Pith. sign in

REVIEW 4 major objections 6 minor 3 references

A single diffusing skyrmion can act as a programmable multi-state probability unit, with voltages tuning its stationary distribution.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A single diffusing skyrmion in a patterned magnetic film can implement tunable multi-value probability distributions, softmax-like sampling, and invertible OR logic without wiring many probabilistic bits together.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A credible experimental proof of concept for skyrmion-based multi-value probabilistic computing, with a real invertible OR gate, but the Boltzmann/softmax mapping rests on an untested detailed-balance assumption and energy parameters fitted to the same data. the 4 major comments →

arxiv 2508.19623 v1 pith:OV7OUQ2G submitted 2025-08-27 cond-mat.mtrl-sci cond-mat.dis-nn

Multi-value Probabilistic Computing with current-controlled Skyrmion Diffusion

Thomas B. Winkler , Yuean Zhou , Grischa Beneke , Fabian Kammerbauer , Sachin Krishnia , Mario Carpentieri , Davi R. Rodrigues , Mathias Kl\"aui
show 1 more author
Johan H. Mentink
This is my paper
classification cond-mat.mtrl-sci cond-mat.dis-nn
keywords magnetic skyrmionsprobabilistic computingmulti-value logicinvertible logicsoftmaxspin-orbit torqueBrownian diffusionBoltzmann distribution
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper claims that the random thermally activated motion of a single magnetic skyrmion through a thin film containing several pinning sites can serve as a multi-value probabilistic computing element. The central idea is that the fraction of time the skyrmion spends in each pinning site is a Boltzmann probability, so the spatial occupancy map directly is a discrete probability distribution; an applied voltage, through spin-orbit torque, changes each site's energy approximately linearly and thus reshapes the distribution. Because Boltzmann weights equal softmax outputs, the device can approximate the softmax function used in neural networks, and by assigning truth-table rows to pinning sites it can run invertible logic without interconnecting many binary probabilistic bits. The evidence is long Kerr-microscopy trajectories at several voltages (six pinning sites), a Markov-state model of the hopping, Monte Carlo training of a small network on the Iris dataset using the diffusion-based softmax, and an experimental four-site invertible OR gate whose distributions are selected by minimizing KL divergence to the clamped target. If the linearly-tunable-Boltzmann picture holds, this points to compact, low-power hardware that stores uncertainty as physical dwell-time statistics.

Core claim

On its own terms, the paper establishes that a skyrmion diffusing in a pinning-dominated landscape is a physical realization of a K-outcome probability distribution. The stationary occupancy of pinning site i is claimed to be π_i(U) = e^{-E_i^0 - c_i U} / Σ_j e^{-E_j^0 - c_j U}, with the coefficients c_i set by the current-density path; this is exactly a softmax over the voltage-modified energies. The authors verify this model against measured occupancy for a six-site device under voltages from -1.5 mV to 2 mV, then use the same model to implement an invertible OR gate in a four-site device, obtaining KL divergences below 0.1 for the 1-clamped and unclamped cases. They also show by Monte Car

What carries the argument

The central object is a single magnetic skyrmion treated as a Brownian quasi-particle moving through an effective energy landscape with K pinning sites. The key identity is the Boltzmann–softmax equivalence π_i = e^{-E_i}/Σ_j e^{-E_j} = softmax(-E_i), which turns dwell-time statistics into a read-out probability vector. The tuning mechanism is current-generated spin-orbit torque modeled as a conservative force, so a voltage U shifts each site energy linearly, E_i(U)=E_i^0+c_i U, preserving detailed balance and keeping the distribution Boltzmann. The analysis pipeline is Markov state modeling: trajectories are coarse-grained into K sites by maximizing self-transition 'crispiness', and voltage

Load-bearing premise

The load-bearing premise is that the spin-orbit torque acts as a conservative force and Joule heating is negligible, so a voltage-biased skyrmion obeys detailed balance with a Boltzmann dwell-time distribution E_i(U)=E_i^0+c_i·U—if non-conservativity or heating appears, the softmax and invertible-logic mappings collapse; note the paper attaches this Boltzmann claim to a citation '(Brems et al., 2025)' that is absent from its reference list.

