acceptodds
Under review as a conference paper at ICLR 2027

DS-OPT: Optimization-Native Reasoning via Dynamical System Machines

Abstract

Many structured reasoning tasks, such as logic puzzles and constraint satisfaction, require solving combinatorial optimization problems over many interacting discrete variables. Neural networks are built from primitives such as matrix multiplication, which excel at representation learning but are not native mechanisms for combinatorial search, so neural models must grow in size or computation to emulate reasoning over global constraints. We propose DS-OPT, a neural reasoning framework that adds a Dynamical System Machine (DSM) as a layer of the network. The neural layers learn to construct, for each input, a multi-state energy whose low-energy states answer the task, and the DSM minimizes this energy through its native discrete dynamics, using noisy winner-take-all units that reduce to an Ising machine in the two-state case. A differentiable relaxation makes the layer trainable end to end, while inference always runs the discrete DSM. On Zebra puzzles, visual Sudoku and binary image completion, DS-OPT is more accurate than neural baselines of comparable size. On Zebra it solves up to 98.4% of the puzzles, against 82.6% for a Transformer with five times more parameters, at half its inference time and energy, with DSM costs estimated by circuit-level simulation. These results suggest that DSM hardware can complement neural networks as an efficient substrate for combinatorial reasoning.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Dynamical Systems, Optimization, brain-inspired AI

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_25H1OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents a hybrid model that uses a transformer combined a differentiable DSM layer where coupling weights are learned. This allows the model to construct and solve a wide variety of Ising and Potts model problems. Crucially, a relaxation is used to make the DSM layer differentiable during training, and then inference in the DSM layer is done in the standard discrete manner. The model is tested on a small but diverse set of problems, showing improved performance over vanilla transformer models.

Critical strengths

The presented model is compelling, and the allegory with the brain is nice additional motivation. For the most part, the approach to making DSMs learnable seems very natural, and thus provides a good baseline for understanding how DSMs can be combined with neural networks.



The proposed model offers clear benefits over standard transformers. This shows that the addition of the DSM layers is indeed performing useful computation.

This is a particularly clearly written paper that is easy to understand and follow. 



The combination of reasoning and perception in the Sudoku problem gives an especially nice example of the power of these hybrid systems. This really highlights the novelty.

Critical weaknesses

The most useful application of this model appears to be in combinatorial optimization, but evaluation is limited to easy instances that classical algorithms can easily solve. 



Empirical comparisons are strong against traditional transformers, but appear somewhat mixed against more sophisticated models. It would be nice to have a conceptual comparison against the most closely related neural combinatorial optimization models, and to place the model in the context of neural algorithmic reasoning more broadly.

Although the DSM model seems effective, it is unclear where it will be useful over existing approaches, making it hard to judge practical impact.

Questions for authors

For the experiments, is there currently available DSM hardware that matches the performance of the simulated circuit? If so, why not use it? If not, what barriers remain? 



Have the authors considered hybrid models that use the DSM layer in conjunction with other neural algorithmic reasoning architectures, e.g., a recurrent transformer? The composability of the DSM layer seems to be a significant strength.



Can you give motivation for your choice of learned coupling rule? There are a wide range of choices that could be made here.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 25H1 (part1)

We thank the reviewer for the positive assessment and the constructive questions.

W1 & W3 (evaluation on easy combinatorial instances; where DS-OPT is useful).

Our target is not to solve explicitly stated combinatorial problems faster than classical solvers, which indeed handle Sudoku or Zebra puzzles easily once the rules are written down. It is reasoning tasks in which the problem is hidden in raw input and must first be constructed: a classical solver cannot read the digit images of visual Sudoku, the free-text clues of a logic puzzle, or write down the rules of image completion. Our results show that this construction step, not the solving, is where the difficulty lies. In visual Sudoku, solving the LeNet's fixed readings with the same DSM succeeds on only 61.1% of the boards, while joint reading and reasoning solves 97.2%. To make this point clearer, we have added ZebraLogic (Lin et al., 2025), which states Zebra puzzles in natural language and is challenging even for large reasoning models (new Table and Section 4.4):

ModelParamsAcc. (%)
DS-OPT (SmolLM2 interface + DSM)138M97.4
without couplings (J=0J=0)138M0.3
Transformer, same interface (one pass)138M77.6
Qwen2.5-0.5B, full fine-tuning0.50B81.6
Qwen2.5-1.5B, LoRA1.6B79.9
Qwen2.5-7B, LoRA7.2B70.1
Qwen2.5-32B-Instruct, QLoRA on 100k puzzles32B63.2
DeepSeek-R1 (zero-shot)671B78.7
Llama-3.1-405B (zero-shot)405B32.6

A small language model only needs to read the clues and construct the energy, while the DSM performs the reasoning; models up to 50 times larger that must also reason in their forward pass stay far below. As for practical impact, DS-OPT targets this class of perception- and language-grounded reasoning, and the DSM gives it an efficient substrate: circuit-level simulation estimates its inference time and energy at a fraction of those of neural models of similar accuracy (Tables 1–2), a step toward complex reasoning on power-limited edge devices. Scaling to larger and harder combinatorial instances, across several DSM cores, is a natural next step.

W2 (comparison with more sophisticated models; neural combinatorial optimization and algorithmic reasoning).

