pith. sign in

arxiv: 2605.23934 · v1 · pith:JHXFL2NYnew · submitted 2026-04-24 · 💻 cs.AI · quant-ph

Practical Quantum CIM Empowerment via All-Domestic-Core Agentic Large Model

Wang Rui , Lu Diannan This is my paper

Pith reviewed 2026-07-04 19:30 UTC · model glm-5.2

classification 💻 cs.AI quant-ph
keywords quantum computingCoherent Ising Machinelarge language modelsagentic systemsQUBOIsing modelcombinatorial optimizationflexible job-shop scheduling
0
0 comments X

The pith

LLM solves scheduling optimum after quantum hardware fails

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

The paper documents a system where a large language model (Doubao-1.5-Pro) orchestrates a Coherent Ising Machine (CIM) — a specialized quantum optimizer — through a ReAct-style agentic workflow built on LangGraph and LangChain. The agent handles natural-language problem intake, QUBO/Ising model construction, constraint-weight tuning, and result diagnosis across two tasks: a mass-spectrometry-inspired combinatorial problem and a flexible job-shop scheduling problem (FJSP). The system's primary intended contribution is demonstrating that an LLM can serve as the cognitive driver for quantum hardware, lowering the expertise barrier for quantum optimization. The paper's secondary and more provocative claim is the observation of a 'quantum-enhanced agent paradigm': after the CIM produced only a suboptimal makespan of 14 on the FJSP instance, the LLM — now carrying contextual experience from the full quantum iteration history — autonomously reasoned its way to the global optimum of 11 in a follow-up dialogue, without any further hardware calls. A control group given the same problem and prompt but without the quantum-iteration context failed to produce a feasible solution. The authors propose that the quantum iteration process, even when it does not converge to an optimum, supplies structured feedback that enhances the LLM's own problem-solving capability — a direction they call 'quantum-iteration-empowered LLM reasoning.'

Core claim

The central object is a closed-loop agentic workflow where an LLM orchestrates physical CIM hardware for combinatorial optimization. The paper's most distinctive claim is that the byproduct of failed or incomplete quantum optimization — the diagnostic context, constraint-violation patterns, and weight-tuning history generated during iteration — can serve as a form of experiential context that improves the LLM's subsequent reasoning. In the FJSP case, the CIM achieved makespan 14 with machine conflicts; after iterative weight tuning resolved the conflicts but not the suboptimality, the LLM used the accumulated context to derive the global optimum (makespan 11) through pure reasoning. The same

What carries the argument

The system architecture consists of: (1) a Central Model Agent (Doubao-1.5-Pro) handling decision-making and multi-turn dialogue reasoning; (2) a LangGraph workflow with eight ReAct nodes (Parsing, Modeling, Solving, Result Interpretation, Evaluation, Memory, Decision-Making, Weight Adjustment); (3) a LangChain tool layer encapsulating QUBO/Ising construction and CIM hardware invocation via the KaiwuAPI; and (4) a physical femtosecond-laser-pumped CIM from QBoson. The CIM maps combinatorial problems onto Ising spin variables via degenerate optical parametric oscillators, returning up to ten candidate spin configurations per solve. The agent's role is confined to constraint-weight tuning and

If this is right

  • The paper suggests that quantum hardware need not produce optimal solutions to be useful — the process of attempting quantum optimization generates structured diagnostic context that can improve classical AI reasoning, which is a distinct value proposition from the standard quantum-speedup narrative.
  • If the quantum-iteration-empowered reasoning effect generalizes, it would imply that even noisy, low-precision quantum hardware (the CIM's 8-bit quantization caused substantial constraint violations) could still contribute to problem-solving by providing exploratory context rather than exact solutions.
  • The hard-coded modeling architecture acknowledges a current capability ceiling: LLMs cannot yet autonomously generate correct QUBO matrices for novel problems, limiting the system's generality to pre-structured problem classes.
  • The finding that domestic (Chinese) LLMs can match international models on quantum-assistant tasks, despite token-efficiency and prompt-cache gaps, has implications for technology sovereignty in quantum-AI integration pipelines.

Where Pith is reading between the lines

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

  • The observed 'quantum enhancement' may be indistinguishable from a context-window effect: the LLM that saw iteration history had more problem-specific information than the control, regardless of whether that information came from quantum hardware or from any systematic search process. Disentangling quantum-specific feedback from generic structured search feedback would require a control group that
  • The FJSP instance (3 jobs, 3 machines, 9 operations) is small enough that an LLM with sufficient context could plausibly enumerate or near-enumerate solutions, raising the question of whether the effect would persist on instances large enough to be genuinely hard for classical reasoning but still within CIM's qubit range.
  • If the effect is real and quantum-specific, one could test this by comparing LLM post-iteration reasoning quality across hardware with different noise characteristics — if noisier hardware produces richer diagnostic context, the LLM's post-iteration reasoning should improve more, which would be a counterintuitive and testable prediction.

