{"id":"3e711393-3670-4f7d-a26b-55c5db6cf3ee","arxiv_id":"2607.16183","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Tunable energy landscapes whose thermal averages equal sigmoid, softmax, and matrix-vector products can, in principle, form the basis of a low-energy analog computer, with a superconducting double-well device as a first experimental step.","lead":"This paper lays out a design for a new kind of computer that uses the random thermal jiggling of cold superconducting circuits as a computing resource, showing how basic machine-learning operations could be built from tunable energy landscapes. It also reports a first experimental building block — a superconducting double-well \"neuron\" — though the measured thermal-activation energies only partially match the theory.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Experiment cannot distinguish wrong device model from wrong fitted parameters: E_esc plateau ~193 mK vs T_cross 46–76 mK, slope 12 vs 20.8 GHz/K, absorbed by 'very large uncertainty' in C, L, I_c.","rationale":"The reader's weakest_assumption exactly identifies the Langevin/thermal-activation description of the device as the structurally distinct premise that the experimental section stands or falls on. The reader's CONDITIONAL verdict already asks for propagated parameter uncertainties, end-to-end energy accounting, and non-destructive readout. My stress-test confirms this: the E_esc vs k_B T discrepancy (3x plateau, 40% slope) is the single most load-bearing concern in the experimental claim. The framework itself is internally consistent and supported by standard identities and numerical checks; no concern there rises to the level of invalidating the blueprint. The concrete test I propose is the minimal one that would settle whether the device model is wrong or just under-constrained by the current fit: a full uncertainty propagation through the extraction pipeline. If the model survives, the CONDITIONAL verdict stands; if it fails, the experimental claim would need to be downgraded to a qualitative demonstration, not a quantitative validation of the Langevin framework. I agree with the reader rather than partially because the reader's weakest_assumption and the reader's proposed condition (address the E_esc vs k_B T discrepancy with propagated parameter uncertainties) are precisely the same concern I identify, and I do not see another concern of comparable weight.","tokens_in":55010,"tokens_out":1644,"duration_ms":15164,"concrete_test":"Propagate full joint uncertainties of the Hamiltonian fit (C, L, I_c, junction asymmetry, inline inductances) through Eqs. (60), (62), and (65) to produce a confidence band for E_esc(T). If a single consistent parameter draw can reproduce both the ~193 mK plateau and the ~12 GHz/K slope while also matching the resonator-frequency map in Fig. 17, the device model survives; if no draw within the fit's confidence region does, the discrepancy indicates a wrong device model or missing physics and the experimental validation claim must be weakened.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing experimental claim is that the fabricated thermodynamic neuron is described by the underdamped Langevin circuit model of Eq. (61) with a single parallel resistance R, so that escape follows the Kramers/Arrhenius form Eq. (62) with E_esc = k_B T in the thermal regime. The data in Fig. 13(c) do not confirm this: the low-temperature plateau corresponds to ~193 mK while the independently estimated crossover from Eq. (60) is 46–76 mK (roughly 3x excess), and the slope above 100 mK is ~12 GHz/K versus k_B/h = 20.8 GHz/K (≈40% low). The paper's response is that C, L, and I_c have 'very large uncertainty' and that unmodeled quasiparticle damping could change a_t (Sec. VIII C, Appendix H). That is an in-sample absorption of the discrepancy: the same parameters that fix the plateau also set ω_b, ω_p, and ΔU used to extract E_esc, so a 3x error in E_esc can be explained by shifting parameters without independent verification. Since the device is currently limited to binary readout and cannot continuously measure flux or work statistics, the experiment as presented cannot distinguish a wrong device model (e.g., R not a single temperature-independent resistor, multi-mode effects, or non-thermal activation) from wrong fitted Hamiltonian parameters. The framework-level potential constructions are unaffected, but the paper's claim of a 'preliminary experimental realization' of the Langevin neuron is not established by this data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for energy-based thermodynamic computing based on sampling from the equilibrium distribution of underdamped Langevin dynamics. It introduces elemental potentials whose thermal expectations approximate sigmoid (Eq. 21), softmax (Eq. 29), matrix-vector products (Eq. 27), and addition (Eq. 28), and derives a covariance-based gradient estimator for parameter gradients (Eq. 34). The framework is then embedded in factor graphs, extended to on-chip self-learning via a Born-Oppenheimer force, and illustrated with numerical demos (PGM, GMM, HMM, continuous Ising, thermoformer). A superconducting double-well 'thermodynamic neuron' is presented as a preliminary experimental realization, with escape-rate measurements intended to confirm the Kramers/Arrhenius prediction E_esc = k_B T (Sec. VIII).","tokens_in":55321,"tokens_out":6808,"duration_ms":64041,"significance":"The