acceptodds
Under review as a conference paper at ICLR 2027

Efficient Reasoning with Flow Language Models

Abstract

Flow Language Models (FLMs) have emerged as a continuous-state alternative to discrete diffusion language models, yet the role of their continuous representations in reasoning remains unclear. We investigate this question by comparing the reasoning efficiency of FLMs and discrete diffusion models, measured by solution accuracy under matched denoising steps. Unlike discrete diffusion, which passes categorical states between denoising steps, FLMs evolve a continuous sequence representation throughout denoising and decodes it into discrete tokens only at the end. Our theoretical analysis shows, from a superposition perspective, how information retained in these continuous states can benefit reasoning. Intermediate-state interventions provide further empirical support for this theoretical account, showing that removing information about alternative candidates reduces subsequent solution recovery. Together, these findings show that FLMs allow evidence for multiple candidates to persist and inform subsequent reasoning before a discrete answer is produced. Furthermore, our experiments on maze planning and Sudoku tasks show that FLMs achieve greater reasoning efficiency in the few-step regime: FLMs achieves higher sequence accuracy than discrete diffusion baselines at matched model sizes and small denoising steps. On maze planning tasks, FLMs can also achieve comparable accuracy with smaller models. For example, on Maze15, FLM reaches the 95% accuracy target at 64 denoising steps with 36.5% fewer parameters than MDLM. These findings point to continuous state spaces as a promising foundation for reasoning models that require fewer refinement steps.

Primary area:
generative models
Keywords:
Reasoning, Flow Language Model

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_SqeEOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks whether continuous-state flow language models (FLMs) reason more efficiently than discrete diffusion language models, measuring sequence accuracy under matched denoising budgets on maze planning (Maze10/15/20) and Sudoku-Extreme, plus zero-shot transfer to Standard Sudoku. Theoretically, Theorem 1 shows the FLM ODE state accumulates a temporal convex superposition of candidate token sequences; Theorem 2 shows retained non-top-1 historical support can change prediction margins while the hard state stays fixed, with a tight Ωt/2 bound for any hard-state readout; Theorem 3 bounds the decoder-risk gap between continuous states and their hard projections. Empirically, at matched parameter counts FLM leads in the few-step regime: one-step sequence accuracy is 95.7/95.9/71.4 versus 94.0/52.6/35.6 for the best discrete baseline on Maze10/15/20, and on Maze15 at 64 steps FLM reaches the 95% target with an 88.25M model versus 139.00M for MDLM (36.5% fewer parameters). Mechanism studies (candidate-retention curves, intermediate-state interventions with Digit-soft and top-k controls, matched decoder probes) support the claim that retained non-top-1 mass is used by later denoising steps. On Sudoku the advantage is confined to 1–2 steps; beyond that, enhanced discrete samplers overtake FLM.

Critical strengths

Carefully controlled comparison. Architectures and parameter counts are matched per task and size (Tables 5, 9), step budgets are matched with an explicitly disclosed evaluation convention (K forward passes for FLM vs K+1 for discrete samplers), and the baseline suite includes exactly the methods that could break the claim (entropy/confidence remasking, ReMDM). The few-step advantage reproduces across three maze sizes, three model sizes per task, wall-clock latency (Table 10), and zero-shot transfer to Standard Sudoku (29.5 vs 27.1 at one step).

The theory–experiment loop is genuine. The superposition account (Theorems 1–3) is tested with interventions that include proper controls — Digit-soft isolates the projection operation itself, renormalized vs non-renormalized top-k separate rescaling from information removal, and an independent validation cohort reproduces the effect (47 of 48 correct-candidate events positive) — plus matched decoder probes whose largest Raw–Hard F1 gap occurs at step k=4 of an 8-step trajectory (0.8080 vs 0.5929). This is more than correlation-level evidence that retained candidate mass is computationally used.

Critical weaknesses

The headline comparisons carry no statistical support. Tables 1–2 appear to be single runs, and the test sets are 1,000 mazes per size and 1,000 Sudoku puzzles, where binomial noise alone is roughly ±1.5 points at p≈0.37. The Sudoku few-step margins the abstract leans on (21.0 vs 19.2 at one step, 36.7 vs 36.2 at two steps) lie within that noise, and no seeds, confidence intervals, or repeat runs are reported for any main-table entry.

