Large language model agents for biological intelligence across genomics, proteomics, spatial biology, and biomedicine,”Briefings in Bioinformatics, vol. 27, no. 2, p. bbag110, 2026, https://doi.org/10.1093/bib/bbag110

    work page doi:10.1093/bib/bbag110 2026

  7. [7]

    Artificial intelligence agents for biological research: a survey,

    C. Qi, W. Wang, S. Jiang, Q. Liu, X. Song, H. Fang, and Z. Wei, “Artificial intelligence agents for biological research: a survey,”Briefings in Bioinformatics, vol. 27, no. 1, p. bbag075, 2026, https://doi.org/10. 1093/bib/bbag075

    2026

  8. [8]

    Qwen3.5: Towards native multimodal agents,

    Qwen Team, “Qwen3.5: Towards native multimodal agents,” Feb. 2026, https://qwen.ai/blog?id=qwen3.5

    2026

Show all 35 references
  1. [9]

    Gpt-4o system card,

    OpenAI, “Gpt-4o system card,” 2024, https://doi.org/10.48550/arXiv. 2410.21276

    2024 doi arXiv

  2. [10]

    Introducing gpt-5.2,

    OpenAI Research, “Introducing gpt-5.2,” Dec. 2025, https://openai.com/ index/introducing-gpt-5-2/

    2025

  3. [11]

    Qdockbank: A dataset for ligand docking on protein fragments predicted on utility-level quantum computers,

    Y . Zhang, Y . Yang, C.-C. Lu, W. Jiang, F. Cheng, B. Fang, and Q. Guan, “Qdockbank: A dataset for ligand docking on protein fragments predicted on utility-level quantum computers,” inProceedings of the International Conference for High Performance Computing, Networking, Stora...

    2025

  4. [12]

    Methods of protein structure comparison,

    I. Kufareva and R. Abagyan, “Methods of protein structure comparison,” inHomology Modeling: Methods and Protocols, A. J. W. Orry and R. Abagyan, Eds. Totowa, NJ: Humana Press, 2012, pp. 231–257, https://doi.org/10.1007/978-1-61779-588-6 10

    2012 doi

  5. [13]

    The variational quantum eigensolver: A review of methods and best practices,

    J. Tilly, H. Chen, S. Cao, D. Picozzi, K. Setia, Y . Li, E. Grant, L. Wossnig, I. Rungger, G. H. Booth, and J. Tennyson, “The variational quantum eigensolver: A review of methods and best practices,”Physics Reports, vol. 986, p. 1–128, Nov. 2022, https://doi.org/10.1016/j.phys...

    2022 doi

  6. [14]

    Ai agent systems: Architectures, applications, and evaluation,

    B. Xu, “Ai agent systems: Architectures, applications, and evaluation,” arXiv, 2026, https://doi.org/10.48550/arXiv.2601.01743

    2026 doi

  7. [15]

    A self-correcting multi-agent LLM framework for language-based physics simulation and explanation,

    D. Park, H. Moon, and S. Ryu, “A self-correcting multi-agent LLM framework for language-based physics simulation and explanation,”npj Artificial Intelligence, vol. 2, no. 1, p. 10, Jan. 2026, https://doi.org/10. 1038/s44387-025-00057-z

    2026

  8. [16]

    Verification-aware planning for multi-agent systems,

    T. Xu, D. Zhang, K. Mitra, and E. Hruschka, “Verification-aware planning for multi-agent systems,” inProceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), 2026, pp. 7528–7546, https://doi.org/10.186...

    2026

  9. [17]

    Artificial intelligence for quantum computing,

    Y . Alexeev, M. H. Farag, T. L. Patti, M. E. Wolf, N. Ares, A. Aspuru- Guzik, S. C. Benjamin, Z. Cai, S. Cao, C. Chamberlandet al., “Artificial intelligence for quantum computing,”Nature Communications, vol. 16, no. 1, p. 10829, 2025, https://doi.org/10.1038/s41467-025-65836-3

    2025 doi

  10. [18]

    Quantum agents,

    E. Sultanow, M. Tehrani, S. Dutta, W. J. Buchanan, and M. S. Khan, “Quantum agents,”arXiv, 2025, https://doi.org/10.48550/arXiv.2506. 01536

    2025 doi

  11. [19]

    Programming quantum computers with large language models,

    E. R. Henderson, J. M. Henderson, J. Ange, and M. A. Thornton, “Programming quantum computers with large language models,”arXiv, 2025, https://doi.org/10.48550/arXiv.2506.18125

    2025 doi arXiv

  12. [20]

    Finding low-energy conformations of lattice protein models by quantum annealing,

    A. Perdomo-Ortiz, N. Dickson, M. Drew-Brook, G. Rose, and A. Aspuru- Guzik, “Finding low-energy conformations of lattice protein models by quantum annealing,”Scientific Reports, vol. 2, no. 1, p. 571, Aug. 2012, https://doi.org/10.1038/srep00571

    2012 doi

  13. [21]

    A quantum alternating operator ansatz with hard and soft constraints for lattice protein folding,

    M. Fingerhuth, T. Babejet al., “A quantum alternating operator ansatz with hard and soft constraints for lattice protein folding,”arXiv preprint arXiv:1810.13411, 2018, https://doi.org/10.48550/arXiv.1810.13411

    2018 doi arXiv

  14. [22]

    Peptide conformational sampling using the Quantum Approximate Optimization Algorithm,