Load-bearing premise

The paper observes a single instance where the LLM derived the FJSP optimum after quantum iteration, and treats this as evidence of a generalizable paradigm. The control comparison tests whether having prior iteration context helps (it does), but does not isolate whether the quantum-specific nature of that context matters versus any structured problem exploration producing equivalent context. The reasoning log was captured only via console output, and the authors themselves (

What would settle it

Run the same FJSP instance and multiple new instances with: (a) LLM + CIM iteration context, (b) LLM + classical simulated annealing iteration context with equivalent diagnostic information, (c) LLM + random search iteration context. If the quantum-iteration group outperforms the classical-iteration group in post-iteration reasoning quality, the quantum-specific enhancement claim is supported; if not, the effect reduces to a generic context-availability phenomenon.

read the original abstract

Quantum computing devices are recognized as powerful tools for solving NP-complete problems. However, the intricacy of their modeling presents notable barriers for non-specialists, while the tedious iteration of constraint weights and modeling methodologies also consumes substantial effort on the part of experts. To address these challenges, this study integrates a femtosecond laser-pumped Coherent Ising Machine (CIM) with an LLM-driven agentic system by leveraging the LangGraph and LangChain frameworks. Comprehensive investigations demonstrate that large language models (LLMs) can effectively perform such tasks in modeling as QUBO/Ising model calibration, constraint weight decision iteration and rapid validation of literature-reported schemes. Notably, all these tasks can be fully implemented based on domestic large models, combined with domestically developed CIM hardware, we truly achieve the practical empowerment of quantum CIM that fully relies on all-domestic agentic large models and hardware. This work successfully realizes robust technological integration, laying a solid foundation for subsequent research. Nevertheless, it also identifies the persisting challenges in the two cutting-edge fields of large models and quantum computing at the current stage. Encouragingly, we unexpectedly discover a promising new paradigm where accumulated knowledge from agent-assisted quantum computing iterations reciprocally enhances the agent's own problem-solving capability, thereby addressing these challenges.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

4 major / 8 minor

Summary. The manuscript presents an integration framework combining a domestic LLM (Doubao-1.5-Pro) with a physical Coherent Ising Machine (CIM) from QBoson, using LangGraph/LangChain orchestration. The system is evaluated on two tasks: (1) a mass-spectrometry-inspired subset-sum QUBO problem, where the agent performs constraint-weight iteration but the physical CIM fails to produce accurate solutions due to 8-bit quantization and hardware noise; and (2) a 3×3 Flexible Job-Shop Scheduling Problem (FJSP) benchmark from Fu et al. [24], where the agent diagnoses sign errors in Ising conversion, tunes penalty weights, and achieves a feasible makespan of 14. The paper's central novelty claim is the 'quantum-enhanced agent paradigm': after the FJSP quantum iterations, the LLM autonomously derived the global optimum (makespan 11) in a follow-up dialogue, which a control group without prior quantum-iteration context could not achieve.

Significance. The system integration of a domestic LLM with physical CIM hardware via a ReAct workflow is a legitimate engineering contribution, and the candid documentation of hardware limitations (8-bit quantization, noise) for fine-grained optimization is valuable. The FJSP error-diagnosis pipeline (§3.2.3) demonstrates a meaningful agent capability: autonomous detection and correction of sign errors in Ising conversion. However, the paper's headline discovery — 'quantum-iteration-empowered LLM reasoning' — rests on a single uncontrolled instance with no statistical validation, and the control comparison is confounded (see Major Comments). The claim is presented as a 'novel paradigm' but the evidence does not support that level of generality.

major comments (4)
  1. §3.3, control group design: The control group was 'fed with exactly the same prompt and instance details' but lacked 'the contextual information of the complete pre-quantum optimization iteration.' This means the control lacked not just quantum-specific feedback but ALL multi-turn reasoning context — constraint analysis, penalty-weight tuning, solution diagnosis, Gantt chart construction. The improvement in the experimental group could be entirely explained by extended multi-turn problem engagement rather than any quantum-specific signal. To attribute the enhancement to 'quantum iteration' specifically, a proper control would need to provide equivalent multi-turn reasoning context without quantum hardware involvement (e.g., classical solver iterations, or even the same reasoning steps without CIM calls). Without this, the central discovery claim is unsupported.
  2. §3.3 and Appendix (LLM reasoning log): The reasoning log shows purely classical combinatorial reasoning — trying machine assignments, checking conflicts, computing completion times. No quantum-specific insight (spin configuration, energy landscape analysis, CIM-specific feedback) appears in the reasoning chain. The paper itself acknowledges the reasoning was 'more as a heuristic enumeration of known suboptimal solutions than a rigorous optimization' (§3.3). This undercuts the claim that 'accumulated knowledge from agent-assisted quantum computing iterations reciprocally enhances the agent's own problem-solving capability' (Abstract, §4). The LLM appears to have simply continued reasoning about the FJSP instance after being primed with problem structure — a well-known context-priming effect that does not require quantum iteration.
  3. §3.3, statistical validity: The 'discovery' rests on a single instance (3×3 FJSP, one experimental run, one control comparison). There is no replication across multiple instances, no statistical comparison, and no mechanism proposed for why quantum iteration context would systematically improve LLM reasoning. The Appendix notes the reasoning 'was only captured via console output and not pre-logged' and that reproduction 'if necessary' is planned. A single anecdotal observation cannot support the claim of a 'new paradigm' (§4).
  4. §3.1: The mass-spectrometry task failed to produce accurate solutions (minimum mass deviation 52.3 Da, >3.6% relative error). While the paper candidly documents this, it means one of the two validation tasks did not achieve its objective. The paper should more clearly frame this as a negative result rather than embedding it within a narrative of successful 'cognitive measurements.'
minor comments (8)
  1. Abstract: The phrase 'we unexpectedly discover a promising new paradigm' overstates the evidence. Consider toning down to 'we observe a phenomenon' pending stronger validation.
  2. §2.6.4: The decision to hard-code QUBO modeling nodes is described as a pragmatic necessity, but the axiom that hard-coding is a 'necessary and sufficient substitute for autonomous code generation' is not independently validated. A brief comparison of hard-coded vs. agent-generated modeling accuracy would strengthen this justification.
  3. Table 1: The processing time table is clear, but the FJSP instance should be explicitly labeled with its source (is this from Fu et al. [24] or newly constructed?). The text says 'we present a new canonical benchmark instance' but the benchmark formulation is taken from [24].
  4. Fig. 5 and Fig. 6: The Gantt charts are referenced but the figure quality and labeling should be verified in the final version — machine conflict visualization (Fig. 5) and the two scheduling solutions (Fig. 6a, 6b) need clear axis labels and legends.
  5. Appendix (LLM reasoning log): The raw reasoning dialogue is lengthy and somewhat difficult to follow. Consider summarizing the key reasoning steps in a structured format (e.g., a table of attempted makespan values and feasibility outcomes) while retaining the full log as supplementary material.
  6. References [2]: Citing the '15th Five-Year Plan' government document as a scientific reference for quantum computing motivation is unusual for a technical journal. Consider replacing with a peer-reviewed source.
  7. §3.2.3: The sign error in Ising conversion (Eq. 6) and the sign inversion of H4 are described as autonomously located by the agent. It would strengthen the claim to show the specific diagnostic reasoning the agent used to identify these errors, rather than just stating the outcome.
  8. Data Availability: 'We will release and continuously update the code and all data required to verify the presented conclusions on Github' — no URL or DOI is provided. This should be included for reproducibility.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for a careful and constructive review. The referee correctly identifies that the engineering contributions (system integration, error-diagnosis pipeline, candid hardware-limitation documentation) are solid, while the headline 'quantum-enhanced agent paradigm' claim is insufficiently supported by the current evidence. We agree with the core of each major comment and will revise accordingly: (1) we will add a proper control group that provides equivalent multi-turn reasoning context without quantum hardware involvement; (2) we will reframe the §3.3 finding as an exploratory observation rather than a 'novel paradigm'; (3) we will conduct additional replication experiments across multiple FJSP instances and report statistical comparisons; and (4) we will reframe the mass-spectrometry task as a negative result. We cannot fully resolve the confound identified in Major Comment 1 with the existing data — this requires new experiments that we commit to performing in the revision.