What would settle it

Take one device and measure the stationary occupancy at a series of increasing voltages long enough to test whether ln(π_i/π_j) is linear in U for every pair of pinning sites. Any systematic curvature, or a net circulation in the Markov-chain transition matrix indicating broken detailed balance, would falsify the Boltzmann-softmax mapping; the same measurement at matched temperatures would separate genuine spin-orbit-torque effects from Joule heating.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A single skyrmion device can represent K discrete probabilities at once, eliminating the network of binary p-bits needed by earlier invertible-logic hardware.
  • Because each truth-table row maps to its own pinning site, invalid logic states are never sampled, which the authors argue yields close to 100% sampling accuracy for invertible gates.
  • The device can act as a physical softmax layer for classifiers and attention mechanisms, with best accuracy when pre-activation differences are small—the regime relevant to attention computations.
  • Adding more current paths or local electrodes generalizes the energy model to E_i = E_i^0 + Σ_l c_{l,i} U_l, allowing more complex, locally tunable probability distributions.
  • The number of computable states grows combinatorially with the number of skyrmions and pinning sites, offering a scaling route toward multi-skyrmion probabilistic processors.

Where Pith is reading between the lines

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

  • If the linear voltage-energy relation holds over a wider voltage range, the device becomes a physical log-probability adder: each independent current path contributes additively to log π_i, so probabilities factor into independently controllable terms—a natural fit for Bayesian networks and factor graphs.
  • The same experiments could probe the validity boundary of the conservative-force assumption: at higher current densities, non-conservative spin-orbit torque and Joule heating should break detailed balance, and one could test whether non-equilibrium operation trades sampling accuracy for higher speed.
  • With focused-ion-beam or laser-written pinning sites, the energy landscape itself could be programmed during fabrication to encode a fixed base distribution, with voltages used only as a continuous reweighting control.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper reports a proof-of-concept implementation of multi-value probabilistic computing (MPC) using a single magnetic skyrmion diffusing among K pinning sites in a triangular thin-film device. The core claim is that the time-averaged spatial occupation of the skyrmion realizes a tunable discrete probability distribution, with the pinning-site energies shifted linearly by an applied voltage through spin-orbit torques. Using this mapping, the authors demonstrate (i) a softmax-like computation, although via Monte Carlo simulations rather than experimentally, and (ii) an experimentally realized invertible OR gate, where truth-table entries are assigned to pinning sites and voltages are selected from the fitted energy model. The manuscript emphasizes that this approach avoids networks of binary p-bits and offers scalability advantages.

Significance. If the central physical claim holds—that a current-biased skyrmion in a pinning landscape reaches a Boltzmann stationary distribution with voltage-tunable energies—then this is a genuinely novel MPC element with high information density: a single physical device can represent a multi-outcome probability distribution and perform invertible logic without coupled stochastic elements. The experimental Kerr-microscopy data are substantial (hours-long trajectories at multiple voltages), the direct truth-table-to-pinning-site mapping is elegant, and the OR-gate voltage selection provides a partial out-of-sample test of the fitted model. These strengths make the paper potentially significant for probabilistic and unconventional computing. However, as detailed below, the validation is partly in-sample and the detailed-balance assumption underpinning the Boltzmann mapping is asserted rather than tested, so the physical interpretation is not yet fully established.