theoretical core of the paper is largely sound and useful: the potential constructions are explicit and concrete, the gradient identity of Eq. (34) is derived from first principles and is potentially hardware-friendly, and the factor-graph assembly is a coherent organizing principle. The numerical demonstrations are illustrative, and the framework makes falsifiable predictions for thermal expectations that could be tested in simulation or hardware. However, the experimental section, which is the paper's claim of a hardware proof-of-principle, does not quantitatively validate the thermal-activation model: the data in Fig. 13(c) show a low-temperature plateau corresponding to ~193 mK against a predicted T_cross of 46–76 mK, and a high-temperature slope of ~12 GHz/K against the expected k_B/h = 20.8 GHz/K. The authors' own explanation relies on parameter uncertainty and unmodeled quasiparticle damping, but no independent calibration is provided. Thus the hardware claim is not established, while the framework-level derivations stand.","major_comments":[{"comment":"The escape-energy data do not confirm the claimed thermal-activation regime. The low-temperature plateau E_esc/h = 4.02 GHz corresponds to ~193 mK, while T_cross estimated from Eq. (60) is 46–76 mK; the slope above 100 mK is ~12 GHz/K versus k_B/h = 20.8 GHz/K. The authors attribute this to 'very large uncertainty' in C, L, I_c and to quasiparticle damping. However, the same fitted parameters determine the ω_b used in T_cross and the ω_p and ΔU used in the Arrhenius fit of Eq. (65), so the comparison is not an independent test. With binary readout only and no direct measurement of R(T) or the thermodynamic neuron's own frequency, the experiment cannot distinguish a wrong device model from wrong fitted parameters.","section":"Sec. VIII C, Fig. 13(c), Eqs. (60)-(62)"},{"comment":"The identification of the low-temperature plateau with k_B T_cross is asserted without derivation. For macroscopic quantum tunneling, the escape rate is not generally of the Arrhenius form with E_esc = k_B T_cross; the effective activation energy depends on the action and damping. Moreover, the missing data between 60 and 100 mK (Appendix J) means the crossover region is never directly measured, so the transition from plateau to linear rise is inferred, not observed. A conclusive test requires resolving the crossover or fitting the full escape-rate expression with independently calibrated parameters.","section":"Sec. VIII C, Eq. (60), Appendix H"},{"comment":"The extraction of E_esc via Eq. (65) relies on ω_p and ΔU computed from the Hamiltonian fit. According to Appendix H, the fit uses 10 free parameters after fixing C and is performed only against readout resonator frequencies, not against the thermodynamic neuron's dynamics. The authors state the fit has 'very large uncertainty' in C, L, I_c, yet they do not propagate this uncertainty into E_esc or report confidence intervals for the slopes in Fig. 13(b). The claim that the measured slope is 40% below k_B/h is therefore not statistically grounded. Please propagate the fit covariance into E_esc and provide an independent estimate of ω_p and ΔU.","section":"Sec. VIII B 2, Appendix H"}],"minor_comments":[{"comment":"The displayed Hamiltonian is corrupted by stray tokens (\"⌟⟨⟨⟪rl⟫l⟩⟩...\") and is unreadable. The equation must be reset to the intended expression.","section":"Eq. (59)"},{"comment":"The variance potential is described as computing the variance of z, but the formula involves auxiliary x_i and y with the term (x_i - (z_i - μ))^2. The role of these auxiliary variables and the domain of validity should be clarified, especially since the text states this potential is not meant to be an equilibrium realization.","section":"Sec. VII E, Eq. (58)"},{"comment":"Several core components are cited only as pending patent applications by the same authors (e.g., Refs. [100-102], [114-115], [134], [147-149], [189], [202-203], [205]). For a journal publication, please provide public references or full technical descriptions in the text so that reviewers and readers can verify these components.","section":"References"},{"comment":"No code or data repository is provided for the numerical experiments (PGM, GMM, HMM, Ising, thermoformer) or for the experimental escape-rate data. The manuscript would be stronger if the authors released code/data, or at least provided the raw fit parameters and uncertainties for the key figures.","section":"General"},{"comment":"The tokens-per-joule projections exclude cryogenic cooling, control electronics, and calibration overhead, as stated in the text. The comparison with H100 GPUs in Fig. 9 should be interpreted with care; a clear caveat in the caption would prevent over-generalization.","section":"Sec. VII, Fig. 9"}],"recommendation":"major_revision","confidential_remarks":"The paper's heavy reliance on pending patent applications by the authors for core components is unusual for a journal submission and may raise concerns about completeness and prior disclosure. The experimental section would benefit from raw data, propagated uncertainties, and an independent calibration of the Hamiltonian parameters. The editor may wish to consider whether the 'Blueprint' framing with a hardware demonstration is appropriate given that the hardware validation is currently inconclusive."