The claimed advantage lives in a narrow regime on narrow tasks. On Sudoku, FLM plateaus after two steps and is clearly beaten from 16 steps onward (at 512 steps: 45.5 vs 62.2 for MDLM+Conf on Sudoku-Extreme); on Maze20 the 258.81M FLM never crosses 95% (92.9 at 512 steps); and all tasks are character-level synthetic outputs with no text-reasoning or language-modeling check, so whether "continuous states reason more efficiently" extends beyond toy planning remains open.

Questions for authors

How many seeds or runs stand behind Tables 1 and 2? In particular, are the 1–2 step Sudoku gaps (21.0 vs 19.2; 36.7 vs 36.2) stable across seeds given the 1,000-example test sets?

What is the rationale for the unequal maze training budgets (1,000 epochs for FLM vs 2,000 for discrete baselines, Table 6)? Does the few-step ordering change under equal epochs or equal wall-clock training compute?

Under strictly matched model evaluations (FLM at 2 steps vs a discrete sampler at 1 step, both = 2 evaluations), does the one-step advantage persist? The current convention grants discrete samplers one extra evaluation at every budget.

Is there any evidence the effect survives outside character-level synthetic tasks, e.g. a small text-reasoning benchmark or general language-modeling perplexity check?

3 comments
AuthorsOpenReview ↗

Response to Q1 and Q2 (Part 1/2)

We thank the reviewer for recognizing our carefully controlled comparisons and the genuine interplay between theory and experiments. We also appreciate the reviewer’s valuable feedback, which has helped us improve the quality of our paper.

We have uploaded a revised version of the paper, with all changes highlighted in blue. All references to pages, sections, figures, tables, and equations use the revised version's numbering.

[W1 & Q1 Results across multiple seeds]

Our original experiments used seed 0, and we added Sudoku experiments with seeds 1 and 2. The two tables below report the mean results across all three seeds, showing FLM’s consistent reasoning advantage with few denoising steps. These results are also included in the revised paper’s appendix.

Sudoku-Extreme: mean ± sample standard deviation (%).

MethodN=1N=2N=4N=8N=16N=64N=256N=512
D3PM9.80 ± 1.9812.90 ± 1.5618.05 ± 0.7825.00 ± 1.9829.85 ± 3.0438.40 ± 2.6940.60 ± 2.4041.45 ± 2.33
SEDD17.65 ± 1.2020.85 ± 0.4925.35 ± 0.9232.05 ± 0.6436.80 ± 0.4244.00 ± 3.9647.80 ± 1.2746.55 ± 1.91
SEDD + Ent.17.80 ± 1.4134.60 ± 2.2644.25 ± 0.2150.15 ± 1.2050.70 ± 2.4053.15 ± 0.6452.25 ± 2.1953.10 ± 1.70
SEDD + Conf.17.80 ± 1.4125.10 ± 1.8433.55 ± 1.7739.80 ± 0.7145.15 ± 1.0656.05 ± 0.2157.80 ± 0.1457.05 ± 0.49
MDLM18.60 ± 0.7120.90 ± 1.1325.80 ± 0.8532.25 ± 0.7838.60 ± 1.9845.30 ± 3.5447.65 ± 0.0747.60 ± 2.26
MDLM + Ent.18.65 ± 0.7832.00 ± 2.1241.80 ± 0.7147.65 ± 2.1949.05 ± 3.0451.40 ± 1.5650.60 ± 1.7052.05 ± 2.19
MDLM + Conf.18.65 ± 0.7825.90 ± 0.8535.70 ± 2.2641.05 ± 1.0645.80 ± 2.5557.55 ± 3.7560.35 ± 2.7660.45 ± 2.47
FLM22.95 ± 2.7639.55 ± 4.0345.70 ± 2.1247.05 ± 2.6246.75 ± 0.6446.45 ± 3.6147.30 ± 1.8446.90 ± 1.98

Standard Sudoku: mean ± sample standard deviation (%).