The published models that surpass DS-OPT use far more parameters or data: on Zebra, masked diffusion and flow models reach 98.5–100% with 14–19× more parameters and 5× more training puzzles, and on visual Sudoku IRED reports 98.3% with at least 36× more parameters. At comparable scale DS-OPT is the most accurate, and we have added comparisons that make this explicit: TRM and HRM scaled to our compute (Section 4.5) and the Recurrent Transformer trained with the same data as DS-OPT (Table 2). We have also measured the published models on an A100: their inference energy is 2–1,700× our circuit-level estimate. Following the reviewer's suggestion, we have added a paragraph to the related work on neural combinatorial optimization and neural algorithmic reasoning. Both learn how to solve a problem that is stated explicitly, either by constructing solutions or heatmaps for routing and graph problems, or by imitating the steps of classical algorithms. DS-OPT addresses the complementary setting in which the problem must be constructed from raw input, and leaves the solving to the native dynamics of the DSM.

Q1 (available DSM hardware).

No existing chip yet matches the multi-state DSM of DS-OPT exactly; we are developing this chip and expect to tape it out next year. As described in Appendix C.4, we have also run the 2-state form of DS-OPT on an existing coupled-oscillator Ising chip, where its behavior agreed well with the simulation.

Q2 (combining the DSM layer with other reasoning architectures).

We appreciate the reviewer for pointing out that the composability is actually a strength of DS-OPT. We have tested in a new experiment, where we place DS-OPT inside a recurrent loop in which the DSM state is fed back to the Transformer for a second pass. The looped model solves 99.1% of the Zebra test puzzles, compared with 98.4% for a single pass (Table 1), showing that the DSM layer composes with recurrent architectures and benefits from them. We have also mentioned this in the contribution part in the revision.

(Due to the word limit, the rest of response will follow in the next message.)

AuthorsOpenReview ↗

Response to Reviewer 25H1 (part2)

Q3 (choice of the coupling rule).

The couplings are an output of the network, in the same role as attention weights: the coupling between each pair of variables is computed by a symmetric pairwise map of their features, and this is the rule behind all our models. We learn the couplings rather than prescribe them because the rules are usually hidden in the input, in images, language or implicit structure, and cannot be written down in advance. The constraints on the form of this map come from the hardware rather than from the task:

  • the couplings are symmetric (Jji=Jij⊤J_{ji}=J_{ij}^\top), so that the energy is well defined;
  • interactions are pairwise, the only kind a DSM implements, and relations among three variables are expressed through auxiliary variables;
  • the energy is normalized in scale, so that a given noise amplitude has the same meaning in training and inference.

Thank you for your response. You have addressed all of my concerns and I have raised my score. Not being able to implement this model directly on current hardware is a significant limitation, as other reviewers have mentioned.

We sincerely appreciate your encouraging scores and valuable suggestions. Regarding your additional comment, we fully understand and agree that implementing the method on real hardware is important for a more reliable evaluation. To further strengthen the manuscript at this stage, we have now added and highlighted the experimental results from an existing coupled-oscillator Ising chip in Appendix C.4, providing additional evidence to support the reliability of our simulation-based evaluation. Due to the limited number of available spins, we mainly evaluated a standard synthetic 6×6 Sudoku dataset. The following table compares the hardware measurements with software simulation results, with all reported values averaged over 10 runs.

Table: Comparison of the 2-State Model Between Software Simulation and Measurements on an Oscillator-Based Ising Machine.

DatasetAccuracy (%)Inference (s)Energy (J)
Sudoku 6×6 (Simulated)98.01 ± 0.054.88 × 10⁻³2.44 × 10⁻⁴
Sudoku 6×6 (Measured)98.23 ± 0.034.65 × 10⁻³2.26 × 10⁻⁴

As shown in the table, the hardware measurements from the real chips show good agreement with the simulation results. Interestingly, we observe that the annealing quality (reflected as final accuracy) on real hardware is slightly better than that in simulation. One possible explanation is that the non-idealities in real hardware may perform as an additional noise source, which help the system better escape local minima during annealing. These results provide preliminary hardware evidence supporting the feasibility of our approach and the consistency between simulation and physical implementation.

We also would like to emphasize that we are actively developing a customized chip dedicated to DS-OPT, with tape-out planned for the first half of 2027. Once the chip is fabricated and tested, we plan to replace the corresponding simulation-based results with actual silicon measurements, enabling a more comprehensive hardware evaluation of our proposed approach.

Once again, we sincerely thank you for your time and constructive feedback, which have helped us further strengthen the manuscript. As the discussion phase is approaching its end, we regret that there may not be enough time to receive your further feedback. Nevertheless, we sincerely hope that our response and the additional experimental results have helped address your current concerns.

Reviewer_Erf7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper introduces a neuro-symbolic method that employs a Dynamical System Machine (DSM) as its symbolic layer. This layer operates by minimizing an energy function whose minimum corresponds to a solution to the combinatorial problem of interest. During training, a continuous relaxation of the discrete DSM is used, enabling end-to-end training of the entire architecture.

Critical strengths

I appreciated several aspects of the paper. First, I believe that combining combinatorial optimization with neural networks is a problem of great importance to the community. Moreover, I find the proposed approach very interesting. I was not previously familiar with Ising machines and DSMs, and I must say that the topic sparked my interest. The experimental results are also interesting, particularly those on Visual Sudoku, which is an especially challenging problem.

Critical weaknesses

I believe that the main weakness of the paper lies in its presentation rather than its technical content. Overall, I found the paper difficult to follow for readers who are not already familiar with the specific topic. In my view, a scientific paper, especially one submitted to a general-purpose conference such as ICLR, should be sufficiently self-contained to allow readers to understand the motivation, the proposed method, and the experimental analysis. In this case, I found issues with the presentation of all three aspects. I provide further details below.

Introduction and background.

The introduction and background seem to assume substantial prior knowledge of DSMs and Ising machines, making the paper difficult to understand for a broader audience.