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the theoretical core is genuine and checkable: the double-well sigmoid (Eq. 21), softmax (Eq. 26/29), matrix-vector product (Eq. 25), and addition (Eq. 24) potentials do produce the claimed thermal expectations, and the gradient rule (Eq. 34) is the standard score-function/REINFORCE estimator with zero baseline, correctly derived and naturally matched to the hardware. These constructions are the real contribution. Second, the experimental section is a real first step but does not yet validate the device model. The measured escape-energy plateau corresponds to ~193 mK while the predicted crossover is 46–76 mK, and the slope above 100 mK is 12 GHz/K versus the expected 20.8 GHz/K. The paper absorbs this into 'very large uncertainty' in C, L, and I_c, which is an in-sample absorption: the same parameters set ω_b, ω_p, and ΔU used to extract E_esc. So as presented, the experiment cannot distinguish a wrong device model (temperature-dependent damping, multimode effects, non-thermal activation) from wrong fitted parameters. This does not touch the framework-level potential constructions, which stand on their own.\n\nCredit where due: the paper is unusually candid about its own gaps. It flags that estimation oscillators are hard on superconducting hardware, that the layer-norm potential is not an equilibrium realization, and that the chip energy projections exclude cryogenic cooling, control, and calibration. The factor-graph/PGM synthesis is a useful framing, and the toy thermoformer gives a concrete sense of where sample counts bite.\n\nSoft spots, in proportion: the experimental validation gap is the main one. The heavy reliance on 13 pending patent applications for core components is a mild concern—not a flaw per se, but it means the devices are not independently checkable yet. Eq. (59) contains corrupted inserted text, likely a LaTeX artifact; minor, but it should be cleaned.\n\nWho this is for: people working on stochastic analog hardware, EBM training, or stochastic thermodynamics. It deserves a serious referee: the theory is checkable, the hardware direction is substantive, and the limitations are stated honestly. I'd bring it to reading group and cite the potential constructions if I worked in this area.","headline":"A checkable theoretical blueprint for thermodynamic ML primitives; the hardware prototype is real but its validation does not yet confirm the Langevin model.","tokens_in":55967,"tokens_out":3276,"would_cite":true,"duration_ms":29225,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Thermal equilibrium of a tunable physical system can be programmed to run machine-learning operations, with gradients read from measured covariances — a path to computing near the Landauer limit.","keywords":["thermodynamic computing","energy-based models","Langevin dynamics","Gibbs distribution","probabilistic graphical models","superconducting circuits","contrastive divergence","Landauer limit"],"falsifier":"Measure escape energy E_esc versus temperature on a device whose parameters C, L, I_c are pinned down by direct spectroscopy of the neuron itself, not inferred from resonator fits. If the low-temperature plateau still corresponds to ~193 mK rather than the predicted T_cross of 46-76 mK, or the thermal slope still falls short of k_B/h = 20.8 GHz/K, the single-resistor Langevin description of the device is falsified. A complementary numerical check: simulate the circuit with temperature-dependent quasiparticle resistance and see whether any parameter set within the stated uncertainties reproduce","tokens_in":54781,"feed_emoji":"🔥","tokens_out":11065,"duration_ms":88424,"temperature":0.7,"pith_summary":"This paper argues that a physical system evolving under underdamped Langevin dynamics settles into the Gibbs distribution, so any energy potential that can be physically built is automatically a programmable thermal sampler. The authors construct explicit potentials whose thermal averages implement the elementary operations of neural networks — sigmoid, softmax, matrix-vector product, and addition — and they derive a gradient rule, ∂E[y]/∂θ = −Cov(y, ∂E_θ/∂θ), that turns backpropagation into a covariance measurement on the same hardware. On these primitives they assemble a blueprint for differentiable machine learning, including factor-graph models and a thermodynamic transformer, and report a first experimental building block: a superconducting double-well 'thermodynamic neuron.' The payoff, if the framework holds, is sampling-based machine learning with energy consumption potentially orders of magnitude closer to the Landauer limit than deterministic digital computing. The authors themselves flag that the transformer's layer-norm block is not an equilibrium realization and that projected chip energies exclude cryogenic, control, and calibration overhead.","feed_headline":"Thermal noise runs machine learning in new computing blueprint","feed_subtitle":"A tunable potential's thermal averages do sigmoid, softmax, and matrix products; gradients are measured covariances.","key_machinery":"The load-bearing object is the Gibbs distribution as the exact steady state of the underdamped Langevin equation: fluctuation-dissipation-matched damping and noise guarantee that the equilibrium of any physically built potential U_θ is the Boltzmann weight e^(−βU_θ), however complicated the landscape. Riding on it