acceptodds
Under review as a conference paper at ICLR 2027

Diffusion Language Models as Open Quantum Systems

Abstract

Diffusion language models generate text by reversing a classical diffusion process over token sequences. We show that their continuous-time forward and reverse dynamics can be embedded in the framework of open quantum systems, which describes quantum systems interacting with their environment. This connection provides a basis for extending diffusion sampling to superpositions of sentence states and for modifying the underlying generation dynamics. We construct a -leaping unraveling scheme that simulates stochastic trajectories of these states while updating multiple token positions at each step. This enables generation conditioned on superposed prompts or thoughts, which we call Superposition of Prompts (SoP) and Superposition of Thoughts (SoT). The same formulation yields Quantum-Inspired Remasking (QIR): perturbing the masked-diffusion dynamics produces a confidence-based rule for revising previously generated tokens. Experiments with pretrained diffusion language models show that superposition-based generation can retain information from conditioning components not selected at the final measurement, while QIR improves LLM-assessed text quality when applied at an appropriate stage of generation.

Primary area:
generative models
Keywords:
Diffusion Language Models, Discrete Diffusion Models, Remasking, Open Quantum Systems

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_S1QCOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The authors show that the dynamics of diffusion models can be embedded into quantum Lindbladian evolution. This gives a natural generalization of diffusion models, namely, quantum open system dynamics. The authors then construct multiple quantum-inspired classical learning algorithms based on Lindbladian evolution. Finally, the authors evaluate the behavior of their quantum-inspired algorithms; for their SoP and SoT algorithms, they show qualitative differences in sentence generation. For their quantum-inspired remasking technique, the authors show that in various settings it can either improve or worsen the performance of standard pretrained diffusion models with different choices of hyperparameters.

Critical strengths

In my opinion, quantum-inspired classical learning algorithms are an underexplored area in machine learning algorithm design. The authors develop this in a relevant setting---diffusion modeling---and empirically show how these quantum-inspired methods can yield qualitatively different behavior than traditional methods.

Critical weaknesses

While the embedding of diffusion into Lindbladian dynamics is technically new to the best of my knowledge, it is not very surprising given Lindbladian dynamics can be used to describe, e.g., the general stochastic evolution of probability distributions. It is nice to see the exact correspondence written out, however, which mitigates this weakness.

To me, the biggest weakness of the work is the inefficiency of the introduced algorithms. Generally there seems no way to circumvent this as, in the end, the algorithms are based on simulating quantum Lindbladian evolution. The overhead the authors' generalized algorithms introduced might be fine in certain settings, but in my opinion the authors do not do enough to justify in which settings this overhead might be worth it; in fact, they barely discuss the complexity of their algorithms outside of a brief mention in the conclusion. One potential experiment the authors can do to address this weakness would be to perform a comparison between their approach and traditional approaches when the time complexity of each algorithm is fixed. This controls for the overhead induced by considering general Lindbladian evolution and gives more evidence that their algorithms are doing something beyond what existing methods can already achieve.

Unfortunately, I think the main weakness---the inefficiency of the authors' introduced algorithms, limiting the scope---does not do enough to overcome the strengths, and I therefore recommend weak rejection.

Questions for authors

If one fixes the computational complexity of the quantum-inspired approach compared with traditional approaches, do traditional approaches outperform the quantum-inspired approach?

Is there an interesting "middle ground" where the Lindbladian dynamics are generated by efficiently-simulable quantum processes, e.g., stoquastic dynamics or the like?

3 comments

We thank the reviewer for the thoughtful comments on computational cost and efficiently simulable dynamics. These questions helped us sharpen the discussion of the framework’s practical scope. The revised manuscript includes a comparison under approximately matched computational cost and Hamiltonian Prompt Mixing as an example of efficiently simulable dynamics in which coherence affects generation. We discuss these additions below.

The updated supplementary material includes a PDF of the revised manuscript with revisions highlighted in blue.

Critical Weaknesses

While the embedding of diffusion into Lindbladian dynamics is technically new to the best of my knowledge, it is not very surprising given Lindbladian dynamics can be used to describe, e.g., the general stochastic evolution of probability distributions. It is nice to see the exact correspondence written out, however, which mitigates this weakness. To me, the biggest weakness of the work is the inefficiency of the introduced algorithms. Generally there seems no way to circumvent this as, in the end, the algorithms are based on simulating quantum Lindbladian evolution. The overhead the authors' generalized algorithms introduced might be fine in certain settings, but in my opinion the authors do not do enough to justify in which settings this overhead might be worth it; in fact, they barely discuss the complexity of their algorithms outside of a brief mention in the conclusion. One potential experiment the authors can do to address this weakness would be to perform a comparison between their approach and traditional approaches when the time complexity of each algorithm is fixed. This controls for the overhead induced by considering general Lindbladian evolution and gives more evidence that their algorithms are doing something beyond what existing methods can already achieve. Unfortunately, I think the main weakness—the inefficiency of the authors' introduced algorithms, limiting the scope— does not do enough to overcome the strengths, and I therefore recommend weak rejection.

We thank the reviewer for raising this concern. Our methods exploit restricted state representations rather than simulating arbitrary Lindbladian dynamics. The revised manuscript adds HPM as a concrete example: coherences affect generation while each trajectory retains only a shared body and two prompt amplitudes (Appendix J). We also compare QIR with existing remasking methods under approximately matched generation-time budgets (Appendix H). QIR performs comparably to the WINO-style method, and both outperform ReMDM and no remasking. Together, these results demonstrate useful extensions within efficiently simulable dynamics and competitive performance after accounting for computational cost.