read point-by-point responses

  1. Referee: §3.3, control group design: The control group lacked ALL multi-turn reasoning context, not just quantum-specific feedback. The improvement could be explained by extended multi-turn problem engagement rather than any quantum-specific signal. A proper control would need equivalent multi-turn reasoning context without quantum hardware involvement.

    Authors: The referee is correct. The current control group is confounded: it lacks the entire multi-turn reasoning context (constraint analysis, penalty-weight tuning, solution diagnosis, Gantt chart construction), not just quantum-specific feedback. Therefore, the observed improvement cannot be attributed specifically to quantum iteration. We accept this criticism fully. In the revision, we will design and conduct a proper control experiment in which the agent receives equivalent multi-turn reasoning context — including the same constraint analysis, weight-tuning steps, and solution diagnosis — but with CIM calls replaced by a classical solver (or the same reasoning steps without any hardware invocation). This will isolate whether quantum-specific feedback contributes beyond the multi-turn engagement effect. We acknowledge that with the existing data, we cannot rule out the alternative explanation the referee proposes. The revised manuscript will either provide evidence for a quantum-specific contribution or, if the classical-control group performs comparably, we will report that honestly and further downgrade the claim. revision: yes

  2. Referee: §3.3 and Appendix (LLM reasoning log): The reasoning log shows purely classical combinatorial reasoning with no quantum-specific insight. The paper itself acknowledges the reasoning was 'more as a heuristic enumeration of known suboptimal solutions than a rigorous optimization.' This undercuts the claim that quantum iteration 'reciprocally enhances the agent's own problem-solving capability.' The LLM appears to have simply continued reasoning after being primed with problem structure — a well-known context-priming effect.

    Authors: We agree with the referee's observation. The reasoning log indeed contains no quantum-specific insight (no spin configuration analysis, no energy landscape reasoning, no CIM-specific feedback utilization). The reasoning is purely classical combinatorial enumeration. We also agree that context priming is a well-known phenomenon that could explain the observed improvement without requiring any quantum-specific mechanism. In the revision, we will (a) explicitly acknowledge this alternative explanation, (b) present the finding as an exploratory observation about the value of structured multi-turn problem engagement (whether quantum or classical) for LLM reasoning, and (c) remove or substantially soften the language claiming that 'quantum iteration' specifically enhances LLM reasoning. The abstract and §4 will be revised to reflect this more cautious framing. We will not claim a 'novel paradigm' based on the current evidence. revision: yes

  3. Referee: §3.3, statistical validity: The 'discovery' rests on a single instance (3×3 FJSP, one experimental run, one control comparison). No replication, no statistical comparison, no mechanism proposed. The Appendix notes the reasoning 'was only captured via console output and not pre-logged' and reproduction 'if necessary' is planned. A single anecdotal observation cannot support the claim of a 'new paradigm.'

    Authors: This is a fair criticism. A single unreplicated observation cannot support a paradigm-level claim. We will address this in two ways. First, we will downgrade the claim from 'novel paradigm' to 'exploratory observation warranting further investigation.' Second, we will conduct additional experiments: (a) multiple runs on the same 3×3 FJSP instance to assess run-to-run variability, (b) additional FJSP instances of varying sizes to test generalizability, and (c) the proper control group described in our response to Major Comment 1. We will also pre-log and properly archive all reasoning traces for reproducibility. If the additional experiments do not consistently replicate the effect, we will report that transparently. The revised manuscript will present this section as a preliminary finding with appropriate hedging, not as an established discovery. revision: yes

  4. Referee: §3.1: The mass-spectrometry task failed to produce accurate solutions (minimum mass deviation 52.3 Da, >3.6% relative error). The paper should more clearly frame this as a negative result rather than embedding it within a narrative of successful 'cognitive measurements.'