are the potential 'gadgets' that translate ML operations into energy landscapes (double-well → sigmoid, simplex-constrained wells → softmax, tilted Gaussians → matrix-vector products and sums), the covariance gradient identity ∂E[y]/∂θ = −Cov(y, ∂E_θ/∂θ) that converts backpropagation into a sampling task, and the factor-graph rule that composes gadgets into larger models. On the ex","core_discovery":"The central claim: equilibrium physics is a sufficient substrate for differentiable machine learning. Underdamped Langevin dynamics (Eq. 8) has the Gibbs distribution π_θ(x) ∝ e^(−βU_θ(x)) as its steady state, so a tunable potential U_θ turns hardware into an energy-based model whose samples come from the physics itself. Specific potentials make thermal expectations equal the sigmoid (Eq. 21), softmax (Eq. 29), matrix-vector products (Eq. 27), and addition (Eq. 28); the identity ∂E[y]/∂θ = −Cov(y, ∂E_θ/∂θ) (Eq. 34) turns training into covariance estimation. These blocks compose as factor graphs into mixture models, HMMs, continuous Ising machines, and a thermodynamic transformer. A fabricate","pith_inferences":["The energy projections exclude cryogenic cooling, control electronics, and calibration overhead; if those dominate, as they do in today's superconducting systems, the practical energy advantage over digital hardware remains open even if the equilibrium-computation claim is correct.","The covariance-gradient identity is substrate-agnostic: any fluctuating physical system whose steady state is Boltzmann — optomechanical, CMOS, photonic — could run the same training rule, so the blueprint's core could outlive superconducting hardware.","The roughly 3x low-temperature escape-energy excess is directly testable: independent characterization of C, L, and I_c by neuron spectroscopy would separate a wrong device model from wrong fitted parameters — a decisive experiment the current data cannot yet adjudicate.","A natural next test is whether the sub-k_B/h slope above crossover (12 vs 20.8 GHz/K) is caused by temperature-dependent quasiparticle resistance; a device with a normal-metal shunt of known resistance would make the damping model checkable."],"forward_implications":["If equilibrium sampling is as programmable as claimed, energy-based models become trainable on hardware that draws each sample in roughly one thermalization time, removing the sampling bottleneck that makes EBMs intractable digitally.","Gradient training of a computation graph reduces to estimating a cross-covariance per parameter block — a native analog operation — so backpropagation can be implemented without digital differentiation.","Precision becomes a continuously tunable knob: error scales as N^(−1/2) in sample count, with relative error uniform across magnitudes, unlike floating-point's fixed mantissa.","Idealized work per coupling operation is of order k_B T (against a Landauer floor of ln 2·k_B T), the paper's quantitative argument that thermodynamic ML could run orders of magnitude below digital energy per operation.","The same primitives compose as factor graphs, so any model expressible as a probabilistic graphical model — transformers, HMMs, mixtures — inherits the paradigm."],"fun_headline_variants":["Equilibrium physics runs differentiable machine learning","Thermal noise yields gradients for hardware ML","Thermodynamic hardware computes ML with heat","Langevin dynamics as a differentiable ML substrate","Stochastic analog circuits train ML from thermal averages"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the fabricated superconducting neuron is exactly the idealized underdamped Langevin system, with a single parallel resistance modeling both loss and noise so that escape obeys the Arrhenius law with E_esc = k_B T; the paper's own Fig. 13(c) shows a low-temperature plateau equivalent to ~193 mK against a predicted crossover of 46-76 mK and a thermal slope of ~12 GHz/K against k_B/h = 20.8 GHz/K, discrepancies the authors attribute to the very l","fun_headline_variants_meta":{"raw":{"variants":["Equilibrium physics runs differentiable machine learning","Thermal noise yields gradients for hardware ML","Thermodynamic hardware computes ML with heat","Langevin dynamics as a differentiable ML substrate","Stochastic analog circuits train ML from thermal averages"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000319,"raw_usage":{"total_tokens":1620,"prompt_tokens":712,"completion_tokens":908,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":456,"completion_tokens_details":{"reasoning_tokens":841}},"tokens_in":456,"tokens_out":908,"duration_ms":7713,"temperature":1.0,"reasoning_tokens":841,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T21:07:15.155316+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure escape energy E_esc versus temperature on a device whose parameters C, L, I_c are pinned down by direct spectroscopy of the neuron itself, not inferred from resonator fits. If the low-temperature plateau still corresponds to ~193 mK rather than the predicted T_cross of 46-76 mK, or the thermal slope still falls short of k_B/h = 20.8 GHz/K, the single-resistor Langevin description of the device is falsified. A complementary numerical check: simulate the circuit with temperature-dependent quasiparticle resistance and see whether any parameter set within the stated uncertainties reproduce","supporting_citations":[],"review_version":1}