Questions for Authors

Question 1

If one fixes the computational complexity of the quantum-inspired approach compared with traditional approaches, do traditional approaches outperform the quantum-inspired approach?

To address this question, we have added a comparison under approximately matched computational cost (Appendix H). We match generation time by allocating more sampling steps to the less expensive ReMDM and no-remasking baselines. QIR and the WINO-style method achieve comparable Qwen preference scores, and both outperform ReMDM and no remasking. Thus, QIR remains competitive when computational cost is approximately matched.

Question 2

Is there an interesting "middle ground" where the Lindbladian dynamics are generated by efficiently-simulable quantum processes, e.g., stoquastic dynamics or the like?

Yes. We have added an example of such a middle ground to the revised manuscript: Hamiltonian Prompt Mixing (HPM), in which coherences affect generation while the dynamics remain efficiently simulable on classical hardware. Its Hamiltonian mixes two prompt components while preserving their shared generated body, and the jump updates preserve this two-component representation. Each trajectory therefore requires only one body and two amplitudes, with two prompt-conditioned model evaluations per step. Figure 4 reports the resulting changes in generated text, and Appendix J gives the construction and implementation.

We'd like to thank the authors for their comments and for answering my questions.

The authors' rebuttal to what I perceived as the main weakness of the work then runs into the weakness perceived by Reviewer nMAd; namely, if you are considering this restricted class of Lindbladian operators where this is efficient to classically implement, there is nothing really "quantum-inspired" about it at all (there is not even interference in these special cases). I agree with Reviewer nMAd that the statistics in this setting could be reproduced by other classical samplers. For this reason, I will keep my score.

Thank you for the follow-up. We would like to clarify two distinct points.

First, classical simulability does not preclude a method from being quantum-inspired. Our claim is that the Lindblad framework provides a way to design classical generation algorithms, not that their outputs are impossible to reproduce classically. The efficiently simulable “middle ground” raised in your earlier question is precisely the setting we aim to explore.

Second, the absence of coherence effects applies to our original zero-Hamiltonian construction, but not to the newly added Hamiltonian Prompt Mixing (HPM). In HPM, the Hamiltonian contribution satisfies dρ11dτ=−g(ρ12+ρ21)\frac{d\rho_{11}}{d\tau}=-g(\rho_{12}+\rho_{21}), so off-diagonal coherences explicitly affect populations. For example, the states (∣F,x⟩±∣S,x⟩)/2(|F,\mathbf{x}\rangle\pm|S,\mathbf{x}\rangle)/\sqrt{2} have identical component weights but yield different component-measurement probabilities after the rotation. A description retaining only those weights therefore cannot reproduce the dynamics. HPM remains efficiently simulable because each trajectory retains only a shared body and two amplitudes, including their relative phase—not because coherences are absent.

Appendix J derives this mechanism, and Figure 4 shows how adding the Hamiltonian changes the generated-text statistics. We hope this distinction clarifies why classical simulability does not negate either the quantum-inspired construction or the role of coherence in HPM, and ask that you reconsider your assessment in light of this addition.

Reviewer_1nFNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper embeds continuous-time discrete diffusion language models into the Lindblad formalism. With a zero Hamiltonian and jump operators L = sqrt(Qtok(m, n)) |m><n| acting on single token positions, the Lindblad equation restricted to the diagonal is exactly the classical master equation of the diffusion process, and time-reversed jump operators reproduce the reverse diffusion generator. The authors then use this as a design space: a τ-leaping unraveling sampler that updates several positions per step, superpositions of prompts (SoP) and of reasoning traces (SoT), and a perturbed masked-diffusion dynamics that yields Quantum-Inspired Remasking (QIR), where tokens the model is less confident about are remasked more often.

Critical strengths

1.The embedding itself is correct and cleanly presented. Restricting a Lindblad equation with diagonal jump operators to the diagonal sector gives the classical master equation, and the paper works out the reverse generator explicitly for the uniform and masked cases. As a piece of formalism it is tidy and the notation is consistent. 2.The τ-leaping unraveling is a sensible adaptation of an established simulation technique, and it addresses a real inefficiency of one-token-per-step dynamics. Stating a first-order consistency result rather than claiming exactness is the right level of claim. 3.QIR is cheap and plausible: it reuses the model's own reconstruction score as the remasking rate and needs no retraining. The empirical caution that early remasking lowers perplexity while increasing repetition, and that judge preference peaks at a middle schedule, is a useful warning for a field that over-reads perplexity. 4.The paper is careful in places: it calls the methods quantum-inspired rather than quantum, reports the adverse regime, and does not claim a speedup.