MethodN=1N=2N=4N=8N=16N=64N=256N=512
D3PM17.88 ± 2.3021.15 ± 2.4025.58 ± 1.8030.63 ± 2.4436.73 ± 2.6543.93 ± 2.5144.70 ± 1.9846.55 ± 1.91
SEDD24.98 ± 0.0427.90 ± 0.4231.55 ± 0.2136.83 ± 0.3942.73 ± 0.7449.63 ± 1.5950.85 ± 0.3552.43 ± 0.95
SEDD + Ent.25.23 ± 0.3940.83 ± 0.8150.98 ± 1.4554.10 ± 0.4257.40 ± 0.7157.25 ± 0.7857.55 ± 0.7156.70 ± 1.06
SEDD + Conf.25.23 ± 0.3931.75 ± 0.4939.35 ± 0.2144.25 ± 0.0050.50 ± 0.2860.38 ± 0.2561.60 ± 0.2862.05 ± 0.14
MDLM25.90 ± 1.7028.53 ± 1.3132.65 ± 1.7037.58 ± 2.0243.50 ± 2.1250.65 ± 1.9151.48 ± 2.5853.00 ± 0.00
MDLM + Ent.25.90 ± 1.7037.93 ± 2.2346.83 ± 2.0252.55 ± 1.8454.18 ± 1.9456.23 ± 1.1057.23 ± 1.5255.98 ± 2.30
MDLM + Conf.25.90 ± 1.7032.68 ± 1.3140.93 ± 1.0345.23 ± 1.6650.33 ± 1.6662.58 ± 2.1664.28 ± 1.8763.75 ± 0.78
FLM29.35 ± 0.2144.60 ± 4.5349.53 ± 2.5150.90 ± 2.2650.60 ± 1.5650.68 ± 0.3251.28 ± 0.1151.48 ± 0.32

[Q2 Training-budget fairness]

We use different training durations because discrete methods converge more slowly. While FLM has converged by 1,000 epochs, the discrete baselines have not, so we extend their training to 2,000 epochs until convergence. Additional results for MDLM, SEDD, and D3PM at epoch 1,000 on Maze 20 are shown below.

ModelDenoising StepsAccuracy
D3PM12.5%
D3PM89.5%
MDLM11.9%
MDLM87.9%
SEDD16.8%
SEDD817.0%
AuthorsOpenReview ↗

Response to Q3 and Q4 (Part 2/2)

[Q3 Strictly matched model evaluation]

We report the results under strictly matched numbers of model evaluations (NFE) below. FLM retains its advantage in the low-NFE regime. We have also included these results in Appendix D.8 (p. 32) of the revised paper.

DatasetMethodNFE=2NFE=4NFE=8NFE=16NFE=32NFE=64
Maze20FLM88.191.392.192.492.592.5
Maze20MDLM24.029.735.842.651.363.0
Maze20MDLM + Ent.24.022.022.622.623.123.2
Maze20MDLM + Conf.24.031.234.239.445.855.2
Maze20SEDD32.940.647.355.064.873.9
Maze20SEDD + Ent.32.958.082.890.290.888.5
Maze20SEDD + Conf.32.962.171.176.782.384.4
Maze20D3PM26.832.939.948.057.868.9
Maze20D3PM + Ent.26.827.228.028.028.528.0
Maze20D3PM + Conf.26.850.555.462.970.579.9
Sudoku-ExtremeFLM36.744.245.246.344.743.9
Sudoku-ExtremeMDLM20.324.931.039.544.746.7
Sudoku-ExtremeMDLM + Ent.20.342.448.449.953.454.2
Sudoku-ExtremeMDLM + Conf.20.328.036.641.949.554.0
Sudoku-ExtremeSEDD18.824.030.238.044.745.7
Sudoku-ExtremeSEDD + Ent.18.843.448.953.453.453.9
Sudoku-ExtremeSEDD + Conf.18.826.333.739.545.750.4
Sudoku-ExtremeD3PM11.616.524.331.438.741.4
Sudoku-ExtremeD3PM + Ent.11.630.940.041.842.446.6
Sudoku-ExtremeD3PM + Conf.11.619.530.536.741.248.1

[W2 & Q4 Scope and generalizability]

This paper primarily focuses on structured reasoning tasks, particularly planning and logical reasoning. These settings allow us to perform targeted interventions on intermediate states to further investigate the underlying mechanism. We consider extending our study to natural-language reasoning, including math and coding tasks, as future work. Moreover, our key strength is providing a theoretical explanation for FLM’s reasoning advantage with few inference steps.

We have made this scope more explicit throughout the paper to clarify the scope of our findings. Specifically, we have revised the abstract (p. 1, Lines 11--14), introduction (p. 1, Lines 47--48), and conclusion (p. 10, Lines 533--534) to emphasize our focus on structured reasoning and identify extensions to natural-language reasoning as future work.

