{"id":"43fe9ab7-6cbb-476a-891f-ce39e1c067b2","arxiv_id":"2412.10317","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A probabilistic delay cell using a superparamagnetic tunnel junction samples exponential dwell-time distributions in the time domain, with proposed circuits for Metropolis-Hastings and weighted random sampling.","lead":"This paper builds a circuit that samples random numbers from an exponential distribution by measuring when a tiny magnetic switch flips. The goal is hardware that directly produces the distributions needed for Monte Carlo sampling and weighted random draws.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Drift in the rate parameter is acknowledged but not bounded; a quantitative bound on drift-induced bias in the weighted sampler would settle whether the applications hold.","rationale":"The reader's weakest_assumption is exactly the stationary-lambda assumption underlying the applications, with Appendix A and the tau(I) reduced chi-squared as evidence. My stress-test agrees and sharpens the attack: the exponential first-passage claim (Fig. 1(c)) can be true while Eq. (4) fails, because Eq. (4) requires jointly fixed rates across devices and across time. The paper itself flags the drift explicitly in Sec. IV and Appendix A, so the concern is not manufactured; it is a limitation the authors acknowledge but do not quantify in application-level terms. I considered whether the drift is merely a device-maturity issue outside the paper's claims. It is not: the applications section (Sec. V) asserts the circuits implement Metropolis-Hastings and weighted random sampling, and those implementations inherit the drift sensitivity. The paper gives no drift magnitude bound, no compensation scheme, and no demonstration of tolerance, so the conditional recommendation stands. I do not escalate to REJECT because the experimental exponential claim is independently supported by the CDF fit, the tunability data, and honest reporting, and a drift-compensated or recalibrated version of the proposal may well work. A single concrete experiment—measuring lambda(t) and simulating the sampler under observed drift—would settle whether this concern is decisive or merely a calibration detail.","tokens_in":19221,"tokens_out":2597,"duration_ms":20868,"concrete_test":"Re-analyze the Appendix A drift trace and the Fig. 4 tau(I) data to extract a time-resolved rate lambda(t) (e.g., moving-window mean dwell time over the 205 s run). Then simulate 10^6 draws of the two-device weighted sampler with lambda_1 and lambda_2 drawn independently from the empirically observed drift distribution, and compare the resulting selection-probability histogram against the nominal lambda_1/(lambda_1+lambda_2). If the relative selection error exceeds the application tolerance (e.g., 1%), the drift should be flagged as preventing the claimed sampler functionality. Repeat with interleaved short calibration runs as a control to confirm the drift, not the exponential single-shot noise, is responsible.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that each current step produces exponentially distributed first-switching times with rate lambda, and that the proposed circuits use these samples for Metropolis-Hastings steps and exponential-clocks weighted sampling. The paper's own Appendix A and Sec. IV document substantial drift: the AP-state probability wanders over 205 s with a spread of 0.042 s versus the 0.015 s expected from statistics, and the tau(I) fit gives reduced chi-squared = 10.68, which the authors attribute to drift. For the weighted sampler, Eq. (4) P_j = lambda_j / sum(lambda_k) is only correct if all lambda_j are fixed for the duration of a draw. Under drift, each device's effective rate varies, so the realized selection probability is E[lambda_j / sum(lambda_k)] over the drift distribution, not lambda_j / sum(lambda_k). The discrepancy is first-order in the drift variance and can be substantial: a 10% RMS drift in one rate among n=2 devices changes selection probabilities by several percent, directly biasing the claimed 'n-sided die'. Similarly, the Metropolis-Hastings circuit's Eq. (3) acceptance probability becomes E[exp(-beta Delta E * lambda / lambda_0)] rather than exp(-beta Delta E), so detailed balance against the target Boltzmann distribution is not guaranteed. The paper provides no bound on drift magnitude, no compensation scheme, and no demonstration that the applications tolerate the observed drift. This is a correctness risk for the central proposal, not merely a device-engineering footnote: the single-shot exponential claim can survive (Fig. 1(c)) while the sampler-level claim fails.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a probabilistic delay cell built from a superparamagnetic magnetic tunnel junction (SMTJ): the device is biased into the parallel state, a current step puts it into the superparamagnetic regime, and the time of the first P-to-AP switching event is measured with on-board counter electronics. The central experimental claim is that these first-switch times are exponentially distributed with a current-tunable rate. After reporting CDF fits (Fig. 1(c), reduced chi-squared 1.92; Fig. 3), the paper documents device drift (Appendix A, Sec. IV) and then proposes temporal-computing circuits for Bernoulli decisions, a Metropolis-Hastings stepper (Eq. (3)), and a weighted random sampler based on the exponential-clocks method (Eq. (4)). The applications are presented as proposals rather than as measured end-to-end demonstrations.","tokens_in":19488,"tokens_out":8867,"duration_ms":88919,"significance":"If the exponential-switching claim and the stationarity issue can be settled, the probabilistic delay cell is a genuinely useful primitive: a single switching event yields an analog exponential sample, potentially avoiding the repeated Bernoulli trials and Bernoulli-factory overhead that state-encoded SMTJ sampling