Critical weaknesses

1.The state never leaves the diagonal in any observable sense. With H = 0 and jump operators that are diagonal in the sentence basis (rank-one in |m><n| with position-wise support), the off-diagonal coherences never feed back into the diagonal: a coherence between two sentences maps to a coherence between the two transitioned sentences, and can only reach the diagonal when the two sentences coincide, which the transition structure forbids. The Born-rule distribution over sentences is therefore a classical mixture of the per-branch distributions with time-varying weights. SoP and SoT can be reproduced by running the two conditional generations and mixing them with the same weights - a cheap control the paper does not run. Until that control is shown to fail, 'the components influence one another' is an assertion rather than a demonstrated effect. 2.The superposition experiment shows what a mixture predicts. Outputs lying between the two single-prompt distributions, with higher similarity to the unselected prompt than a single-prompt sample, is exactly what sampling from a weighted mixture gives. It is not evidence of interference or of a non-classical effect. The claim in the abstract - retaining information from conditioning components not selected at measurement - is true of any mixture and therefore does not distinguish the framework. 3.The embedding theorem is standard. Every classical continuous-time Markov chain is a Lindblad equation with diagonal jump operators; this is textbook. The paper presents it as 'we show that diffusion language models can be embedded', which oversells the content. The non-trivial parts are the reverse-time construction and the sampler, and even there the contribution is careful bookkeeping rather than a new identity. 4.QIR is existing confidence-based remasking with a new derivation. Remasking by reconstruction confidence is already used - LLaDA's confidence-based remasking, WINO's verification threshold, and the remasking sampler of Wang et al. The γ-perturbation derivation is elegant, but the resulting algorithm is in the same family, and the experiments compare only against 'no remasking'. Without WINO, ReMDM or LLaDA-style remasking baselines, the 62% preference cannot be attributed to QIR rather than to remasking in general. 5.The evaluation is thin on both halves. SoP/SoT rest on a single prompt family, a handful of samples, and MiniLM similarities with no task-level metric. QIR uses 256 unconditional sequences per condition, an LLM judge, and no human evaluation; 'middle is best' is selected post hoc over three schedules; γ = 0.02 is fixed with no sensitivity study. The judge is shown same-seed baselines, which is good, but LLM judges are known to reward length and fluency, and length is not controlled.

Questions for authors

1.Can you run the mixture control - sample branch 1 or 2 with the same time-varying weights, generate, and compare the output distribution to SoP? If the two differ, what quantity distinguishes them? 2.Is there any observable in your construction that a classical mixture cannot reproduce? If not, in what sense is the state a superposition rather than a reweighted ensemble? 3.How often do the weights collapse to a single branch without the cβ softening, and what does that imply about the mechanism? 4.How does QIR compare with confidence-based remasking as used in LLaDA, WINO and ReMDM at the same schedule and compute budget? 5.How sensitive are the QIR results to γ, and to the number of sampled positions K = 8?

3 comments

We thank the reviewer for the careful assessment and constructive suggestions. The comments helped us clarify the role of coherence and strengthen the experimental evaluation through comparisons with existing remasking methods and a parameter-sensitivity study. We respond to each point below.

The updated supplementary material includes a PDF of the revised manuscript with revisions highlighted in blue.

Critical Weaknesses

The state never leaves the diagonal in any observable sense. With H = 0 and jump operators that are diagonal in the sentence basis (rank-one in |m><n| with position-wise support), the off-diagonal coherences never feed back into the diagonal: a coherence between two sentences maps to a coherence between the two transitioned sentences, and can only reach the diagonal when the two sentences coincide, which the transition structure forbids. The Born-rule distribution over sentences is therefore a classical mixture of the per-branch distributions with time-varying weights. SoP and SoT can be reproduced by running the two conditional generations and mixing them with the same weights - a cheap control the paper does not run. Until that control is shown to fail, 'the components influence one another' is an assertion rather than a demonstrated effect.

We thank the reviewer for highlighting this limitation. For the original zero-Hamiltonian dynamics at β=1\beta=1, coherences do not affect sentence-measurement statistics, which agree with a classical mixture in the continuous-time limit. The revised manuscript adds Hamiltonian Prompt Mixing (HPM), whose nonzero Hamiltonian couples coherences to sentence probabilities. Figure 4 reports the resulting changes in generated text relative to no rotation. We explain the distinction in our response to Question 2 below and derive the dynamics in Appendix J.

The superposition experiment shows what a mixture predicts. Outputs lying between the two single-prompt distributions, with higher similarity to the unselected prompt than a single- prompt sample, is exactly what sampling from a weighted mixture gives. It is not evidence of interference or of a non- classical effect. The claim in the abstract - retaining information from conditioning components not selected at measurement - is true of any mixture and therefore does not distinguish the framework.

The embedding theorem is standard. Every classical continuous-time Markov chain is a Lindblad equation with diagonal jump operators; this is textbook. The paper presents it as 'we show that diffusion language models can be embedded', which oversells the content. The non-trivial parts are the reverse-time construction and the sampler, and even there the contribution is careful bookkeeping rather than a new identity.

We agree that embedding a classical Markov process into Lindblad dynamics is an established construction. Our contribution is to use this correspondence as a framework for extending pretrained diffusion language models, rather than to claim a new general embedding theorem. We develop a token-wise unraveling sampler and concrete modifications of the inference dynamics: QIR modifies the jump operators, while the newly added HPM introduces a Hamiltonian through which coherences affect generation. These constructions and their empirical evaluation demonstrate how the framework can guide sampler design beyond the original diffusion dynamics.

QIR is existing confidence-based remasking with a new derivation. Remasking by reconstruction confidence is already used - LLaDA's confidence-based remasking, WINO's verification threshold, and the remasking sampler of Wang et al. The γ\gamma-perturbation derivation is elegant, but the resulting algorithm is in the same family, and the experiments compare only against 'no remasking'. Without WINO, ReMDM or LLaDA-style remasking baselines, the 62% preference cannot be attributed to QIR rather than to remasking in general.