We sincerely hope that we have addressed the concerns of the reviewer satisfactorily in the revised version and would kindly ask the reviewer to update their score accordingly.

Dear Reviewer,

As the discussion period draws to a close, we hope that our responses and revised manuscript have satisfactorily addressed your concerns. We would appreciate any further feedback and would be happy to discuss any remaining questions.

Thank you for your time and consideration.

Reviewer_t5MMOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript deals with discrete reasoning problems via embedding the discrete target into continuous embedding space. Their main claim is that having the process evolving in continuous space helps reasoning, since its continuous state at a time t contains information about several possible solutions, while e.g. in MDLM a state only contains the information if it is masked, or where it is committed.

Critical strengths

The question is interesting and relevant in order to understand the different behavior of reasoning models depending on whether they stay in discrete space or embed the tokens into continuous space. The statements about the information on intermediate times are interesting and the experiments give some interesting insights into if the additional information of an intermediate continuous state is beneficial. Weaknesses: The theoretical statements only concern the learned process, which is not made clear enough. The experiments are not yet sufficient to back their main claim. The connection to better few step generation is not clear. Details, see questions.

Critical weaknesses

The theoretical statements only concern the learned process, which is not made clear enough. The experiments are not yet sufficient to back their main claim. The connection to better few step generation is not clear. Details, see questions.

Questions for authors
  • In the numerical examples, sudoku and Maze, there is only one solution in the dataset per clue. Thus e.g. in Theorem 1, there is, for the perfectly learned network, no mixing going on. Why should this mixing etc. be an advantage in the case of an approximation?
  • For sudoku, the approaches with discrete intermediate states outperform the FLM starting at N=16. Why does the information in the intermediate states not help in the case of more steps? Especially since for N=1,2 FLM doesnt even benefit much from intermediate states, and as Figure 4 suggest, a benefit of FLM should be, that it can correct "wrong" intermediate states, which should help especially for N not too small
  • In Figure 4 it would be interesting to know, in how big the difference between the top 2 highest scoring probability were in creating this plot. Also it would be interesting to know, if only the intermediate states were "wrong" at the intervention time, or also the P_t.
  • In order to make it more readable, e.g. in Intermediate-state intervention. there could be less text, but more formulae or even a schematic algorithm.

Minor:

  • some notation is used before introduced, e.g. CG,qC_{G,q}
5 comments
AuthorsOpenReview ↗

Response to Q1 and Q2 (Part 1/2)

We sincerely thank the reviewer for the constructive feedback, which has helped us to substantially improve the clarity of our paper.

We have uploaded a revised version of the paper, with all changes highlighted in blue. All references to pages, sections, figures, tables, and equations use the revised version's numbering.

[Q1 Clarification of our superposition-based theoretical analysis]

An ideal predictor would output the same correct answer at every step. However, even when the solution is unique, a trained finite-capacity model may not recover it exactly and may need multiple denoising steps to approach it. Our focus is therefore on how superposition in intermediate FLM denoising states, before the final answer is decoded from the continuous state, can guide subsequent reasoning and thus improve the final solution.

Furthermore, Theorem 3 provides a theoretical account of how superposition can help recover information about the correct solution under a small denoising budget. Figure 4 provides empirical evidence for this mechanism.

To connect the proposed few-step mechanism to final solving performance, we further extend Theorem 3 in Appendix C.5 (Corollary 3; p. 26, lines 1350--1359), giving a sufficient condition under which retaining superposition information improves FLMs' final solution accuracy.

We also directly test how interventions on intermediate states affect final reasoning accuracy across different denoising budgets and intervention times (Table 3; pp. 9--10, lines 468--485). Delete removes non-winning history without renormalization; Delete+Transfer transfers that mass to the winner. Results are shown below, with t⋆=k/Nt^\star=k/N denoting intervention time.

Nt⋆NaturalDeleteDelete+Transfer
20.5036.0333.3332.60
40.2542.0741.9041.67
40.5042.0741.2040.80
40.7542.0741.9341.80
80.2543.8043.6343.47
80.5043.8043.2343.00
80.7543.8043.6343.50

Both interventions reduce final accuracy in every setting, supporting the contribution of retained history to few-step reasoning.