A first example concerns the discussion of hardware, which is introduced without adequately explaining the relationship between DSMs, dedicated hardware implementations, and software simulations. In particular, it is not made clear why hardware is relevant to the proposed method. This discussion, together with Figure 2, occupies considerable space, despite appearing to play only a marginal role in the methodological and empirical contributions of the paper. I believe this space could be used more effectively to introduce the concepts necessary to understand the proposed method.

A second example concerns the presentation of Ising machines, which is somewhat circular. In the background section, Ising machines are defined as models that minimize the Ising energy described by Equation 3, which, however, only appears later in the section describing the proposed method. In that section, Equation 3 is used to show that a DSM can represent an Ising machine as a special case, while assuming that the reader is already familiar with Ising machines and introducing symbols such as bb and Ji,jJ_{i,j} without sufficiently clear explanations.

A third example concerns the concept of spin, which is used without ever being explicitly defined. Its meaning becomes clearer in the description of the method, just before Equation 3, but I believe that new terminology should be properly defined when first introduced.

Description of the method and distinction between training and inference.

Equation 4 introduces Gaussian noise applied to the values fi,vf_{i,v}. However, in Equation 5, which describes the continuous relaxation used during training, the noise seems to disappear, without a clear explanation of this choice. I would have expected the noise to be retained, similarly to the Gumbel-Softmax approach. The situation becomes even less clear in the description of the Straight-Through Estimator. At line 316, the authors state that the discrete state is sampled from m(k)m^{(k)}, but that the gradient is computed as if m(k)m^{(k)} had been used directly. It is therefore unclear whether m(k)m^{(k)} represents a probability distribution from which samples are drawn or a deterministic value. The former interpretation would seem consistent with the description of sampling and Equation 4, while the latter appears to be suggested by Equation 5.

The abstract also seems to suggest that noise is an important component of the dynamics, but it is not clarified whether it is used only during inference or also during training.

More generally, I believe the paper should distinguish much more explicitly between the operations performed during training and those used during inference. For example, although the overall context suggests that training relies on continuous relaxed dynamics, some statements seem to imply that the discrete dynamics are also used directly during training (e.g., lines 161–164).

Experimental analysis and comparison with baselines.

One aspect that I consider particularly important concerns the comparison with the baselines on the Visual Sudoku task. The authors report better results than NeSy programming and emphasize that their evaluation criterion is stricter, without clearly explaining in which way it is stricter, and what this difference entails.

More importantly, I believe that the comparison does not sufficiently account for the differences between the respective training procedures. NeSy Programming employs an initial image clustering phase and subsequently learns the Sudoku rules from the resulting clusters. These clusters do not necessarily correspond to the correct classes and may contain errors. In contrast, the proposed method employs a symbolic pretraining phase in which the rules are learned directly from grids represented using correct discrete symbols.

Consequently, NeSy Programming faces a more challenging rule-learning problem, as it must infer logical relationships from potentially noisy symbolic representations (with permuted symbols). Furthermore, its approach is more general in this respect, as it does not require access to additional correctly labeled symbolic grids beyond the visual examples.

This does not necessarily invalidate the experimental comparison, nor does it constitute a limitation of the proposed method in itself. However, I believe these differences are essential for correctly interpreting the results. Ideally, comparisons should be conducted under more comparable training conditions. At the very least, the authors should explicitly discuss these differences and their implications. Simply stating that NeSy Programming uses a pretraining phase (Table 2) and that its evaluation criterion is less strict (line 420) does not seem sufficient.

A similar issue concerns the neural baselines. The descriptions of their respective training procedures are not sufficiently detailed to clearly understand the conditions under which the comparison is conducted. In particular, it is unclear whether these baselines also benefit from a separate learning phase using symbolic grids or whether they are trained directly on images.

Questions for authors
  1. Why is Gaussian noise used instead of Gumbel or logistic noise? In the context of categorical variables, the Gumbel-Max trick provides a particularly natural connection to discrete sampling, while Gumbel-Softmax allows for a differentiable continuous relaxation. The authors mention this aspect in the paper but opt for Gaussian noise. Are there theoretical or empirical reasons for preferring Gaussian noise? Is this choice potentially related to hardware implementations of DSMs?
  2. Is the noise introduced in Equation 4 also used in the continuous relaxation employed during training, or is it only used during inference? In particular, how can the introduction of noise in Equation 4 be reconciled with its apparent absence in Equation 5?
  3. What is the actual role of hardware in the proposed contribution?
  4. It would be interesting to understand whether hardware implementations of DSMs could also be used during training, potentially in combination with a neural network running on conventional hardware. Is this possible? If so, how could communication between the two components be handled and, more importantly, how could gradients be propagated through the DSM?
  5. In the Sudoku task, does the model actually learn a separate 9×99 \times 9 matrix for each of the 81×8181 \times 81 pairs of cells, or are inductive biases or structural constraints introduced to explicitly exploit the structure of the problem, such as rows, columns, and blocks?
  6. Do the neural baselines used for Visual Sudoku also benefit from a separate learning phase using correctly labeled symbolic grids, or are they trained directly on images?
5 comments
AuthorsOpenReview ↗

Response to Reviewer Erf7 (part 1)

We sincerely appreciate the reviewer for the helpful comments, especially on the detailed guidance of how to improve the presentation of this manuscript. We have tried our best to address all the concerns mentioned in both weakness and questions parts. All the improved writings are highlighted in the revision, and we tried to briefly summarize them in this response.

W1 (also answers Q3): Paper presentation issue on Introduction and background.