QIR shares the use of reconstruction confidence with existing methods, but its distinction is not merely a new derivation of the same rule. It assigns stochastic, diffusion-time-dependent remasking rates, whereas WINO uses a confidence threshold. These rules lead to different temporal patterns of remasking and responses to the onset schedule. We have added a comparison of QIR with existing remasking methods at approximately matched generation time, using 1,000 sequences per condition on SEDD-medium (Appendix H). The baselines include ReMDM-cap, ReMDM-conf, a WINO-style confidence-threshold rule, and no remasking. The remasking methods also target comparable numbers of remasking events. We use method-specific onset schedules because the methods respond differently to remasking timing.

The evaluation is thin on both halves. SoP/SoT rest on a single prompt family, a handful of samples, and MiniLM similarities with no task-level metric. QIR uses 256 unconditional sequences per condition, an LLM judge, and no human evaluation; 'middle is best' is selected post hoc over three schedules; γ\gamma = 0.02 is fixed with no sensitivity study. The judge is shown same-seed baselines, which is good, but LLM judges are known to reward length and fluency, and length is not controlled.

We report results for all three onset schedules, including the degradation with early remasking; “middle performs best” summarizes this comparison rather than selectively reporting a favorable condition. Regarding the concern that “LLM judges are known to reward length and fluency,” all conditions generate 1,024 tokens, and the complete decoded texts are supplied to the judge without truncation. Generation length is therefore controlled. Fluency is a relevant aspect of text quality, but we also report repetition and token entropy to assess aspects that a fluency preference may overlook. Each pair is evaluated in both presentation orders, and our conclusions are explicitly stated in terms of LLM-assessed quality.

Questions for Authors

Question 1

Can you run the mixture control - sample branch 1 or 2 with the same time-varying weights, generate, and compare the output distribution to SoP? If the two differ, what quantity distinguishes them?

Thank you for suggesting this control. For the original zero-Hamiltonian dynamics at β=1\beta=1, SoP agrees in the continuous-time limit with sampling a branch once from the initial weights and generating conditionally on that prompt. Thus, the joint distribution of the selected branch and generated text does not distinguish the two procedures in this limit, despite the evolving trajectory weights. The revised manuscript additionally introduces Hamiltonian Prompt Mixing (HPM), which goes beyond this mixture-equivalent dynamics; we discuss its mechanism and experimental results in our response to Question 2 below.

Question 2

Is there any observable in your construction that a classical mixture cannot reproduce? If not, in what sense is the state a superposition rather than a reweighted ensemble?

We agree that, for the original zero-Hamiltonian dynamics at β=1\beta=1, sentence-measurement statistics reduce to those of a classical mixture in the continuous-time limit. To investigate dynamics in which coherence actively affects generation, we introduce Hamiltonian Prompt Mixing (HPM) in the revised manuscript. Its nonzero Hamiltonian couples off-diagonal coherences to diagonal populations, so a description based only on component weights is insufficient. Figure 4 shows that HPM shifts generated text toward the unselected prompt's references relative to no rotation. The dynamics and the role of coherences are derived in Appendix J.

Question 3

How often do the weights collapse to a single branch without the c$\beta$ softening, and what does that imply about the mechanism?

We have added the untempered case (β=1\beta=1) to the Shannon-entropy plot in Appendix F.4 (Fig. 7). For past/future, all 120 trajectories in each model reach max⁡ip(i)≥0.95\max_i p(i)\geq0.95, and all satisfy this criterion at the last recorded step. The median first-passage step is 14.5 for masked SEDD and 2 for uniform SEDD, out of 256 sampling steps. This indicates that softening prolongs the period during which both prompt components contribute appreciably to the trajectory's transition rates.

Question 4

How does QIR compare with confidence-based remasking as used in LLaDA, WINO and ReMDM at the same schedule and compute budget?

We have added a comparison of QIR with existing remasking methods at approximately matched generation time (Appendix H). The baselines include ReMDM-cap, ReMDM-conf, and a WINO-style confidence-threshold rule. We use method-specific onset schedules because the methods respond differently to remasking timing, so a common onset need not be equally suitable for all methods. QIR's onset dependence is evaluated separately in Appendix G.

Question 5

How sensitive are the QIR results to γ\gamma, and to the number of sampled positions K = 8?

We have added a sensitivity study in Appendix I (“QIR Parameter Sensitivity”) using SEDD-medium with 128 samples per setting. We vary γ∈{0.02,0.2,2,20}\gamma \in \{0.02,0.2,2,20\} at K=8K=8, and K∈{4,8,16}K \in \{4,8,16\} at γ=2\gamma=2, with remasking starting halfway through generation. QIR consistently outperforms no remasking in Qwen preference across all tested settings, indicating that the improvement is robust over the examined parameter ranges. Full results and evaluation settings are provided in Appendix I.

Dear Reviewer,

We are following up on our responses and the revised manuscript. We have added comparisons with existing remasking methods under approximately matched computational cost, a QIR parameter-sensitivity study, and Hamiltonian Prompt Mixing, which explicitly couples coherences to generation. We believe these additions address the main weaknesses identified in your review.