[Q2 Clarification of FLM’s few-step advantage]

At N=16N=16, as shown in Table 2, vanilla discrete methods such as MDLM do not outperform FLM (40.0% versus 46.3%). However, sampling strategies can improve inference performance, allowing MDLM+Conf (47.6%) to surpass FLM.

Theorem 3 helps explain how superposition can support few-step reasoning and when its additional benefit diminishes with further refinement. We apologize for not making this connection clear.

Specifically, the right-hand bound in Equation (14) implies that the gap between continuous-state and hard-state readouts vanishes as bt→0b_t\to0. This means that, as the hard state accurately represents the target, continuous values offer limited additional information. Retained superposition information can therefore benefit few-step reasoning while the current token choices remain insufficient to identify the target.

To make this point clearer, we have added the following corollary (p. 5, Lines 233--236 in the revised paper), which follows directly from the proof of Theorem 3:

Corollary 1 (Vanishing additional information benefit). With the notation of Theorem 3, let Y^=1[Π(Zt)]J=vpath\widehat{Y}=\mathbf{1}\\{[\Pi(\mathbf{Z}_t)]_J=v_{\mathrm{path}}\\} and qt=Pr⁡(Y^≠Y)q_t=\Pr(\widehat{Y}\neq Y). The Bayes-risk gap It:=H(Y∣Bt)−H(Y∣At)\mathcal{I}_t:=H(Y\mid B_t)-H(Y\mid A_t) satisfies 0≤It≤hbin(qt)0\leq\mathcal{I}_t\leq h_{\mathrm{bin}}(q_t). Consequently, along any sequence of states,

\quad\Longrightarrow\quad \mathcal{I}_t\longrightarrow 0.$$ We also extended our intermediate-state intervention experiments to $N=64$ to examine how removing historical information affects final reasoning accuracy at larger step budgets. At the midpoint ($k=N/2$), we remove historical information, then continue denoising. Each setting uses 1,000 puzzles and three random seeds. | Steps (N) | Original accuracy | After history removal | Drop (pp) | 95% CI (pp) | | ---: | ---: | ---: | ---: | ---: | | 2 | 36.03% | 33.33% | 2.70 | [2.07, 3.40] | | 4 | 42.07% | 41.20% | 0.87 | [0.40, 1.37] | | 8 | 43.80% | 43.23% | 0.57 | [0.13, 1.03] | | 64 | 44.87% | 44.70% | 0.17 | [-0.23, 0.57] | As the number of steps increases, removing historical information has a smaller effect on final accuracy. We additionally evaluated the methods on Sudoku at $N=4,8$ and added seeds 1 and 2 in the revised paper's Appendix D.7 (p. 33). The three-seed averages below show a consistent few-step advantage for FLM, including at $N=4$. These results support our focus on few-step reasoning. With more iterations, models can form reliable token choices with less reliance on retained historical information, and further computation may compensate for its removal.
AuthorsOpenReview ↗

Response to Q3 and Q4 (Part 2/2)

[Q3 Figure 4]

We evaluated the original Figure 4 Top-1 intervention in a paired follow-up with 300 puzzles, three seeds, and N=64N=64. For cells whose correct digit ranks second in Eˉt\bar E_t, the gaps are generally substantial, and PtP_t usually remains incorrect:

tGap: median [Q1, Q3]Z_t correctP_t correct
0.2030.753 [0.392, 0.963]2.93%7.53%
0.4060.788 [0.410, 0.970]3.23%14.18%
0.5940.820 [0.454, 0.976]3.69%13.79%
0.7970.921 [0.694, 0.990]2.60%3.65%

Gaps are normalized over all nine digits; correctness uses full-vocabulary argmax. When both ZtZ_t and PtP_t are incorrect, final cell recovery is:

tNaturalTop-1Difference (pp)Adjusted interval (pp)
0.20327.389%26.983%+0.405[-0.737, 1.621]
0.40615.223%14.199%+1.024[-0.209, 2.802]
0.5945.522%4.621%+0.900[-0.088, 2.157]
0.7972.380%2.015%+0.365[0.063, 0.733]

Intervals use paired puzzle-cluster bootstrap with Bonferroni adjustment across 36 comparisons. The latest-time effect remains positive after adjustment, supporting that the recovery benefit is not solely explained by an already-correct instantaneous output.

[Q4 Readability of the intervention experiment]

Thank you for this helpful suggestion. We have revised the “Intermediate-state intervention” paragraph to improve readability by shortening the prose and introducing explicit mathematical definitions of the candidate filtering operator, the intervention conditions, and the cell-level recovery rate.