We agree that the paper assumed too much familiarity with DSMs. We have rewritten the introduction and background section to help a reader without this background be able to follow the motivation and the method. The main changes are as below:

Why hardware matters. The introduction now motivates DSMs from the reasoning problem itself. Optimization is a natural primitive for structured reasoning, but on conventional processors it relies on exact solvers that iterate over a large search space, at high time and energy cost. A DSM performs this optimization natively in an analog circuit. Existing AI-oriented DSMs relax the energy into a continuous quadratic form to accept real-valued inputs, which reduces them to solving a linear system and gives up their combinatorial search. DS-OPT resolves this mismatch by letting neural layers handle continuous inputs and keeping the DSM discrete for reasoning.

How the method relates to the hardware. We now state explicitly that DS-OPT is trained on GPUs and that, at inference, the DSM layer runs natively on DSM hardware or is simulated on a GPU; all accuracies in the paper come from this simulation, and the time and energy columns from circuit-level simulation of the hardware. The hardware also determines the design of the method: the layer is built around the discrete noisy winner-take-all update that a DSM executes, and the training relaxation is designed to match it.

Self-contained background. Section 2 now introduces the basic concepts before they are used. A spin is defined at its first occurrence, the Ising energy is written out in Section 2.1 with every symbol explained, and the node update is described in plain terms (each update sets a spin to the sign of its local field, which lowers the energy or leaves it unchanged; added noise lets the state escape poor local minima). Section 3 now only refers back to this definition when showing that the multi-state energy reduces to it for two states, so the definition is no longer circular.

Space. The hardware figure and its circuit details have been moved to Appendix A, and the background has been shortened by removing motivation that is now in the introduction. The space saved is used for the explanations above and for a clearer separation of training and inference in Section 3 (see W2).

W2 (training vs. inference, noise in Eq. 5, the straight-through estimator).

We have revised Section 3 so that training and inference are separated explicitly:

Training and inference are now stated side by side. Section 3.3 opens by stating that training runs only the relaxation of Eq. (5), on a GPU, and never the discrete updates, while inference runs only the discrete noisy WTA updates of Eq. (4), natively on a DSM or simulated on a GPU, and never the relaxation. The sentence at lines 161–164 (in previous manuscript) and the abstract now say explicitly that the discrete noisy dynamics are used at inference and that training uses a noise-free relaxation of them.

Why Eq. 5 has no noise (related to Q2). The noise is not discarded but averaged out. As the reviewer notes, with Gumbel noise the WTA of Eq. (4) selects each state with probability softmax(fi/T)\mathrm{softmax}(f_i/T). The softmax in Eq. (5) is therefore the expected outcome of the noisy WTA, and its temperature TkT_k plays the role of the noise amplitude. Following this average rather than one noisy trajectory keeps the relaxation deterministic and its gradient free of sampling variance. We now explicitly state this in revision after Eq. (5).

What mm is. In Eq. (5), mi(k)m_i^{(k)} is a deterministic point on the probability simplex. In the discrete forward pass used for image completion, the same mi(k)m_i^{(k)} is treated as a categorical distribution: a one-hot state is sampled from it, which is exactly the WTA of Eq. (4) with Gumbel noise applied to ui(k)u_i^{(k)}. The sampled state drives the next step, while the gradient is computed through mi(k)m_i^{(k)} (straight-through). This is the only place where noise enters training. We now explicitly state this in Section 3.3 and Algorithm 1 in revision.

(Due to the word limit, the rest of response will follow in the next message.)

AuthorsOpenReview ↗

Response to Reviewer Erf7 (part 2)

W3 (Experimental analysis and comparison with baselines).

We thank the reviewer for this careful reading. We have revised related statements and add more details in the revision, and the summary is as below:

Evaluation criterion. Our statement was imprecise, and we have corrected it. We count a board as solved only if all 81 cells are correct, including the given digits read from images. This is the same as the total board accuracy of Li et al. (2023), who count boards that are both recognized and solved correctly, and stricter than the blank-cell accuracy reported for the Recurrent Transformer. The revised text states this explicitly and no longer claims a stricter criterion than Li et al.

Training conditions. We agree that this difference is essential for interpreting the results, and Section 4.2 now discusses it explicitly. Our couplings are learned from 1k symbolic Sudoku-Extreme puzzles with correct digits, whereas Li et al. learn the rules jointly with the digit reader from the visual boards alone. Beyond this, we also added two observations which suggest that most of the accuracy gap does not come from the extra data:

  1. Given the same 1k symbolic puzzles, the Recurrent Transformer improves only from 93.5% to 94.1% (94.9% when its test-time recurrence is extended to match our compute), still below DS-OPT's 97.6%.
  2. Li et al. recover the exact Sudoku constraints, so their rules at test time are as complete as ours, and their total accuracy equals their perception accuracy (95.5%): their errors come from misread digits. DS-OPT instead corrects misreadings during reasoning: with the LeNet's readings fixed, the same DSM solves only 61.1% of the boards, while joint reading and reasoning overrides 94.7% of the misreadings.

W4 & Q6: training conditions of the neural baselines.

We have added a table in Appendix G to list for every method, including whether it uses additional symbolic grids or perception pre-training, how the rules are represented, and the evaluation metric. In particular, The published Recurrent Transformer is trained directly on the visual boards, without any symbolic phase; TRM and HRM have no published visual Sudoku results, and we trained them ourselves with the same 1k symbolic puzzles added. To make the comparison fair, we retrained the Recurrent Transformer with the same 1k symbolic puzzles added, and DS-OPT remains the most accurate. The added results are in Table 2 and Table 5 now in revision.

Q1 (Gaussian vs. Gumbel noise).