With the discussion deadline approaching, we ask that you review these updates and reconsider your score. If any major concerns remain unresolved, please let us know while there is still time to discuss them.

Thank you for your time.

Reviewer_nMAdOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper embeds continuous-time discrete diffusion language models (uniform and masked) into the Lindblad equation for open quantum systems. The Hamiltonian is set to zero and one jump operator is used per (position, source token, target token), L = sqrt(σ(t) Q_tok(m,n)) |m⟩⟨n| on one tensor factor. On diagonal density operators this reproduces the classical master equation (Thm 3.1). The time-reversed Lindblad operators of Kwon & Kim (2018) reproduce the reverse CTMC (Thm 3.2). The authors then simulate the reverse dynamics by quantum-jump unraveling and introduce a τ-leaping version that updates several positions per step (Thm D.8). With this machinery they propose three samplers:

  • The state is a weighted superposition Σ_i sqrt(p(i)) |prompt_i, x⟩. One transition draw is shared by all components, the component weights are updated multiplicatively (Thm D.10), and a Born-rule "measurement" at the end selects one component. Superposition of Thoughts (SoT).

  • The same sampler, with fixed <think>…</think> blocks as the components.

  • Quantum-Inspired Remasking (QIR) where weak reverse jump [mask]→i with rate γ is added to the forward process. Its time reversal gives a remasking rate proportional to the reciprocal of the model's score for restoring the current token, i.e. inverse-confidence remasking.

Experiments:

  • SoP with SEDD-small on one forest/snow prompt pair. Superposition outputs are, on average, more similar in MiniLM embedding space to the unselected condition's single-prompt outputs (Table 5).
  • One qualitative SoT sample from iLLaDA-8B-Instruct (Table 1).
  • QIR on unconditional 1024-token generation with SEDD-medium, RADD-medium and iLLaDA-8B-Base. Starting QIR halfway through sampling gives a Qwen3-32B pairwise preference of about 62% over no remasking. Starting it early lowers GenPPL substantially but increases repetition and reduces entropy (Table 2).
Critical strengths
  • The diagonal embedding, the reverse-time Lindblad operators, and the reduction to the reverse CTMC are carefully written out in the appendix, with a self-contained review of both the diffusion and the open-system background. For ML readers, the appendix is a readable introduction to quantum-jump unraveling.
  • Factorizing the jump process by token position so that several positions can change per step is the natural analogue of τ-leaping. It makes the trajectory picture usable at realistic sequence lengths.
  • There are three model families, same-seed pairing, order-swapped judging, a repetition metric and token entropy alongside GenPPL. The negative finding (early remasking "wins" on GenPPL and CoLA but loses on diversity, repetition and judge preference) is reported honestly and is a useful reminder about GenPPL.
  • The paper states the β tempering, the K-candidate subsampling, the probability clipping and the score approximation in QIR. The appendix also hedges several claims appropriately (for example, that the checkpoints in F.4 do not isolate diffusion type).
Critical weaknesses
  • The Hamiltonian is zero, all amplitudes are real and non-negative, and every jump operator maps a basis sentence to a basis sentence. Under condition (78), the diagonal of ρ evolves autonomously (Thm D.1). Off-diagonal terms never feed back into measurement statistics in the sentence basis, so there is no interference anywhere in the method.

  • The consequence for SoP can be stated directly. Theorems C.7, D.3 and D.8 say that the ensemble of unraveled trajectories reproduces the Lindblad solution, whose diagonal is the classical mixture Σ_i p(i) · p(x | prompt_i). Measurement statistics of untempered SoP (β = 1) therefore match, to first order in Δt, those of the trivial baseline "sample i ~ p(i), then generate from prompt i alone".

  • The effect in Table 5 therefore comes entirely from the β = 0.01 tempering, which the paper describes as "slightly departing from theoretical principles" (l. 364). With β → 0 the weights stay near 1/2. Each token is then drawn from an equal average of the two conditional reverse rates, and the final "measurement" becomes a near-coin flip that is only weakly coupled to the text. This is a classical rate- or likelihood-averaging ensemble (closely related to composable/mixture guidance in diffusion, e.g. Liu et al., 2022, Compositional Visual Generation with Composable Diffusion Models). The quantum formalism contributes no capability here. That it produces text "between" the two prompts is expected, not a validation of superposition.

  • This undercuts the central claims that superposition "is fundamentally different" from classical alternatives (ll. 47–51) and that the framework "opens a path toward quantum-native language models in which superposition and interference play an active role" (ll. 482–485). No experiment uses a non-zero Hamiltonian, complex phases, or non-basis-preserving jumps, which are the only ingredients that could produce non-classical statistics.

  • Table 5 compares SoP only against single-condition generation. Given point 1, the following baselines are needed: (a) random prompt selection with p = 1/2, which matches untempered SoP; (b) direct averaging of the two conditional per-token reverse rates or logits (β = 0), with no "measurement"; (c) concatenating or merging the two prompts ("forest and snow"); (d) SoP at β = 1 and several intermediate β values, to show how the effect depends on tempering. Without these comparisons, the experiment cannot attribute the effect to anything specific to the proposed framework.

  • middle lowers GenPPL by about 17% and also lowers token entropy (7.57 → 7.52 for SEDD-medium). It is well known that GenPPL can be traded against diversity, for example through temperature, nucleus sampling, or fewer effective steps. The paper needs MAUVE (standard in this literature) or an entropy-matched comparison, such as baseline sampling at reduced temperature tuned to the same entropy.