[Minor]

Thank you for pointing this out. We have clarified this notation at its first use.

We sincerely hope that we have addressed the concerns of the reviewer satisfactorily in the revised version and would kindly ask the reviewer to update their score accordingly.

Thank you for your detailed answers and the improved manuscript. I still have the following remarks/concerns:

  • I think it would be good to note close to Theorem 1, that for an ideal predictor and only one solution per clue, that there is only one candidate
  • The results of table 3 are interesting, but the effects, except for N=2, are really small. This is true for N=4 ant t=0.25, where we should still be in the case, where " the gap between continuous-state and hard-state readouts" is big. Since the differences are so small, can one analyze for the cases that go wrong the mechanism?
  • Concerning Figure 4, the differences between natural and Top2 are really small. For Top1, can the effect be not explained that since you possibly delete the "correct" token and only retain the wrong one, that you go a bit out of distribution? Also just keeping the wrong one also pushes the current state more to the wrong state, is that correct?
Reviewer_EfDcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the comparison between the reasoning efficiency of FLMs and discrete diffusion models. The authors also provide a theoretical analysis to show how information retained in FLMs' continuous states can benefit reasoning. The paper finds that FLMs allow evidence for multiple candidates to persist and inform subsequent reasoning. Experiments on maze planning and Sudoku show that FLMs achieve greater reasoning efficiency in the few-step regime.

Critical strengths
  1. The paper studies an interesting question about the reasoning efficiency of FLMs and discrete diffusion models. The authors compare different methods with different model sizes and denoising steps. The FLMs have good performance in the few-step regime.
  2. The authors provide detailed theoretical analysis. They provide some insights about how continuous states retain information about multiple candidates and how this information affects subsequent predictions.
Critical weaknesses
  1. The experiments are limited. The paper only evaluates FLMs on maze planning and Sudoku tasks. There are no experiments on practical tasks like math and coding. So the generalization and practical usefulness of the conclusions are unclear.
  2. On Sudoku, FLMs only outperform discrete diffusion models slightly in the few-step regime, and perform worse with more steps. And the theoretical analysis does not fully explain the improvements. The proposed mechanism of retaining information across denoising steps cannot fully explain the one-step strong performance on maze planning tasks.
Questions for authors
  1. Please see the weaknesses.
  2. From Appendix D.2.2, the FLMs use argmax decoding, but the discrete diffusion models use sampling with temperature 1.0. Could this affect the fairness of the comparisons?
4 comments

We thank the reviewer for the valuable feedback, which has greatly helped us improve the clarity of our paper.

We have uploaded a revised version of the paper, with all changes highlighted in blue. All references to pages, sections, figures, tables, and equations use the revised version's numbering.

[W1 Scope and generalizability]

This paper primarily focuses on structured reasoning tasks, particularly planning and logical reasoning. These settings allow us to perform targeted interventions on intermediate states to further investigate the underlying mechanism. We consider extending our study to natural-language reasoning, including math and coding tasks, as future work. Moreover, our key strength is providing a theoretical explanation for FLM’s reasoning advantage with few inference steps.

We have made this scope more explicit throughout the paper to clarify the scope of our findings. Specifically, we have revised the abstract (p. 1, Lines 11--14), introduction (p. 1, Lines 47--48), and conclusion (p. 10, Lines 533--534) to emphasize our focus on structured reasoning and identify extensions to natural-language reasoning as future work.

[W2 Few-step gains and weaker performance with more steps on Sudoku]

Our Theorem 3 explains how superposition can support few-step reasoning and when its additional benefit vanishes with further refinement. We apologize for not making this clear before.

Specifically, the right-hand bound in Equation (14) implies that the gap between continuous-state and hard-state readouts vanishes as bt→0b_t\to0. This means that, as the hard state accurately represents the target, continuous values offer limited additional information. Retained superposition information can therefore benefit few-step reasoning while the current token choices remain insufficient to identify the target.