    Authors: We accept this criticism. The mass-spectrometry task did not achieve its objective — the minimum mass deviation of 52.3 Da (3.6% relative error) is far from the <0.1% precision achievable in simulation, and the task cannot be considered successful. In the revision, we will reframe §3.1 explicitly as a negative result: the agent successfully performed the modeling and iteration workflow, but the physical CIM hardware could not deliver solutions meeting the task's precision requirements due to 8-bit quantization and hardware noise. We will remove the language framing this as 'cognitive measurements' that validated the system's solving capability, and instead present it as a candid documentation of current hardware limitations. The value of this section lies in the agent's autonomous diagnostic capability (identifying quantization and noise as bottlenecks) and in the honest reporting of hardware constraints — not in successful problem-solving. revision: yes

Circularity Check

0 steps flagged

No significant circularity; one minor self-citation that is not load-bearing for the central claims.

full rationale

The paper's two main claims are empirical demonstrations, not derivations that reduce to their inputs by construction. (1) The FJSP QUBO formulation, penalty weights (α=150, β=100, γ=100, δ=15), and variable pruning are explicitly attributed to Fu et al. [24] as a benchmark reimplementation task (§3.2), not claimed as an independent first-principles derivation. (2) The 'quantum-enhanced reasoning' claim (§3.3) is an empirical observation: the LLM derived makespan 11 after quantum iteration context, while a control without that context failed. The reasoning log (Appendix) shows purely classical combinatorial reasoning with no quantum-specific insight, and the paper itself acknowledges this was 'more as a heuristic enumeration of known suboptimal solutions than a rigorous optimization.' The control group is confounded (it lacks all prior multi-turn reasoning context, not just quantum-specific feedback), but this is an experimental design weakness, not circularity in the derivation sense. The one self-citation is reference [48] (R. Wang's QUBO-VAE GitHub project), used to justify that the KaiwuAPI can be integrated into Markov decision workflows — an infrastructure claim, not load-bearing for either central result. No step in the paper's chain reduces to its own inputs by definition or by fitted-parameter renaming. The claims may be overstated relative to the evidence, but that is a correctness concern, not a circularity concern. Score 2 reflects the minor, non-load-bearing self-citation and the absence of any construction-level circularity in the derivation chain.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The framework introduces one invented conceptual entity (quantum-iteration-empowered reasoning) with no independent evidence beyond a single anecdotal instance. The free parameters are penalty weights iteratively tuned by the agent, which is standard for QUBO but means the 'solutions' are products of fitting. The key ad-hoc axiom is that hard-coding modeling is an acceptable substitute for autonomous code generation, which limits the system's generality.

free parameters (4)
  • λ_pos (position one-hot penalty weight) = iteratively tuned by agent
    The agent performs gradient-based tuning of this weight across multiple iterations (§3.1).
  • λ_mass (mass matching penalty weight) = iteratively tuned by agent
    The agent tunes this weight to balance against the position constraint (§3.1).
  • α, β, γ, δ (FJSP penalty weights) = 150, 100, 100→500, 15
    Initial values from Fu et al. [24]; γ is iteratively tuned from 100 to 500 by the agent (§3.2.3).
  • t_max (maximum time parameter for FJSP) = 18
    Set by the agent for the 3×3 FJSP instance (§3.2.2).
axioms (4)
  • domain assumption The Doubao-1.5-Pro LLM is capable of understanding QUBO/Ising modeling and performing iterative weight decisions.
    This is the foundational assumption of the entire framework; the paper provides qualitative evidence but no systematic benchmark of the model's reasoning accuracy (§2.6.2).
  • domain assumption The QBoson CIM hardware correctly implements the Ising Hamiltonian minimization.
    The framework relies on the CIM returning ground-state solutions; the paper documents significant precision limitations (8-bit quantization, noise) that undermine this assumption in practice (§3.1).
  • standard math The ReAct paradigm (Yao et al. [50]) is suitable for quantum optimization workflows.
    The LangGraph workflow is built on this paradigm; the paper provides qualitative evidence of its applicability (§2.3).
  • ad hoc to paper Hard-coding QUBO modeling nodes is a necessary and sufficient substitute for autonomous code generation.
    The authors state that 'fully autonomous code generation by the central model for such mathematically delicate tasks proved both token-intensive and error-prone' (§4), justifying the hard-coded approach, but this limits the generality of the 'agentic' claim.
invented entities (1)
  • Quantum-iteration-empowered LLM reasoning no independent evidence
    purpose: A proposed new paradigm where quantum computing iteration feedback enhances LLM reasoning capability.
    This is the paper's key conceptual contribution, but it is supported by a single uncontrolled instance (§3.3). The LLM reasoning log shows heuristic enumeration, not a mechanism tied to quantum-specific feedback. No falsifiable prediction is made beyond the observed instance.

pith-pipeline@v1.1.0-glm · 28413 in / 3235 out tokens · 197053 ms · 2026-07-04T19:30:22.775644+00:00 · methodology

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. Comprehensive pKa Data Augmentation from Limited Real Data through an Engineered Models-Quantum Framework

    physics.chem-ph 2026-06 unverdicted novelty 5.0

    A quantum-assisted molecular generation method targets extreme pKa values to augment limited experimental data, showing better tail sampling on coherent Ising machines than classical VAE-RNN approaches.