We use Gaussian noise because it is what analog circuits provide naturally, with the same variance as Gumbel noise (π2/6\pi^2/6). And we do some ablation study on this to check whether the noise type matter in practical. As shown in the new Appendix F, we re-run inference with Gumbel noise, which makes every update an exact Gibbs step, in the setup of the main results. The two agree within 0.1 points: Zebra 96.47% (Gaussian) vs. 96.44% (Gumbel), mean of three models; visual Sudoku 97.56% vs. 97.61%, mean of four models. This is less than the variation from the decoding seed alone and thus we believe the noise type is not important here. And the insight is, the main role of the noise is to let the state escape local minima, and thus the exact distribution of noise doesn’t matter a lot.

Q2 (noise in training).

As we already explained in response to W2, the noise of Eq. (4) is used only at inference. Training uses the noise-free relaxation of Eq. (5), whose SoftMax is the expected outcome of the noisy WTA, so the temperature replaces the noise.

Q3: refer to W1

(Due to the word limit, the rest of response will follow in the next message.)

AuthorsOpenReview ↗

Response to Reviewer Erf7 (part 3)

Q4: DSM hardware during training.

This is a very good and interesting question. While it’s not in the scope of this work itself, we tried our best to provide a fair analysis as below.

On the one hand, A basic DSM is not suited to take part in training directly, because the gradient of the loss cannot be obtained from its discrete hardware dynamics as easily as from the relaxed dynamics simulated on a GPU. This is why we designed the training method of Section 3: the network learns through a differentiable relaxation on the GPU, and the DSM is used at inference.

On the other hand, training on the DSM itself becomes possible with additional circuitry. DS-TPU (Wu et al., 2025), for example, adds real-time feedback circuits to a DSM that compute the loss as a feedback current and adjust the programmable couplings and fields on the device, enabling online training. The same principle could in principle be applied to DS-OPT. Since the fields and couplings are produced by the neural layers, the DSM would return the gradient of the loss with respect to (h,J)(h, J) to the conventional hardware, which back-propagates it into the network. The added communication is of the same size as the (h,J)(h, J) sent from the GPU to the DSM in the forward pass, which our simulation already accounts for and which has a minor effect on the overall cost. On-device training is therefore feasible in principle, with the main added cost in the extra training circuitry; we leave its detailed design and analysis to future work.

Q5 (structure of the Sudoku couplings).

The model does not learn a separate D×DD\times D matrix per cell pair. Sudoku couplings are same-state (Potts) couplings, JijI9J_{ij} I_9, i.e. one scalar per cell pair, computed from learned position embeddings (Eq. 6). No row, column or block structure is built in: the couplings are learned from the symbolic puzzles by pseudo-likelihood with an ℓ1\ell_1 penalty. The learned couplings nevertheless recover the Sudoku constraint graph exactly: all 810 cell pairs that share a row, column or block receive a same-digit penalty (Jij<0J_{ij}<0), and these are precisely the 810 most strongly penalized pairs, with the weakest of them 1.16 times stronger than any coupling outside the graph. We have added this analysis to Appendix C.2.

Q6: refer to W4

Reviewer_pi4AOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors introduce DS-OPT a neural reasoning framework, combining a neural network to perceive inputs which are passed to a Dynamical System Machine (DSM) to perform energy minimization. DS-OPT integrated the DSM directly in the network and relaxes the discrete optimization task to be able to train the model end to end. This relaxation is discarded at test time, where discrete winner-takes-all annealing is performed. For evaluation, the proposed method is applied on zebra puzzles, visual sudoku and binary image completion. Variable couplings are created differently per task. For sudoku, couplings are constant across all problems and are therefore statistically learned before training. Zebra puzzles and binary image completion however do not share couplings across instances and are therefore constructed by the model alongside the external fields.

Critical strengths

++ The proposed relaxation of the discrete optimization problem is well motivated and subsequent evaluation shows that:

  1. the learnt couplings meaningfully enhance system performance. The ablation with J=0 is informative
  2. the DSM is able to correct mistakes made by the perception module
  3. it is possible to effectively transfer from the relaxed training target to the strictly discrete inference setup
  • The paper is very well written and structured, motivating their approach. Intuition and setup is built from the ground up and the relaxation and inference setup is clearly outlined.
Critical weaknesses
  • The results do not provide a fair and comprehensive overview of methods. While single pass transformer models are provided, it seems unfair comparing a model unrolling over multiple relaxation steps to a transformer performing a single forward pass. Comparing against TRM [1], HRM [2] would give insights how the proposed method fares against models also iteratively refining the prediction. Also, compared methods are roughly parameter matched, however the proposed method performs lots of refinement steps on hardware, making compute matched comparisons vital for a comprehensive comparison. With inference time and energy consumption left blank for most of the related literature, it is not possible to validate the efficiency claims against recent prior work in the field.

  • For the sudoku task, the shared couplings are computed from 1k Sudoku Extreme examples which the baselines you compare to apparently never see. Consequently, your method is exposed to more and different data in the training setup compared to the baselines.

  • It is not clear which parts of the proposed method are novel contributions. The division of labor [3] as well as the input-dependent energy construction [4] seems to be precedented in prior work. Therefore, It would be important to clearly outline the concrete, novel contributions in the introduction.

[1] Jolicoeur-Martineau, A. (2025). Less is more: Recursive reasoning with tiny networks. arXiv preprint arXiv:2510.04871. [2] Wang, G., Li, J., Sun, Y., Chen, X., Liu, C., Wu, Y., ... & Yadkori, Y. A. Hierarchical reasoning model, 2025. URL https://arxiv. org/abs/2506.21734. [3] Defresne, M., Gambardella, R., Barbe, S., & Schiex, T. (2026). Scaling neuro-symbolic problem solving: Solver-free learning of constraints and objectives. [4] Sercu, T., Verkuil, R., Meier, J., Amos, B., Lin, Z., Chen, C., ... & Rives, A. (2021). Neural potts model.