major comments (4)
  1. [Markov State Modeling] The central mapping from time-averaged occupation to a Boltzmann distribution rests entirely on the assertion that the spin-orbit torque is conservative and Joule heating is negligible. This is stated, not demonstrated. For the triangular geometry the current density is spatially nonuniform, and a non-conservative force or a thermally driven probability current would make the stationary distribution non-Gibbsian; the linear energy model E_i(U)=E_i^0+c_i U would then be an empirical curve fit rather than a physical softmax/invertible-logic substrate. I request an explicit experimental check of detailed balance, e.g., verifying the Kolmogorov cycle condition or comparing pi_i W_ij with pi_j W_ji from the tracked trajectories at each voltage. A KLD value around 0.1 for one OR-gate setting is not sufficient to confirm the mechanism.
  2. [Methods: Numerical minimization of the energy model] The parameters E_i^0 and c_i are fitted to the same experimental occurrence maps pi_exp(U_m) used for the 'good agreement' shown in Fig. 3b. The loss function L uses pi_exp at exactly the measured voltages, so Fig. 3b is an in-sample fit, not an independent validation of the Boltzmann/linear-energy hypothesis. This circularity weakens the claim that the modeled distributions are predictions. Please reframe Fig. 3b as model calibration, and add a genuine out-of-sample test—for example, fit on a subset of voltages and predict the held-out voltages, or compare the inferred energies with independent transition-matrix estimates from the same trajectories. The OR-gate voltage scan is a step in this direction and should be highlighted as such, but it currently covers only a few voltages and a single device.
  3. [Invertible Logic Gate with skyrmion diffusion based MPC] The headline 'competitive performance' of the invertible OR gate relies on KLD values below 0.1 for the 1-clamped and unclamped cases, while the 0-clamped case is explicitly bounded by the chosen voltage range and not below 0.1. No error bars or repeated-measurement statistics are reported for the experimental stationary distributions. To support the claim of competitive accuracy, please provide bootstrap or split-half uncertainties on pi_exp and KLD, and report the KLD for all three target distributions with their statistical spread. As it stands, the comparison to s-MTJ implementations (Zhang et al., 2025) is made without a common statistical basis.
  4. [Skyrmion-diffusion based softmax calculation] The softmax demonstration is a Monte Carlo simulation of a rigid skyrmion-like particle, not an experimental realization. The abstract and introduction state that the system 'demonstrates softmax computation,' which could be read as an experimental claim. Please clarify in the text and abstract that the experimental contribution is the voltage-tunable stationary distribution, whereas the softmax/neural-network result is a numerical proof-of-concept based on the modeled energy landscape. This distinction is important for assessing what has been demonstrated at the hardware level.
minor comments (6)
  1. [Markov State Modeling] The Boltzmann formula is written as "pi_exp(x)=e(-(V(x)))/S" with the factor beta omitted and the notation garbled. Please define the discrete normalized distribution explicitly: pi_i = exp(-beta E_i)/Z, with Z = sum_j exp(-beta E_j). Also clarify the units of energy given that beta is later absorbed.
  2. [Methods: Clustering algorithm] The thresholds dmax=0.2, clim=1, and olim=3 appear to be chosen ad hoc. A sensitivity analysis would strengthen confidence that the number of pinning sites and the transition matrix are robust to these choices.
  3. [Invertible Logic Gate, Fig. 5] The notation "pi(U)=pi_{1-clamped}, pi_{0-clamped} or pi_{unclamped}" is unclear. Please define the target distributions explicitly, e.g., pi_goal for output clamped to 0, output clamped to 1, and unclamped uniform over valid OR states.
  4. [Methods: Experimental details] The reference to "micmag2" is a GitHub link without a version or DOI. Please provide a citable reference or a more complete description of the current-density simulations so that the initial slope values can be reproduced.
  5. [Bibliography] There is a typo: "Fischer, 1936" should be "Fisher, R. A. (1936)". Also, some references (e.g., Virtanen et al.) are missing author lists or use incomplete formatting.
  6. [Figure 4] In Fig. 4d, please specify whether the error bars are standard deviations or standard errors and over what exactly the 10 runs are averaged. The current caption is ambiguous.