Reference graph

Works this paper leans on

66 extracted references · 66 canonical work pages · cited by 1 Pith paper · 1 internal anchor

  1. [1]

    Position One-Hot Constraint (Hpos)This term enforces that exactly one amino acid is selected at each sequence position. The penalty function is defined as: Hpos =λ pos SX s=1 1− X a∈A xs,a !2 where S is the total number of sequence positions, A denotes the set of candidate amino acids, and xs,a ∈ {0,1} is the binary variable indicating whether amino acid ...

    work page

  2. [2]

    Total Mass Matching Constraint ( Hmass)This term minimizes the deviation between the total mass of the selected sequence and the target mass Mtarget. The penalty function is defined as: Hmass =λ mass SX s=1 X a∈A maxs,a −M target !2 where ma is the monoisotopic mass of amino acid a, and λmass is the penalty weight for the mass matching constraint

    work page

  3. [3]

    QUBO Standard Form and Matrix ConstructionThe combined objective function is reformulated into the standard QUBO form: min x∈{0,1}N xT Qx where N=S× |A| is the total number of binary variables. The entries of the QUBO matrix Q are derived by combining the linear and quadratic coefficients from both penalty terms: •Linear terms (diagonal elementsQ ii): For...

    work page

  4. [4]

    The decision variables are defined as: •x i,j,h ∈ {0,1} : Equals 1 if operation oj,h is processed on machinei, 0 otherwise

    Mixed-Integer Programming (MIP) ModelWe first introduce the classical MIP model for FJSP, which serves as the basis for QUBO formulation. The decision variables are defined as: •x i,j,h ∈ {0,1} : Equals 1 if operation oj,h is processed on machinei, 0 otherwise. •y i,j,h,j′,h′ ∈ {0,1} : Equals 1 if operation oj,h is pro- cessed beforeo j′,h′ on machinei, 0...

    work page

  5. [5]

    Quadratic Unconstrained Binary Optimization (QUBO) ModelTo adapt the FJSP for quantum computing via a coherent Ising machine (CIM), we transform the MIP model into a QUBO formulation by discretizing the timeline and encoding all constraints as penalty terms in the objective Hamiltonian. The core binary variable for QUBO modeling is defined as: ki,t,oj,h =...

    work page

  6. [6]

    H2 = X j∈J hj−1X h=1 X t∈T X i∈Mj,h ki,t,oj,h · X t′<t+pi,j,h X i′∈Mj,h+1 ki′,t′,oj,h+1 (17)

    Operation Sequence Constraint (H2): Guarantees non- overlapping sequential operations within the same job, i.e., an operation can only start after the previous operation of the same job is completed. H2 = X j∈J hj−1X h=1 X t∈T X i∈Mj,h ki,t,oj,h · X t′<t+pi,j,h X i′∈Mj,h+1 ki′,t′,oj,h+1 (17)

    work page

  7. [7]

    Machine Conflict Constraint ( H3): Prevents multiple operations from being processed on the same machine at overlapping time intervals. H3 = X (j,h,j′,h′,t,t′)∈G∪H X i∈Mj,h∩Mj′,h′ ki,t,oj,hki,t′,oj′,h′ (18) where the sets are defined as: G={(j, h, j ′, h′, t, t′)|j, j ′ ∈J, j̸=j ′, t, t ′ ∈T, 0⩽t−t ′ ⩽p i,j,h}, H={(j, h, j ′, h′, t, t′)|j, j ′ ∈J, j̸=j ′,...

    work page

  8. [8]

    Makespan Minimization Objective (H4): Penalizes late completion of operations to minimize the overall makespan. H4 = X j∈J X i∈Mj,h X t∈T ki,t,oj,h · t+p i,j,h −P oj,h (19) where Poj,h is the minimum predecessor time of operation oj,h, defined as the sum of the minimum processing times of all preceding operations ofo j,h: Poj,h = X loj,h′ <loj,h min i′∈Mj...

    work page

  9. [9]

    The standard QUBO form is given in Eq

    Conversion from QUBO to Ising ModelThe CIM solver requires the problem to be formulated as an Ising model with spin variables σi ∈ {−1,+1} . The standard QUBO form is given in Eq. (1): min x∈{0,1}n xT Qx+c T x where Q∈R n×n is the quadratic coefficient matrix, and c∈R n is the linear coefficient vector. We perform the variable substitution to map binary Q...

    work page

  10. [10]

    Variable Pruning ApproachTo reduce the number of qubits required for quantum computation, we implement two variable pruning strategies to eliminate invalid binary variablesk i,t,oj,h:

    work page

  11. [11]

    Machine Eligibility Pruning: For any operation oj,h and machine i that is not in the eligible machine set Mj,h, the variable is fixed to 0 for all time steps: ki,t,oj,h = 0,∀t∈T, i /∈M j,h, j∈J(23)

    work page

  12. [12]

    Time Window Pruning: An operation can only start pro- cessing after all its preceding operations are completed, and must finish before the minimum start time of its subsequent operations. We fix the variable to 0 for all time steps outside the valid time window: ki,t,oj,h = 0,∀t∈Twitht < P oj,h ort > S oj,h (24) where Poj,h is the minimum predecessor time...

    work page

  13. [13]

    -We first test the feasibility of a makespan of 12

    to explore the optimal solution. -We first test the feasibility of a makespan of 12. We calculate the total processing time of all operations by selecting the shortest processing time for each job: 8 for Job 1, 9 for Job 2, and 7 for Job 3, giving a total processing time of 24 and an average machine load of 8 across the three machines. The theoretical low...

    work page

  14. [14]

    Further exploration is still required for a schedule with a makespan of 10

    Analysis suggests the start time of Job 1’s operations must be⩽ 2 and end time ⩽ 5, while all operations of Job 2 finish by 10, with no obvious conflicts. Further exploration is still required for a schedule with a makespan of 10. -We first analyze the processing times ofO1,1 and O2,1 on M1 and find they overlap, meaning they cannot both be assigned to M1...

    work page

  15. [15]

    X. B. Zhu, Z. Y. Lu, and J. W. Pan, Quantum computing: A solution for computing power improvement in the post-Moore era (in Chinese), Study Times, 2022. https://www.cas.cn/zjs/2 02203/t20220302 4826718.shtml

    work page 2022

  16. [16]

    Outline of the 15th Five-Year Plan (2026–2030) for National Economic and Social Development of the People’s Republic of China (in Chinese), Column 8, Chap. 8, Sec. 2, Xinhua News Agency, Mar. 2026. https://www.gov.cn/yaowen/liebiao/2026 03/content 7062633.htm

    work page 2026

  17. [17]