Questions for authors
  1. Can you clarify the concrete novel contributions of your work to this field as claims of your contribution seem to be covered by prior research (as stated in the weaknesses section)?
  2. How do (possibly compute matched) TRMs/HRMs compare to your method?
5 comments
AuthorsOpenReview ↗

Response to Reviewer pi4A (part 1)

We sincerely thank the reviewer for the valuable comments which helped us further improve the manuscript. We have tried our best to address the concerns in both weakness and questions as detailed below. The new results and contents are also reflected in the revision of paper and are highlighted in red.

W1&Q2: Comparison with TRMs/HRMs under compute matched setup, and inference & energy consumption of related works.

Comparison with compute-matched TRM/HRM-style methods

We compare DS-OPT with TRM and HRM under matched compute on Zebra, Visual Sudoku, and an additional dataset, ZebraLogic, which takes natural-language puzzle descriptions as input.

Scaling protocol. The original TRM/HRM configurations require substantially more compute than DS-OPT, so they must be scaled down for a fair comparison. Though we cannot measure the trade-off between model size and the number of recurrent iterations perfectly, we evaluate three scaling strategies for each baseline for fairness:

  • Original size: the model size is kept unchanged, and the number of iterations is reduced as far as possible.
  • Original iterations: the number of iterations is kept unchanged, and the model size is reduced until its compute matches DS-OPT.
  • Middle: an intermediate configuration between the two, giving a rough picture of the trade-off.

Shared heads. To rule out any effect of the input/output modules on accuracy, all TRM/HRM variants use the same Transformer, LeNet, and LLM(a 135M SmolLM2 for ZebraLogic) heads as DS-OPT on the respective tasks. Relative compute is computed with these heads excluded, so that it reflects only the difference between the methods themselves.

Zebra

MethodParametersCompute (× ours)Accuracy (best of 3 seeds)
DS-OPT0.09M1×98.4%
TRM (original size)6.8M3.24×94.9%
TRM (original iterations)27K1.26×28.7%
TRM (middle)410K1.09×85.2%
HRM (original size)27.3M5.40×93.9%
HRM (original iterations)170K0.91×26.4%
HRM (middle)3.2M0.98×77.0%

Visual Sudoku

MethodParametersCompute (× ours)Accuracy (best of 3 seeds)
DS-OPT1.23M1×97.6%
TRM (original size)8.10M1.60×55.7%
TRM (original iterations)1.24M0.97×0.0%
TRM (middle)2.02M1.03×0.0%
HRM (original size)28.5M2.70×64.2%
HRM (original iterations)1.58M1.00×3.1%
HRM (middle)7.78M1.00×3.6%

ZebraLogic (natural-language input)

MethodParametersCompute (× ours)Accuracy (best of 3 seeds)
DS-OPT0.09M1×97.4%
TRM (original size)6.8M3.23×84.0%
TRM (original iterations)27K1.26×43.8%
TRM (middle)410K1.08×65.7%
HRM (original size)27.3M5.38×80.4%
HRM (original iterations)170K0.91×45.1%
HRM (middle)3.2M0.98×69.9%

Summary. Across all three datasets, TRM/HRM-style methods achieve substantially lower accuracy than DS-OPT under comparable or even more compute. There are two main reason behind this results. First, only a very small part of DS-OPT are unrolled and iteratively computed, while TRM/HRM methods iterate their entire transformer model. Second, the DSM layer in DS-OPT naturally perform NP-hard solving, while TRM/HRM use transformer recurrent chains to emulate the solving process, resulting in much higher parameter/compute efficiency. We have added new sections to include these experimental results and analysis in Section 4.5 and Appendix D.

(Due to the word limit, the rest of response will follow in the next message.)

AuthorsOpenReview ↗

Response to Reviewer pi4A (part 2)

inference & energy consumption of related works

We have tried our best to reproduce most of the recent prior work and measured their training time, inference time, and energy cost on the same hardware configuration as ours (a single A100 GPU). The results are updated in the tables of the revision and are highlighted below.

Visual Sudoku

ModelParamsAcc. (%)TrainingInferenceEnergy
DS-OPT (ours, LeNet + DSM)1.23M97.63.6 min0.13 s33 J
Recurrent Transformer (Yang et al., 2023)0.70M93.52.1 h0.27 s68 J
IRED (Du et al., 2024)≥44.9M98.330.8 h74 s18.5 kJ
SATNet (Topan et al., 2021)14M64.81.8 h3.1 s786 J
SATNet (Chang et al., 2020)1.05M0.01.8 h3.1 s786 J

Zebra

ModelParamsAcc. (%)TrainingInferenceEnergy
DS-OPT (ours)1.33M98.41.1 h5.1 s1.3 kJ
AR Transformer (Shah et al., 2024)42M79.451 h5.1 min77 kJ
Masked diffusion (Kim et al., 2025)19M98.5238 h8461 s2.1 MJ
Flow Reasoning Model (2026)25M100.076.7 h8628 s2.2 MJ

On Visual Sudoku, the Recurrent Transformer has lower accuracy than DS-OPT and still requires about 35× more training time and about 2× the inference time and energy. IRED reaches accuracy comparable to DS-OPT (98.3% vs. 97.6%), but at roughly 500× the training time and more than 550× the inference time and energy. SATNet costs about 30× more to train and about 24× more at inference, while achieving much lower accuracy. On Zebra, the AR Transformer requires about 46× the training time and about 60× the inference time and energy of DS-OPT, while being nearly 20 points less accurate. Masked diffusion and the Flow Reasoning Model match or slightly exceed our accuracy. However, they require 216× and 70× our training time, respectively, and more than 1,600× our inference time and energy. These results further show that DS-OPT achieves competitive accuracy at a cost up to three orders of magnitude lower than recent prior works.