  • The only "quality" signal favoring middle over start is a single LLM judge, Qwen3-32B, with greedy decoding. No human evaluation, second judge, or calibration is reported. The order-disagreement rate is defined (G.3) but never reported, which matters for interpreting a 62% preference. Finally, all QIR experiments are unconditional. Since iLLaDA-8B is available, a conditional or downstream evaluation (e.g., GSM8K, HumanEval, or instruction following) would greatly strengthen the practical case.

  • Representing a classical CTMC as a Lindblad equation with H = 0 and jump operators sqrt(Q(m,n)) |m⟩⟨n| is textbook (the Pauli master equation as the diagonal sector of a Lindblad equation). It is also the basis of the quantum stochastic walk literature (e.g., Whitfield, Rodríguez-Rosario & Aspuru-Guzik, 2010), which is not cited. Thm 3.2 is a direct application of Kwon & Kim's time reversal (a Petz-type recovery) to diagonal states. Theorems 3.1 and 3.2 are correct but follow quickly from known results. The contribution then rests on the samplers, and points 1–5 apply to those. The related-work paragraph on "Quantum Computation and Diffusion Models" (ll. 108–112) does not discuss any diffusion work. Recent work on "superposition"-style reasoning in LLMs (continuous or soft thoughts, e.g., Coconut; Zhu et al., 2025, Reasoning by Superposition) is relevant to SoT and should be discussed.

  • Use of the term "quantum". The term is used throughout for a method that runs on classical hardware without any quantum resource. The Limitations section acknowledges this. The abstract and introduction should be similarly explicit.

  • "synthesally" (l. 371), "teh" (l. 1056), "denoizing", "Hermite" (should be "Hermitian"), and "reduces to be equivalent" (Thm 3.1). In Thm D.1, the two lines of condition (78) appear identical as typeset. The Thm D.8 proof is very hard to follow; a cleaner per-position argument would help.

  • The joy/sadness prompts are identical: the "sadness" prompt is a copy of the "joy" prompt.

  • The uniform-diffusion checkpoint local_uniform/164931 is undocumented (training data, size, steps). The F.4 comparison cannot be reproduced.

  • Code is not stated to be released, although the appendix refers to "the existing analysis script" and "the accompanying manuscript" (G.1), apparently left over from internal notes.

  • Hirono et al. 2024a and 2024b are the same paper.

Questions for authors
  1. Do you agree that at β = 1, SoP's joint distribution over (selected component, generated text) equals that of sampling the component first and generating conditionally, to first order in Δt? If not, which step of the argument in Weakness 1 fails? Can you report SoP at β = 1 together with the random-prompt baseline on the forest/snow experiment?

  2. How does SoP at β = 0.01 differ, empirically and conceptually, from averaging the two conditional reverse rates at every step (β = 0)? Please include that baseline, and prompt concatenation, in Table 5.

  3. Is there any configuration of your framework (non-zero H̃, complex amplitudes, non-basis-preserving jump operators) that gives measurement statistics no classical mixture sampler can reproduce? If so, why is none of it explored? If not, what does the quantum formulation buy over a Bayesian mixture-of-conditionals sampler?

  4. For superposed initial states, which forward process and which ρ_t define the reverse operators used in Thm D.10? How does this square with Thm D.2's diagonal-ρ_t assumption?

  5. What were the results of the iLLaDA SoT experiments described in F.2–F.3 (forest/city, past/future, calc_reasoning, ghost/cyberpunk)? How was the Table 1 sample chosen?

  6. Is the forward [mask]→i rate γ per token (total ≈ γ|V|) or γ/|V|? How does QIR behave as γ varies for each onset? Is the start degradation explained by the (1 − e^{−σ̄})/e^{−σ̄} factor near t = 1?

5 comments

We thank the reviewer for the careful reading and detailed comments. The questions about the role of coherence helped us clarify the original construction and develop Hamiltonian Prompt Mixing, while the suggested controls enabled a more informative evaluation of SoP. We address the concerns and describe the revisions below.

The updated supplementary material includes a PDF of the revised manuscript with revisions highlighted in blue.

Critical Weaknesses

The Hamiltonian is zero, all amplitudes are real and non-negative, and every jump operator maps a basis sentence to a basis sentence. Under condition (78), the diagonal of ρ\rho evolves autonomously (Thm D.1). Off-diagonal terms never feed back into measurement statistics in the sentence basis, so there is no interference anywhere in the method. The consequence for SoP can be stated directly. Theorems C.7, D.3 and D.8 say that the ensemble of unraveled trajectories reproduces the Lindblad solution, whose diagonal is the classical mixture Σip(i)⋅p(x∣prompti)\Sigma_i p(i) \cdot p(x \mid \mathrm{prompt}_i). Measurement statistics of untempered SoP (β\beta = 1) therefore match, to first order in Δt\Delta t, those of the trivial baseline "sample i ~ p(i), then generate from prompt i alone".

We thank the reviewer for highlighting this limitation. For the original zero-Hamiltonian dynamics at β=1\beta=1, coherences do not affect sentence-measurement statistics, which agree with a classical mixture in the continuous-time limit. The revised manuscript adds Hamiltonian Prompt Mixing (HPM), whose nonzero Hamiltonian couples coherences to sentence probabilities. Figure 4 reports the resulting changes in generated text relative to no rotation. We explain the distinction in our response to Questions 1 and 3 below and derive the dynamics in Appendix J.