    Y. D. Zhang, S. H. Wang, and G. L. Ji, A comprehensive survey on particle swarm optimization algorithm and its applications, Math. Probl. Eng., vol. 2015, p. 931256, 2015

    work page 2015

  18. [18]

    Hauke, H

    P. Hauke, H. G. Katzgraber, W. Lechner, H. Nishimori, and W. D. Oliver, Perspectives of quantum annealing: methods and implementations,Rep. Prog. Phys., vol. 83, no. 5, p. 054401, 2020

    work page 2020

  19. [19]

    Bauer, S

    B. Bauer, S. Bravyi, M. Motta, and G. K. L. Chan, Quantum algorithms for quantum chemistry and quantum materials science,Chem. Rev., vol. 120, no. 22, pp. 12685–12717, 2020

    work page 2020

  20. [20]

    P. A. Deymier, K. Runge, M. A. Hasan, J. A. Levine, and P. Cu- tillas, Setting the stage for materials simulation using acoustic metamaterials digital quantum analogue computing platforms, Mod. Simul. Mater. Sci. Eng., vol. 30, no. 8, p. 084003, 2022

    work page 2022

  21. [21]

    Camino, J

    B. Camino, J. Buckeridge, P. A. Warburton, V. Kendon, and S. M. Woodley, Quantum computing and materials science: A practical guide to applying quantum annealing to the configu- rational analysis of materials,J. Appl. Phys., vol. 133, no. 22, p. 221102, 2023

    work page 2023

  22. [22]

    Tudorovskaya and D

    M. Tudorovskaya and D. M. Ramo, Quantum computing sim- ulation of a phase change in a cavity quantum electrodynamics hamiltonian, inProc. IEEE Int. Conf. Quantum Comput. Eng. (QCE), vol. 2, 2024, pp. 561–562

    work page 2024

  23. [23]

    Fauseweh, Quantum many-body simulations on digital quantum computers: State-of-the-art and future challenges, Nat

    B. Fauseweh, Quantum many-body simulations on digital quantum computers: State-of-the-art and future challenges, Nat. Commun., vol. 15, no. 1, p. 2123, 2024

    work page 2024

  24. [24]

    X. H. Lv, S. Rani, S. Manimurugan, A. Slowik, and Y. H. Feng, Quantum-inspired sensitive data measurement and secure transmission in 5G-enabled healthcare systems,Tsinghua Sci. Technol., vol. 30, no. 1, pp. 456–478, 2025

    work page 2025

  25. [25]

    C. Wang, J. J. Yu, Z. Pei, Q. D. Wang, and C. L. Hong, A first successful factorization of RSA-2048 integer by D-wave quantum computer,Tsinghua Sci. Technol., vol. 30, no. 3, pp. 1270–1282, 2025

    work page 2048

  26. [26]

    Jiang, C.-Q

    Y.-Y. Jiang, C.-Q. Deng, H. Fan, B.-Y. Li, L.-Y. Sun, X.-S. Tan, W.-T. Wang, G.-M. Xue, F. Yan, H.-F. Yu, Y.-S. Zhang, Y.-R. Zhang, and C.-L. Zou, Advancements in superconducting quantum computing,Nat. Sci. Rev., vol. 12, no. 8, p. nwaf246, 2025

    work page 2025

  27. [27]

    Nguyen, T

    N. Nguyen, T. W. Watts, B. Link, K. S. Williams, Y. R. Sanders, S. J. Elman, M. Kieferova, M. J. Bremner, K. J. Morrell, J. Elenewski, E. B. Isaacs, S. D. Johnson, L. Mathieson, K. M. Obenland, M. Otten, R. Sundareswara, and A. Holmes, Quantum computing for corrosion simulation: workflow and resource analysis,npj Quantum Inf., vol. 12, no. 1, p. 27, 2025

    work page 2025

  28. [28]

    Z. Pei, C. L. Hong, F. Xia, and C. Wang, An innovative algo- rithm for attacking symmetric ciphers using D-wave quantum annealing,Tsinghua Sci. Technol., vol. 30, no. 5, pp. 2184– 2194, 2025

    work page 2025

  29. [29]

    Larocca, S

    M. Larocca, S. Thanasilp, S. Wang, K. Sharma, J. Beaumont, P. J. Coles, L. Cincio, J. R. McClean, Z. Holmes, and M. Cerezo, Barren plateaus in variational quantum computing, Nat. Rev. Phys., vol. 7, no. 4, pp. 174–189, 2025

    work page 2025

  30. [30]

    Zhao, H.-S

    X.-H. Zhao, H.-S. Zhong, F. Pan, Z.-H. Chen, R. Fu, Z.-L. Su, X.-T. Xie, C.-X. Zhao, P. Zhang, W.-L. Ouyang, C.-Y. Lu, J.-W. Pan, and M.-C. Chen, Leapfrogging Sycamore: harnessing 1432 GPUs for 7x faster quantum random circuit sampling, Nat. Sci. Rev., vol. 12, no. 3, p. nwae317, 2025

    work page 2025

  31. [31]

    Mehmood, N

    K. Mehmood, N. I. Chaudhary, Z. A. Khan, K. M. Cheema, and M. A. Z. Raja, Design of quantum computing-based avain navigation optimization algorithm for parameter estimation of input nonlinear output error model with key term separation, Mod. Phys. Lett. A, vol. 40, no. 11N12, p. 2550019, 2025

    work page 2025

  32. [32]

    Kim, I.-S

    S. Kim, I.-S. Suh, T. S. Humble, T. Beck, E. Lee, and T.-F. Luo, Harnessing quantum computing for energy materials: Opportunities and challenges,ACS Energy Lett., vol. 11, no. 2, pp. 1182–1192, 2026

    work page 2026

  33. [33]