W2: Training setup alignment for sudoku task.

We thank the reviewer for pointing out this issue. We now state explicitly in Table 2 that, unlike the baselines, our model is also trained on the 1k Sudoku-Extreme puzzles, from which the shared couplings are learned. We also further retrained the Recurrent Transformer (Yang et al., 2023) with the 1k Sudoku-Extreme puzzles added: it solves 94.1% of the boards (mean of 3 seeds), and 94.9% when its test-time recurrence is extended to match our compute, both below DS-OPT (97.6%).

ModelParamsAcc. (%)TrainingInferenceEnergy
DS-OPT (ours, LeNet + DSM)1.23M97.63.6 min0.13 s33 J
Recurrent Transformer (Yang et al., 2023)0.70M93.52.1 h0.27 s68 J
+ Sudoku-Extreme 1k0.70M94.19.7 h0.31 s77 J
+ Sudoku-Extreme 1k, 141 recurrences0.70M94.9—3.0 s755 J

Also, in our new experiments with HRM and TRM (see our response to W1 & Q2), these puzzles are also added to the training data of HRM and TRM, so that all models are trained on exactly the same data.

W3&Q1 Clarification of the novelty and contributions

We thank the reviewer for these references and agree that assigning reasoning to a combinatorial layer with an input-dependent energy is not new by itself. We now cite Defresne et al. (2026) and Sercu et al. (2021) in the introduction and state how DS-OPT differs. We have also revised the contributions in the introduction accordingly. The first two contributions and differences with prior works are detailed below:

(1) Training for the hardware. Both works fit the energy by (pseudo-)likelihood: Defresne et al. minimize it with an exact solver on a CPU, and Sercu et al. do not minimize it. We train through a differentiable relaxation of the DSM’s own noisy winner-take-all dynamics, so the answer is the low-energy state reached by the hardware. NP-hard minimization thus becomes the native operation of the DSM, which cuts inference time and energy in circuit-level simulation (Tables 1–2).

(2) A composable layer. A solver-based layer trained by pseudo-likelihood must be the final layer of the network (Defresne et al., 2026). Our DSM layer is trained end to end without solver calls, so it can also sit inside a network. In a new experiment (added in Table 1) we place it in a recurrent loop between Transformer passes, and Zebra accuracy rises from 97.1% to 99.1% with two passes.

More broadly, we also want to emphasize that one of the aims of this work is to let DSMs, which minimize energies natively at low power, take over the optimization-heavy part of AI reasoning. It could enable new potentials for complex AI tasks eventually to run on power-limited platforms such as edge and implantable devices.

I acknowledge the additional experiments and clarification in the responses as well as the softening of the contribution claims in the revised submission.

Nevertheless, I maintain my "weak reject" score since the paper lacks any empirical evidence on dedicated hardware that support the a bit ambiguous claim of the abstract (These results suggest that DSM hardware can complement neural net- works as an efficient substrate for combinatorial reasoning.) and the misleading statement in the methodology section (... DSM-compatible optimization layer that turns the energy minimization performed natively by DSM hardware into a trainable reasoning primitive).

The only piece of information I could find is in appendix C.4: We have also run the 2-state form of DS-OPT on an existing coupled-oscillator Ising chip (Cılasun et al., 2025), where its behavior agreed well with the simulation. However, the 'agreed well' statement lacks any evidence.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Summary of the main revisions

Dear Area Chair and Reviewers,

We thank all reviewers for their thoughtful feedback and helpful suggestions. We have carefully revised the manuscript to address their concerns, with all changes highlighted in red. Below is a summary of the main revisions. More details can be found in our responses to individual reviewers.

1. Introduction, background, and method (Reviewer Erf7). We rewrote the introduction to better explain the motivation for using DSMs in reasoning tasks and how our approach connects to the hardware. Specifically, training is performed on GPUs, while the DSM handles energy minimization during inference. We also made Section 2 self-contained by defining spins and Ising energy when they are first introduced, and moved the hardware figure to the appendix. In Section 3, we clarified the training and inference procedures, including the roles of noise and the straight-through estimator.

2. Novelty and contributions (Reviewer pi4A). We added discussions of Defresne et al. (2026) and the Neural Potts Model in the introduction. We also revised the contribution statements to better highlight what makes DS-OPT different: the energy function is trained specifically for the DSM dynamics used to minimize it, and the DSM layer can be combined with other neural network layers.

3. Additional baselines and fairer comparisons (Reviewers pi4A, Erf7, 25H1).

  • We added TRM and HRM baselines on Zebra, visual Sudoku, and ZebraLogic, with three ways of scaling them to match our compute budget (new Section 4.5 and appendix E).

  • We retrained the Recurrent Transformer, TRM, and HRM using the same 1k symbolic Sudoku puzzles as DS-OPT. We also added an appendix table summarizing the training settings of all visual Sudoku baselines.

  • We measured the training time, inference time, and energy consumption of published methods on an A100 GPU using their released code.

4. New benchmark (Reviewer 25H1). We added ZebraLogic, a benchmark of logic puzzles written in natural language. For this task, we built DS-OPT on top of a small pretrained language model and compared it with fine-tuned and zero-shot large language models (new Section 4.4).