The effect in Table 5 therefore comes entirely from the β\beta = 0.01 tempering, which the paper describes as "slightly departing from theoretical principles" (l. 364). With β\beta →\to 0 the weights stay near 1/2. Each token is then drawn from an equal average of the two conditional reverse rates, and the final "measurement" becomes a near-coin flip that is only weakly coupled to the text. This is a classical rate- or likelihood-averaging ensemble (closely related to composable/mixture guidance in diffusion, e.g. Liu et al., 2022, Compositional Visual Generation with Composable Diffusion Models). The quantum formalism contributes no capability here. That it produces text "between" the two prompts is expected, not a validation of superposition.

This undercuts the central claims that superposition "is fundamentally different" from classical alternatives (ll. 47– 51) and that the framework "opens a path toward quantum-native language models in which superposition and interference play an active role" (ll. 482–485). No experiment uses a non-zero Hamiltonian, complex phases, or non-basis-preserving jumps, which are the only ingredients that could produce non-classical statistics.

Table 5 compares SoP only against single-condition generation. Given point 1, the following baselines are needed: (a) random prompt selection with p = 1/2, which matches untempered SoP; (b) direct averaging of the two conditional per-token reverse rates or logits (β\beta = 0), with no "measurement"; (c) concatenating or merging the two prompts ("forest and snow"); (d) SoP at β\beta = 1 and several intermediate β\beta values, to show how the effect depends on tempering. Without these comparisons, the experiment cannot attribute the effect to anything specific to the proposed framework.

middle lowers GenPPL by about 17% and also lowers token entropy (7.57 →\to 7.52 for SEDD-medium). It is well known that GenPPL can be traded against diversity, for example through temperature, nucleus sampling, or fewer effective steps. The paper needs MAUVE (standard in this literature) or an entropy-matched comparison, such as baseline sampling at reduced temperature tuned to the same entropy.

Our claim of improved generation quality is not based on lower GenPPL alone. Indeed, early remasking lowers GenPPL further but increases repetition and reduces judge preference, illustrating the distinction emphasized by the reviewer. MAUVE estimates were unstable across clustering configurations and random seeds in our evaluation, so we do not regard them as reliable evidence in this setting.

The only "quality" signal favoring middle over start is a single LLM judge, Qwen3-32B, with greedy decoding. No human evaluation, second judge, or calibration is reported. The order-disagreement rate is defined (G.3) but never reported, which matters for interpreting a 62% preference. Finally, all QIR experiments are unconditional. Since iLLaDA-8B is available, a conditional or downstream evaluation (e.g., GSM8K, HumanEval, or instruction following) would greatly strengthen the practical case.

Our evaluation assesses text quality from multiple perspectives using GenPPL, token entropy, repetition rates, and the same-seed change in repetition, ΔR\Delta R, alongside Qwen preference. For the middle onset, order-disagreement rates are 69.9% for SEDD-medium, 70.3% for RADD-medium, and 55.9% for iLLaDA. Each pair is evaluated in both presentation orders, and the scores are averaged within each seed. This balanced evaluation cancels a fixed preference for either presentation position: a judge that selects the same position in both orders contributes 0.5, favoring neither method. Thus, order disagreement alone does not invalidate the averaged comparison. Human evaluation or another judge would provide complementary evidence, while our reported assessment combines the judge results with independent text statistics.

Representing a classical CTMC as a Lindblad equation with H = 0 and jump operators sqrt(Q(m,n)) ∣m⟩⟨n∣|m\rangle\langle n| is textbook (the Pauli master equation as the diagonal sector of a Lindblad equation). It is also the basis of the quantum stochastic walk literature (e.g., Whitfield, Rodríguez-Rosario & Aspuru-Guzik, 2010), which is not cited. Thm 3.2 is a direct application of Kwon & Kim's time reversal (a Petz-type recovery) to diagonal states. Theorems 3.1 and 3.2 are correct but follow quickly from known results. The contribution then rests on the samplers, and points 1–5 apply to those. The related-work paragraph on "Quantum Computation and Diffusion Models" (ll. 108– 112) does not discuss any diffusion work. Recent work on "superposition"-style reasoning in LLMs (continuous or soft thoughts, e.g., Coconut; Zhu et al., 2025, Reasoning by Superposition) is relevant to SoT and should be discussed.

We agree that embedding a classical Markov process into Lindblad dynamics is an established construction. Our contribution is to use this correspondence as a framework for extending pretrained diffusion language models, rather than to claim a new general embedding theorem. We develop a token-wise unraveling sampler and concrete modifications of the inference dynamics: QIR modifies the jump operators, while the newly added HPM introduces a Hamiltonian through which coherences affect generation. These constructions and their empirical evaluation demonstrate how the framework can guide sampler design beyond the original diffusion dynamics.

Use of the term "quantum". The term is used throughout for a method that runs on classical hardware without any quantum resource. The Limitations section acknowledges this. The abstract and introduction should be similarly explicit.

Thank you for this comment. We have clarified in the Conclusion that our quantum-inspired methods run on classical hardware and use the Lindblad framework to design generation dynamics. “Quantum-inspired” refers to this theoretical framework, not to a claim of quantum computational advantage. Extending these constructions to quantum hardware remains a direction for future work.