    Monroe, W

    C. Monroe, W. C. Campbell, L.-M. Duan, Z.-X. Gong, A. Gorshkov, P. W. Hess, R. Islam, K. Kim, N. M. Linke, and G. Pagano, Programmable quantum simulations of spin systems with trapped ions,Rev. Mod. Phys., vol. 93, no. 2, p. 025001, 2021

    work page 2021

  34. [34]

    Liu, K.-Z

    B.-W. Liu, K.-Z. Wang, D.-M. Xiao, and Z. Yu, On connec- tion among quantum-inspired algorithms of the Ising model, Commun. Math. Sci., vol. 21, no. 7, pp. 2013–2028, 2023

    work page 2013

  35. [35]

    Palacios, A

    A. Palacios, A. Garcia-Saez, B. Juli´a-D´ıaz, and M. P. Estarellas, Scalable 2-local architecture for quantum annealing of Ising models with arbitrary dimensions,Phys. Rev. Appl., vol. 23, no. 5, p. 054070, 2025

    work page 2025

  36. [36]

    Sun, G.-B

    B. Sun, G.-B. Cao, X.-D. Hu, and D.-B. Zhang, Revealing quantum operator scrambling via measuring Holevo informa- tion on digital quantum simulators,Phys. Rev. A, vol. 112, no. 1, p. 012403, 2025

    work page 2025

  37. [37]

    Wang, M.-H

    W.-L. Wang, M.-H. Tang, K. Wang, M. A. Basit, Y. Tian, and M. S. Haque, A resource efficient Ising model-based quantum sudoku solver,Softw. Pract. Exper., 2026, doi: 10.1002/spe.70063

    work page doi:10.1002/spe.70063 2026

  38. [38]

    K. Fu, J. Liu, M. Chen, and H. Zhang, Solving flexible job-shop 20 scheduling problems based on quantum computing,Entropy, vol. 27, p. 189, 2025

    work page 2025

  39. [39]

    C. Qiu, P. Zhu, and L. Wei, A beam search framework for quantum circuit mapping,Entropy, vol. 27, p. 232, 2025

    work page 2025

  40. [40]

    F.-A. Wang, S. Chen, Y. Xuan, J. Liu, Q. Gao, H. Zhu, J. Hou, L. Yuan, J. Cheng, C. Yi, H. Wei, Y. Ma, T. Xu, K. Wen, and Y. Li, Quantum-boosted high-fidelity deep learning, 2025, arXiv:2508.11190 [cs.LG]

    work page arXiv 2025

  41. [41]

    H. Wei, C. Ai, P. Guo, B. Jia, L. Yuan, H. Song, S. Chen, C. Cao, J. Wu, C. Ju, Y. Ma, J. Fan, M. Hu, C. Wang, and K. Wen, A versatile coherent Ising computing platform,Light Sci. Appl., vol. 15, pp. 1–26, 2026

    work page 2026

  42. [42]

    Biesner, T

    D. Biesner, T. Gerlach, R. Sifa, C. Bauckhage, and B. Kliem, Solving subset sum problems using quantum inspired opti- mization algorithms with applications in auditing and financial data analysis, inProc. IEEE 21st Int. Conf. Mach. Learn. Appl. (ICMLA), 2022, pp. 903–908

    work page 2022

  43. [43]

    J. Y. Zha, J. Q. Su, T. E. Li, C. Y. Cao, Y. Ma, H. Wei, Z. G. Huang, L. Qian, K. Wen, and J. Zhang, Encoding molecular docking for quantum computers,J. Chem. Theory Comput., vol. 19, no. 24, pp. 9018–9024, 2023

    work page 2023

  44. [44]

    Mahroo and A

    R. Mahroo and A. Kargarian, Learning infused quantum- classical distributed optimization technique for power genera- tion scheduling,IEEE Trans. Quantum Eng., vol. 4, p. 3102314, 2023

    work page 2023

  45. [45]

    Y. Y. Zhang, Y. H. Li, H. J. Li, F. X. Jiao, Z. M. Zhang, and X. B. Zhang, Accelerating quadratic unconstrained binary optimization solution for logistics distribution routing problem with graph attention network, inProc. 8th CAA Int. Conf. Veh. Control Intell. (CVCI), 2024

    work page 2024

  46. [46]

    Krikidis, MIMO with 1-b pre/postcoding resolution: A quantum annealing approach,IEEE Trans

    I. Krikidis, MIMO with 1-b pre/postcoding resolution: A quantum annealing approach,IEEE Trans. Quantum Eng., vol. 5, p. 2100409, 2024

    work page 2024

  47. [47]

    Volpe, G

    D. Volpe, G. A. Cirillo, M. Zamboni, M. Graziano, and G. Turvani, Improving the exploitability of simulated adiabatic bifurcation through a flexible and open-source digital archi- tecture,ACM Trans. Quantum Comput., vol. 6, no. 1, p. 7, 2025

    work page 2025

  48. [48]

    E. Ak, D. V. Huynh, and T. Q. Duong, Quantum-enhanced optimization for LNG ship routing: Integrating digital twins with QAOA, inProc. IEEE Int. Conf. Commun. Workshops (ICC Workshops), 2025, pp. 769–774

    work page 2025

  49. [49]

    Triuzzi, R

    E. Triuzzi, R. Mengoni, F. Micucci, D. Bonanni, D. Ottaviani, A. R. Beccari, and G. Palermo, Molecular docking via weighted subgraph isomorphism on quantum annealers,Quantum Sci. Technol., vol. 10, no. 4, p. 045049, 2025

    work page 2025

  50. [50]