5. Composability (Reviewers pi4A, 25H1). We added an experiment that integrates the DSM layer into a recurrent loop. This further improves accuracy on Zebra, as shown in Table 1.

6. Additional analyses and corrections (Reviewer Erf7). We added a comparison of Gaussian and Gumbel noise in the appendix, along with an analysis showing that the learned Sudoku couplings exactly recover the constraint graph. We also corrected the description of the evaluation criterion for visual Sudoku.

7. Related work (Reviewer 25H1). We expanded the related work discussion to cover neural combinatorial optimization and neural algorithmic reasoning.

We hope these revisions can address the reviewers' concerns and make the paper clearer and stronger. We sincerely appreciate the reviewers' time and feedback, and welcome any further questions.

Sincerely,

The Authors

AuthorsOpenReview ↗

Discussion Summary and Overview of Manuscript Revisions

Dear Area Chair and Reviewers,

We sincerely thank all the reviewers for your valuable time, helpful suggestions, and constructive feedback. We are glad to see that most of the reviewers' concerns have been addressed during the discussion phase, and we believe our manuscript has been significantly improved. We are particularly encouraged by the positive feedback and progress made throughout the discussion:

  1. Reviewer pi4A maintained their score, recognizing the well-motivated relaxation, informative ablation studies, and clear presentation of our approach. We have carefully addressed their comments and provided additional hardware evaluation results in our latest response. As the discussion phase is coming to an end, we regret that there may not be enough time to receive further feedback, but sincerely hope these additions can help address their remaining concerns and strengthen their confidence in our work.

  2. Reviewer Erf7 raised their score from 2 to 3, expressing satisfaction with our responses and improvements to the manuscript's presentation. They highlighted the importance of combining combinatorial optimization with neural networks, found our approach particularly interesting, and appreciated the challenging Visual Sudoku experiments. We are especially grateful for their detailed suggestions on improving the paper's clarity and organization, which have greatly helped us improve the manuscript.

  3. Reviewer 25H1 raised their score from 3 to 4, showing strong support after our clarifications and additional experiments. They recognized the clear benefits over standard Transformers and highlighted the novelty of combining perception and reasoning through DSMs. They also appreciated the natural design and composability of our framework, which is further supported by our new experimental results. We sincerely appreciate their encouraging feedback and strong support for our work.

As the discussion phase concludes, we would like to summarize the major revisions made in response to the reviewers' feedback. All changes have been highlighted in red in the revised manuscript, and more details can be found in our individual responses.

1. Improved Introduction, Background, and Method (Reviewer Erf7)

We rewrote the introduction to better explain the motivation for using DSMs in reasoning tasks and their connection to hardware. We also made Section 2 more self-contained by introducing the necessary background on spins and Ising energy, and moved the hardware figure to the appendix. In Section 3, we further clarified the training and inference procedures, including the roles of noise and the straight-through estimator.

2. Clarification of Novelty and Contributions (Reviewer pi4A)

We expanded the discussion of related approaches, including Defresne et al. (2026) and the Neural Potts Model, and revised our contribution statements to better highlight the novelty of DS-OPT. In particular, we emphasized that the energy function is trained specifically for the DSM dynamics used to minimize it, as well as the flexibility of integrating DSM layers with other neural network components.

3. Additional Baselines and Fairer Comparisons (Reviewers pi4A, Erf7, 25H1)

  • We added TRM and HRM baselines on Zebra, Visual Sudoku, and ZebraLogic, with three different scaling strategies to match our compute budget (new Section 4.5 and Appendix E).
  • We retrained the Recurrent Transformer, TRM, and HRM using the same 1k symbolic Sudoku puzzles as DS-OPT, and added a detailed comparison of training settings for all Visual Sudoku baselines.
  • We measured the training time, inference time, and energy consumption of published baselines on an A100 GPU using their released implementations.

4. New Benchmark and More Complex Applications (Reviewer 25H1)

We introduced ZebraLogic, a more challenging benchmark of logic puzzles expressed in natural language. We extended DS-OPT by integrating it with a small pretrained language model and compared its performance with fine-tuned and zero-shot large language models (new Section 4.4).

5. Improved Composability (Reviewers pi4A, 25H1)

We added a new experiment that integrates the DSM layer into a recurrent loop, demonstrating the flexibility and composability of our framework. This integration further improves accuracy on Zebra, as shown in Table 1.

(Due to the word limit, the rest of summary will follow in the next message)

6. Additional Hardware Validation (Reviewers pi4A, 25H1)

To further strengthen the reliability of our simulation-based evaluation, we expanded Appendix C.4 with additional hardware details and experimental results measured on an existing coupled-oscillator Ising chip. Specifically, we compared the real chip measurements with simulation results on 6×6 Sudoku, including accuracy, inference time, and energy consumption. The results show good agreement between hardware measurements and simulation, providing preliminary hardware evidence to support the feasibility of our approach and the reliability of our simulation-based evaluation.

7. Additional Analyses and Corrections (Reviewer Erf7)

We added a comparison between Gaussian and Gumbel noise, along with an analysis showing that the learned Sudoku couplings exactly recover the constraint graph. We also corrected and clarified the evaluation criterion for Visual Sudoku.

8. Expanded Related Work (Reviewer 25H1)

We expanded the related work section to include more discussions on neural combinatorial optimization and neural algorithmic reasoning, providing a more comprehensive overview of related research.

We are glad to see that these revisions have addressed most of the reviewers' concerns and significantly improved the quality and clarity of our manuscript. Once again, we are truly grateful for your time, thoughtful suggestions, and encouraging feedback throughout the review and discussion process.

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.