    S. Boulebnane, X. Lucas, A. Meyder, S. Adaszewski, and A. Montanaro, “Peptide conformational sampling using the Quantum Approximate Optimization Algorithm,”npj Quantum Information, vol. 9, no. 1, p. 70, Jul. 2023, https://doi.org/10.1038/s41534-023-00733-5

    2023 doi

  15. [23]

    Quantum-audit: Evaluating the reasoning limits of llms on quantum computing,

    M. Afane, K. Laufer, W. Wei, Y . Mao, J. Farooq, Y . Wang, and J. Chen, “Quantum-audit: Evaluating the reasoning limits of llms on quantum computing,”arXiv, 2026, https://doi.org/10.48550/arXiv.2602.10092

    2026 doi

  16. [24]

    A perspective on protein structure prediction using quantum computers,

    H. Doga, B. Raubenolt, F. Cumbo, J. Joshi, F. P. DiFilippo, J. Qin, D. Blankenberg, and O. Shehab, “A perspective on protein structure prediction using quantum computers,”Journal of Chemical Theory and Computation, vol. 20, no. 9, pp. 3359–3378, 2024, https://doi.org/10. 1021/...

    2024

  17. [25]

    Folding lattice proteins with quantum annealing,

    A. Irb ¨ack, O. Knuthson, M. Meli, and S. Mohanty, “Folding lattice proteins with quantum annealing,”Physical Review Research, vol. 4, no. 4, p. 043013, 2022, https://doi.org/10.1103/PhysRevResearch.4.043013

    2022 doi

  18. [26]

    Quantum optimization within lattice gauge theory model on a quantum simulator,

    Z. Yan, Z. Zhou, Y .-H. Zhou, Y .-C. Wang, X. Qiu, Z. Y . Meng, and X.-F. Zhang, “Quantum optimization within lattice gauge theory model on a quantum simulator,”npj Quantum Information, vol. 9, no. 1, Sep. 2023, http://dx.doi.org/10.1038/s41534-023-00755-z

    2023 doi

  19. [27]

    Comparative protein modelling by satisfaction of spatial restraints,

    A. ˇSali and T. L. Blundell, “Comparative protein modelling by satisfaction of spatial restraints,”Journal of Molecular Biology, vol. 234, pp. 779–815, 1993, https://doi.org/10.1006/jmbi.1993.1626

    1993

  20. [28]

    A direct search optimization method that models the objective and constraint functions by linear interpolation,

    M. J. D. Powell, “A direct search optimization method that models the objective and constraint functions by linear interpolation,” inAdvances in Optimization and Numerical Analysis, S. Gomez and J.-P. Hennart, Eds. Dordrecht: Springer Netherlands, 1994, pp. 51–67, https://doi....

    1994 doi

  21. [29]

    Autodock vina: Improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading,

    O. Trott and A. J. Olson, “Autodock vina: Improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading,”Journal of Computational Chemistry, vol. 31, no. 2, pp. 455–461, 2010, https://doi.org/10.1002/jcc.21334

    2010 doi

  22. [30]

    Molprobity: More and better reference data for improved all-atom structure validation,

    C. J. Williams, J. J. Headd, N. W. Moriarty, M. G. Prisant, L. L. Videau, L. N. Deis, V . Verma, D. A. Keedy, B. J. Hintze, V . B. Chen, S. Jain, S. M. Lewis, W. B. Arendall III, J. Snoeyink, P. D. Adams, S. C. Lovell, J. S. Richardson, and D. C. Richardson, “Molprobity: More ...

    2018 doi

  23. [31]

    Judging llm-as-a-judge with mt-bench and chatbot arena,

    L. Zheng, W.-L. Chiang, Y . Sheng, S. Zhuang, Z. Wu, Y . Zhuang, Z. Lin, Z. Li, D. Li, E. P. Xing, H. Zhang, J. E. Gonzalez, and I. Stoica, “Judging llm-as-a-judge with mt-bench and chatbot arena,” 2023, https://doi.org/10.48550/arXiv.2306.05685. [Online]. Available: https://a...

    2023 doi arXiv

  24. [32]

    Individual comparisons by ranking methods,

    F. Wilcoxon, “Individual comparisons by ranking methods,”Biometrics Bulletin, vol. 1, no. 6, pp. 80–83, 1945, https://doi.org/10.2307/3001968. [Online]. Available: http://www.jstor.org/stable/3001968

    1945

  25. [33]

    A solution for the best rotation to relate two sets of vectors,

    W. Kabsch, “A solution for the best rotation to relate two sets of vectors,” Acta Crystallographica Section A, vol. 32, pp. 922–923, 1976, https: //doi.org/10.1107/S0567739476001873

    1976 doi

  26. [34]

    IBM Quantum breaks the 100-qubit processor barrier,

    J. Chow, O. Dial, and J. Gambetta, “IBM Quantum breaks the 100-qubit processor barrier,” IBM Research Blog, 2021, https://www.ibm.com/ quantum/blog/127-qubit-quantum-processor-eagle

    2021

  27. [35]

    Evidence for the utility of quantum computing before fault tolerance,

    Y . Kim, A. Eddins, S. Anand, K. X. Wei, E. van den Berg, S. Rosen- blatt, H. Nayfeh, Y . Wu, M. Zaletel, K. Temme, and A. Kandala, “Evidence for the utility of quantum computing before fault tolerance,” Nature, vol. 618, no. 7965, pp. 500–505, 2023, https://doi.org/10.1038/ s...

    2023

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.