To make this point clearer, we have added the following corollary (p. 5, Lines 233--236 in the revised paper), which follows directly from the proof of Theorem 3:

Corollary 1 (Vanishing additional information benefit). With the notation of Theorem 3, let Y^=1[Π(Zt)]J=vpath\widehat{Y}=\mathbf{1}\\{[\Pi(\mathbf{Z}_t)]_J=v_{\mathrm{path}}\\} and qt=Pr⁡(Y^≠Y)q_t=\Pr(\widehat{Y}\neq Y). The Bayes-risk gap It:=H(Y∣Bt)−H(Y∣At)\mathcal{I}_t:=H(Y\mid B_t)-H(Y\mid A_t) satisfies 0≤It≤hbin(qt)0\leq\mathcal{I}_t\leq h_{\mathrm{bin}}(q_t). Consequently, along any sequence of states,

\quad\Longrightarrow\quad \mathcal{I}_t\longrightarrow 0.$$ With more denoising steps, discrete methods such as MDLM can reduce uncertainty about the token at each position, potentially improving solution accuracy. Sampling strategies also affect inference performance. For example, Table 2 shows that at 16 steps on Sudoku Extreme, FLM outperforms standard MDLM ($46.3\%$ versus $40.0\%$), but is surpassed by MDLM+Conf ($47.6\%$). We also added multi-seed Sudoku experiments to the revised paper’s Appendix D.7 (p. 32). The results show that FLM has a clear advantage within four denoising steps. **One-step Performance.** We agree that superposition cannot explain FLMs' one-step advantage. This advantage may reflect the difference in the training objective, which can affect how effectively the models learn to predict the solution from the task input in a single step. We have clarified this boundary of our theoretical account in the revised paper (p. 5, lines 247--248). **[Q2 Fairness of decoding strategies]** We further evaluate the argmax decoding strategies of discrete baselines on Maze20 and Sudoku-Extreme. The table below compares FLM with all three discrete baselines across different sampling budgets. FLM still outperforms discrete methods with few inference steps. The difference in the final decoding does not affect the fairness of the comparison in these evaluated settings. We have included these results in Appendix D.6 in the revised paper. | Dataset | Model | Decoding strategy | N=1 | N=2 | N=4 | N=8 | N=16 | N=32 | N=64 | N=128 | | --- | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | | Maze20 | FLM | | 71.4 | 88.1 | 91.3 | 92.1 | 92.4 | 92.5 | 92.5 | 92.6 | | Maze20 | MDLM | argmax | 24.0 | 27.3 | 31.0 | 36.9 | 43.5 | 52.1 | 63.0 | 72.4 | | Maze20 | SEDD | argmax | 32.9 | 37.5 | 43.2 | 49.3 | 55.8 | 64.9 | 74.3 | 80.4 | | Maze20 | D3PM | argmax | 26.8 | 30.6 | 35.6 | 41.4 | 48.8 | 58.2 | 68.9 | 77.6 | | Sudoku-Extreme | FLM | | 21.0 | 36.7 | 44.2 | 45.2 | 46.3 | 44.7 | 43.9 | 43.8 | | Sudoku-Extreme | MDLM | argmax | 20.3 | 23.1 | 26.8 | 32.4 | 39.9 | 44.9 | 46.4 | 49.0 | | Sudoku-Extreme | SEDD | argmax | 18.8 | 21.5 | 25.5 | 31.4 | 38.2 | 45.6 | 45.6 | 46.5 | | Sudoku-Extreme | D3PM | argmax | 11.6 | 14.7 | 18.4 | 25.3 | 32.9 | 39.4 | 41.3 | 42.7 | We sincerely hope that we have addressed the concerns of the reviewer satisfactorily in the revised version and would kindly ask the reviewer to update their score accordingly.

Dear Reviewer,

As the discussion period draws to a close, we hope that our responses and revised manuscript have satisfactorily addressed your concerns. We would appreciate any further feedback and would be happy to discuss any remaining questions.

Thank you for your time and consideration.

Thank you for the response. I still have concerns about the theoretical explanation, as the authors acknowledge that superposition cannot explain the substantial one-step performance advantage on Maze15 and Maze20.

Thank you for your time and valuable feedback. We believe that the one-step advantage and the contribution of retained candidate information to subsequent denoising should be examined separately. Our theoretical analysis shows how superposition retained in continuous states can support subsequent solution recovery with few denoising steps. The intermediate-state interventions in Section 4.4, “Superposition Analysis,” provide empirical support for this mechanism. A theoretical explanation of the one-step advantage requires further study. We have clarified the scope of our theoretical analysis in the revised manuscript. This does not affect our theoretical and empirical conclusions on how superposition supports few-step reasoning.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.