Circularity Check

2 steps flagged

The Boltzmann/linear-energy model is validated against the same data used to fit it, and the softmax simulation builds softmax into the Metropolis-Hastings sampler; the experimental invertible-OR demonstration retains independent content.

specific steps
  1. fitted input called prediction [Markov State Modeling / Figure 3b; Methods: 'Numerical minimization of the energy model']
    "Figure 3b shows the modeled stationary distribution and its scaling with current, obtained from the calculation of the Boltzmann distribution. We compare the modeled results to experimental data that is marked as triangular points, showing good agreement."

    The energies E_k^0 and slopes c_k are not independently predicted or measured: the Methods loss L = Σ_m Σ_k (πexp_k(U_m) − πfit_k(U_m))^2 is minimized against the very same experimental occurrence maps plotted in Fig. 3b. Therefore the 'modeled stationary distribution' is the least-squares best fit to those data, and the 'good agreement' is enforced by construction rather than providing an independent validation of the Boltzmann/linear-energy model.

  2. self definitional [Skyrmion-diffusion based softmax calculation / Figure 4]
    "At each time step, the skyrmion motion was determined using the Metropolis-Hastings algorithm, governed by the local energy at each position. ... The local energy in each region was used as pre-activations for the softmax function. The output of the diffusion-based softmax function is based on the dwell time within each region."

    Metropolis-Hastings is a Markov-chain Monte Carlo sampler whose stationary distribution is, by construction, the Boltzmann distribution exp(−E)/Z of the imposed energies. Setting the local energies equal to the negative pre-activations makes the sampled dwell-time distribution converge to softmax(y) by definition. The simulation therefore instantiates softmax inside the sampler rather than testing whether a physical skyrmion device computes softmax; it is a self-definitional demonstration.

full rationale

The experimental invertible-OR result is not itself circular: the voltage is selected using a fitted model, but the final distributions (stars in Fig. 5a) are newly measured at the chosen voltages and compared with the goal via KLD. That part has genuine hardware content. The circularity is partial and located in two supporting claims: (i) the 'modeled stationary distribution' in Fig. 3b is a fit to the same experimental data it is compared against, so it cannot validate the Boltzmann/linear-energy model; and (ii) the softmax simulation uses a Metropolis-Hastings sampler whose target distribution is softmax by construction, so the simulation is a tautological implementation rather than an independent check. The detailed-balance assumption (conservative SOT, no Joule heating) is a stated physical assumption and a correctness risk, not a circular step. No load-bearing self-citation chain was identified.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The paper's quantitative predictions of probability distributions rest on two fitted parameters per pinning site plus voltage offsets and hand-chosen clustering thresholds; no new physical entities are introduced. The softmax simulation is a Monte Carlo implementation of the same energy model.

free parameters (5)
  • E_i^0 (baseline pinning energies) = fitted, not tabulated (see Fig. 3a)
    Baseline energy of each pinning site at zero voltage, fitted to experimental occurrence maps at 5 voltages.
  • c_i (voltage slopes) = fitted, shown in Fig. 3a
    Scaling of each pinning site energy with applied voltage, fitted with initial values from current path simulations (micmag2).
  • G(U) integration constant per voltage = set to 0 for U=0 mV; fitted for other voltages
    Auxiliary offset in the initial-value problem E_init= -ln(pi_exp)+G(U); accounts for gauge freedom in the energy model.
  • Clustering thresholds dmax, clim, olim = 0.2, 1, 3
    Hand-chosen thresholds used to filter transitions/microstates before Markov state clustering; affect the inferred K and distributions.
  • Softmax simulation pre-activation range and sampling steps = -1..1; 100,000 steps
    Choices in the Metropolis-Hastings simulation; larger separations between pre-activations require longer sampling, as the paper notes.
axioms (5)
  • domain assumption The driven skyrmion system satisfies detailed balance and reaches thermal equilibrium with stationary distribution pi ~ exp(-E_i)
    Invoked in 'Markov State Modeling': SOT treated as conservative force, Joule heating neglected.
  • domain assumption Transitions between coarse-grained pinning sites over the camera lag time (1/16 s) are Markovian
    Markov State Modeling used to extract transition matrix W with lag time tau; clustering maximizes self-transitions.
  • domain assumption Skyrmion can be treated as an overdamped rigid particle (Thiele equation) with negligible skyrmion Hall angle
    Used for the softmax Monte Carlo simulation and for the conservative-force model.
  • domain assumption Pinning sites are naturally given by material imperfections and can be robustly identified as K discrete states
    K=6 and K=4 chosen from cumulative occurrence maps; the whole MPC mapping depends on this discrete coarse-graining.
  • ad hoc to paper Linear voltage scaling E_i(U)=E_i^0+c_i U
    Postulated energy model fitted to the target data; not derived from micromagnetic theory.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-value Probabilistic Computing with current-controlled Skyrmion Diffusion." pith.science (2026). https://pith.science/paper/OV7OUQ2G