    Kaiwu SDK Learning Re- sources, https://kaiwu-sdk-docs.qboson.com/zh/latest/, 2026

    QBoson Quantum Technology Co. Kaiwu SDK Learning Re- sources, https://kaiwu-sdk-docs.qboson.com/zh/latest/, 2026

    work page 2026

  51. [51]

    X. L. Zhang, D. Z. R. Wang, L. X. Dou, Q. F. Zhu, and W. X. Che, A survey of table reasoning with large language models, Front. Comput. Sci., vol. 19, no. 9, p. 199348, 2024

    work page 2024

  52. [52]

    K. E. Tezel and G. Kardas ¸, Debugging in the domain-specific modeling languages for multi-agent systems,J. Comput. Lang., vol. 83, p. 101325, 2025

    work page 2025

  53. [53]

    X. Hu, S. Y. Chen, L. T. Chen, H. J. Wang, X. Zhang, and Z. Zhou, Automating structure-activity analysis for electrochem- ical nitrogen reduction catalyst design through multi-agent collaborations,Nat. Sci. Rev., vol. 12, no. 11, p. nwaf372, 2025

    work page 2025

  54. [54]

    Das, S.-C

    A. Das, S.-C. Chen, M.-L. Shyu, and S. Saad, Enabling synergistic knowledge sharing and reasoning in large language models with collaborative multi-agents, inProc. IEEE 9th Int. Conf. Collab. Internet Comput. (CIC), 2023, pp. 92–98

    work page 2023

  55. [55]

    K. N. Li, W. N. Lu, C. Y. Tang, and C. Z. Yuan, Research on three-dimensional global dynamic path planning algorithm for mobile intelligent agent,Tsinghua Sci. Technol., vol. 31, no. 3, pp. 1678–1690, 2026

    work page 2026

  56. [56]

    Z. J. Zhu, Z. Q. Zhang, K. Chen, D. B. Tang, and Q. X. Cai, Resource optimisation method for multi-agent manufactur- ing system based on cloud-edge collaboration architecture, Tsinghua Sci. Technol., vol. 31, no. 2, pp. 1198–1215, 2026

    work page 2026

  57. [57]

    J. J. Luo, J. J. Xia, B. X. Pan, Y. G. Ham, X. F. Li, S. G. Wei, X. Xue, Y. Q. Wang, Y. Yaqiang, B. Mu, Y. Hong, H. Li, X. H. Zhong, K. Dai, L. Bai, F. Ling, N. Boers, C. Bretherton, P. Gentine, Z. J. Guo, X. M. Huang, D. Kang, H. J. Kim, J. H. Kim, L. L. Lei, L. Li, F. Meng, S. H. Oh, B. Qin, Z. X. Shen, Q. M. Sun, Y. H. Tang, X. Tong, B. Wan, L. N. Wang...

    work page 2026

  58. [58]

    L. Jern, V. Uotila, C. Yu, and B. Zhao, Agent-Q: Fine-tuning large language models for quantum circuit generation and optimization, inProc. IEEE Int. Conf. Quantum Comput. Eng. (QCE), Albuquerque, NM, USA, 2025, pp. 1621–1632

    work page 2025

  59. [59]

    Cao, Z.-Y

    Q.-H. Cao, Z.-Y. Hou, Y.-Y. Li, X. Liu, Z.-Y. Song, L.- Q. Zhang, S. Zhang, and K. Zhao, Scalable quantum state preparation via large-language-model-driven discovery, 2025, arXiv:2505.06347 [quant-ph]

    work page arXiv 2025

  60. [60]

    M. Chen, J. Cheng, P. Li, H. Wang, T. Chen, and J. Liu, Sym- bolic analysis of Grover search algorithm via chain-of-thought reasoning and quantum-native tokenization,npj Quantum Inf., vol. 12, p. 48, 2026

    work page 2026

  61. [61]

    N. N. Nguyen, T. T. Vu, J. Le, H. K. Dam, D. H. Duong, and D. T. Hoang, BRIDG-Q: Barren-plateau-resilient initialisa- tion with data-aware LLM-generated quantum circuits, 2026, arXiv:2603.23979 [cs.ET]

    work page arXiv 2026

  62. [62]

    R. Wang. The open-source QUBO-V AE project available on GitHub, https://github.com/wrlengo/QUBO-V AE, 2026

    work page 2026

  63. [63]

    The open-source LangGraph and Langchain project available on GitHub, https://github.com/langchain-ai, continuously updated

    langchain-ai. The open-source LangGraph and Langchain project available on GitHub, https://github.com/langchain-ai, continuously updated

    work page

  64. [64]

    S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, and others, ReAct: synergizing reasoning and acting in language models, inProc. 11th Int. Conf. Learn. Represent. (ICLR), 2023. Practical Quantum CIM EmpowermentviaAll-Domestic-Core Agentic Large Model 21

    work page 2023

  65. [65]

    C. Q. Sun, Y. Q. Shu, and W. J. Mao, Peptidomic analysis re- veals the critical roles of endogenous peptides in the malignant phenotype of lung adenocarcinoma, (in Chinese),J. Nanjing Med. Univ. (Natural Sciences), vol. 42, no. 3, pp. 317–324, 2022

    work page 2022

  66. [66]

    LLM Reasoning Is Latent, Not the Chain of Thought

    W. Wang, LLM reasoning is latent, not the chain of thought, 2026, arXiv:2604.15726 [cs.AI]. Author biography Wang Rui(ORCiD: 0000-0002-2362-3214) is an Assistant Research Fellow (Postdoc- toral Researcher) at the Department of Chem- ical Engineering, Tsinghua University. He earned his Ph.D. in Chemistry from Tsinghua in 2024 and has published multiple fir...

    work page internal anchor Pith review Pith/arXiv arXiv 2026