"synthesally" (l. 371), "teh" (l. 1056), "denoizing", "Hermite" (should be "Hermitian"), and "reduces to be equivalent" (Thm 3.1). In Thm D.1, the two lines of condition (78) appear identical as typeset. The Thm D.8 proof is very hard to follow; a cleaner per-position argument would help.

Thank you for pointing out these typographical and wording errors. We have corrected them in the revised manuscript.

Regarding Theorem D.1, the two conditions in Eq. (79) of the revised manuscript (Eq. (78) in the submitted version) are distinct: the dagger acts on the second jump operator in the first line and on the first jump operator in the second line. Thus, the two lines involve Ltμ∣k⟩⟨k∣(Ltμ)†L_t^\mu |k\rangle\langle k| (L_t^\mu)^\dagger and (Ltμ)†∣k⟩⟨k∣Ltμ(L_t^\mu)^\dagger |k\rangle\langle k| L_t^\mu, respectively.

The joy/sadness prompts are identical: the "sadness" prompt is a copy of the "joy" prompt.

Thank you for pointing out this error. The joy/sadness prompts were not used in the experiments reported in the paper, and we have removed these unused entries from the prompt table. This correction does not affect the reported results.

The uniform-diffusion checkpoint local_uniform/164931 is undocumented (training data, size, steps). The F.4 comparison cannot be reproduced.

Thank you for pointing out the missing training information. We have clarified in Appendix F.4 that we trained the uniform-diffusion model for 500,000 iterations using the default configuration of the official SEDD repository, which is linked there.

Code is not stated to be released, although the appendix refers to "the existing analysis script" and "the accompanying manuscript" (G.1), apparently left over from internal notes.

Thank you for pointing this out. We plan to publicly release the code. The references to “the existing analysis script” and “the accompanying manuscript” have been removed from the revised manuscript.

Hirono et al. 2024a and 2024b are the same paper.

Thank you for noting the duplicate reference. The revised bibliography contains a single entry for this paper, cited as Hirono et al. (2024).

Questions for Authors

Question 1

Do you agree that at β\beta = 1, SoP's joint distribution over (selected component, generated text) equals that of sampling the component first and generating conditionally, to first order in Δt\Delta t? If not, which step of the argument in Weakness 1 fails? Can you report SoP at β\beta = 1 together with the random-prompt baseline on the forest/snow experiment?

We thank the reviewer for identifying this equivalence. We agree that, for the original zero-Hamiltonian dynamics at β=1\beta=1, the joint distribution agrees, to first order in Δt\Delta t, with sampling a component from the initial weights and then generating conditionally on that prompt. Motivated by this observation, we have added Hamiltonian Prompt Mixing (HPM), in which coherences explicitly contribute to the evolution of sentence probabilities. This extension and its experimental results, described in our response to Question 3 below, have strengthened the paper, and we thank the reviewer for prompting this development.

Question 2

How does SoP at β\beta = 0.01 differ, empirically and conceptually, from averaging the two conditional reverse rates at every step (β\beta = 0)? Please include that baseline, and prompt concatenation, in Table 5.

Thank you for suggesting these controls. We have added both baselines in Appendix F.2 (Table 6 and Figure 5). At β=0.01\beta=0.01, the component weights are updated but strongly softened toward equal weighting, whereas β=0\beta=0 fixes them at 1/21/2. The two settings yield very similar similarity statistics. In contrast, prompt concatenation produces a qualitatively different similarity distribution from SoP: concatenated-prompt outputs are more similar to both reference conditions, whereas SoP outputs lie between the single-prompt baselines. This distinction persists when evaluated on the generated body alone.

Question 3

Is there any configuration of your framework (non-zero H~\tilde{H} , complex amplitudes, non-basis-preserving jump operators) that gives measurement statistics no classical mixture sampler can reproduce? If so, why is none of it explored? If not, what does the quantum formulation buy over a Bayesian mixture-of-conditionals sampler?

In the revised manuscript, we introduce HPM as a concrete extension in which coherence actively affects generation. Its nonzero Hamiltonian couples off-diagonal coherences to diagonal populations, so the component weights alone no longer specify the dynamics. Figure 4 shows that HPM shifts generated text toward the unselected prompt's references relative to no rotation. This construction remains efficiently simulable: each trajectory retains a shared body and only two prompt amplitudes. The dynamics and the role of coherences are derived in Appendix J.

Question 4

For superposed initial states, which forward process and which ρt\rho_t define the reverse operators used in Thm D.10? How does this square with Thm D.2's diagonal-ρt\rho_t assumption?

The reverse jump operators are constructed from the original forward diffusion process and its diagonal reference state, ρtref=∑npt(n)∣n⟩⟨n∣\rho_t^{\mathrm{ref}}=\sum_{\mathbf n}p_t(\mathbf n)|\mathbf n\rangle\langle\mathbf n|. This reference state is distinct from the superposed state evolved during generation. The diagonal-state assumption in Theorem D.2 therefore concerns the construction of the reverse operators, not the initial state supplied to the resulting dynamics. SoP applies these operators to a superposed initial state; it does not reconstruct the reverse operators from that state or assert that this evolution is the exact time reversal of a forward trajectory of the superposed state.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.