@misc{pith2026250819623,
  author       = {Pith},
  title        = {Pith review of: Multi-value Probabilistic Computing with current-controlled Skyrmion Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OV7OUQ2G}},
  note         = {Machine review of arXiv:2508.19623}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Magnetic systems are highly promising for implementing probabilistic computing paradigms because of the fitting energy scales and conspicuous non-linearities. While conventional binary probabilistic computing has been realized, implementing more advantageous multi-value probabilistic computing (MPC) remains a challenge. Here, we report the realization of MPC by leveraging the thermally activated diffusion of magnetic skyrmions through an effectively non-flat energy landscape defined by a discrete number of pinning sites. The time-averaged spatial distribution of the diffusing skyrmions directly realizes a discrete probability distribution, which is tunable by current-generated spin-orbit torques, and can be quantified by non-perturbative electrical measurements. Even a very straightforward implementation with global tuning, already allows us to demonstrate the softmax computation - a core function in artificial intelligence. As a key advance, we demonstrate invertible logic without the need to create a network of probabilistic devices, offering major scalability advantages. Our proof of concept can be generalized to multiple skyrmions and can accommodate multiple locally tunable inputs and outputs using magnetic tunnel junctions, potentially enabling the representation of highly complex distribution functions.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

3 extracted references · 1 canonical work pages

  1. [86]

    https://doi.org/10.1214/aoms/1177729694 Lukasiewicz, T. (1999). Probabilistic and truth-functional many-valued logic programming. Proceedings 1999 29th IEEE International Symposium on Multiple-Valued Logic (Cat. No.99CB36329), 236–241. https://doi.org/10.1109/ISMVL.1999.779722 Maharjan, S., Alsadoon, A., Prasad, P. W. C., Al-Dalain, T., & Alsadoon, O. H. ...

    arXiv 1999

  2. [2194]

    decoder-less

    https://doi.org/10.1016/j.joule.2023.09.004 Duffee, C., Athas, J., Shao, Y ., Melendez, N. D., Raimondo, E., Katine, J. A., Camsari, K. Y ., Finocchio, G., & Khalili Amiri, P. (2025). An integrated-circuit-based probabilistic computer that uses voltage-controlled magnetic tunnel junctions as its entropy source. Nature Electronics. https://doi.org/10.1038/...

    arXiv 2023

  3. [8977]

    Multi-valued logic system: New opportunities from emerging materials and devices

    https://doi.org/10.1038/s41467-024-53270-w Onizawa, N., Nishino, K., Smithson, S. C., Meyer, B. H., Gross, W. J., Yamagata, H., Fujita, H., & Hanyu, T. (2021). A Design Framework for Invertible Logic. IEEE Transactions on Computer- Aided Design of Integrated Circuits and Systems, 40(4), 655–665. https://doi.org/10.1109/TCAD.2020.3003906 25 Parr, T., Rees,...

    work page arXiv 2021

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.