requires. The paper's strengths are its direct measurement of roughly 10,000 switching events with a reported goodness-of-fit, the clear circuit design for time-domain readout, the qualitatively demonstrated current tunability, and a candid discussion of the device's drift and of the non-physical parameters obtained from the tau(I) fit. Those same strengths make the missing quantitative drift analysis conspicuous, because the proposed applications require fixed rates during a draw.","major_comments":[{"comment":"The weighted random sampler is analyzed under the assumption that each probabilistic delay cell has a fixed rate lambda_j for the duration of a draw. This is contradicted by the drift documented in Appendix A and Sec. IV: the AP-state probability wanders over 205 s with a spread of 0.042 s versus the 0.015 s average statistical uncertainty in the bins, and the tau(I) fit in Fig. 4 has reduced chi-squared 10.68, which the authors attribute to drift. When lambda_j is time-varying, the realized first-arrival probability is E[lambda_j / sum_k lambda_k] over the drift distribution, not lambda_j / sum_k lambda_k, and the bias is first order in the rate variance. The manuscript gives no bound on this bias, no compensation scheme, and no demonstration that the proposed 'n-sided die' tolerates the observed drift. Please add a quantitative drift analysis, or explicitly restrict the application claim to runs in which stationarity is verified.","section":"Sec. V, Eq. (4); Appendix A; Sec. IV"},{"comment":"The Metropolis-Hastings circuit requires the stochastic delay to be exactly exponential with rate W beta. Under the same device drift, the effective rate is random, so the actual acceptance probability becomes an expectation over the time-varying rate distribution rather than exp(-Theta(beta Delta E_i)). Detailed balance against the target Boltzmann distribution is therefore not guaranteed for the circuit as described. The paper should quantify the drift-induced error in the acceptance probability (or simulate the Metropolis-Hastings step under the drift model of Appendix A) and state the maximum drift for which the sampler remains correct.","section":"Sec. V, Eq. (3); Sec. IV"},{"comment":"The quantitative tunability needed by Eq. (4) is not established by the tau(I) characterization. The fit to Eq. (2) has reduced chi-squared 10.68, yields non-physical values of tau_0, Delta E, and I_c, and the exponent alpha is not actually determined, as the footnote states. Since the applications set weights by current addressing, the paper needs at least a per-device calibration procedure (for example, measuring tau at the operating current immediately before a sampling run) or an explicit statement of how the weights are to be calibrated without relying on Eq. (2). As written, only qualitative tunability is demonstrated.","section":"Sec. IV, Eq. (2) and Fig. 4"},{"comment":"The long-time tail is part of the exponential claim, but the inset of Fig. 3 shows a deviation from the fit that the text attributes to 'device variability' without quantification. The applications compare exponential samples to deterministic thresholds and to one another, so tail behavior matters. Please quantify the long-time deviation (for example, with a two-component fit or a tail-specific test) and state how the proposed circuits tolerate it.","section":"Sec. IV, Fig. 3 inset"}],"minor_comments":[{"comment":"There are typographical issues: 'Kramer's' should be 'Kramers' in Sec. IV, and 'come existing methods' in the Introduction should be 'some existing methods.'","section":"Introduction and Sec. IV"},{"comment":"The axis labels of Fig. 4 appear in the manuscript as corrupted glyph sequences ('/uni000000...'); please ensure the final figure is legible.","section":"Fig. 4"},{"comment":"Report reduced chi-squared or another goodness-of-fit statistic for the three CDF fits in Fig. 3, as is done for Fig. 1(c), rather than relying on visual inspection.","section":"Fig. 3"},{"comment":"The residual 625 ns timing difference between the signal and reference paths is said to cause a systematic one-bit offset; specify whether the reported time values have been corrected for this offset.","section":"Sec. III A"},{"comment":"The Fokker-Planck/macrospin analysis mentioned in Sec. IV is not presented. Either include the governing equation and representative results, or cite the source for the statement that short-time corrections become unimportant for Delta E/kT > 15.","section":"Sec. IV"},{"comment":"The energy and throughput estimates in Sec. VI depend on assumed device parameters such as 1 V, 10 microampere, 100 ns, and eight time bins; label these explicitly as estimates and provide a model or reference for the per-bit energy comparison.","section":"Sec. VI"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is unusually honest about device limitations, which strengthens its credibility, but the same honesty reveals a load-bearing gap: the applications require stationary exponential rates, while the paper's own data show substantial drift. I do not see this as a reason for rejection, because the experimental primitive and the circuit proposals are separable and the drift issue can be addressed with a quantitative model or a restriction of the claims. However, the current version overreaches in Sec. V, so a major revision is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. The central experimental result is real: after a current step, first-switch times of their SMTJ follow an exponential CDF with a good fit (reduced chi-squared = 1.92 on ~10,000 events), and the probabilistic delay cell circuit is a clean, integrable way to digitize those times. That part deserves credit. The temporal encoding angle is genuinely new; prior work used SMTJs as Bernoulli sources, not as native exponential samplers. The paper is also honest about its limits: it reports drift in Appendix A, admits the tau(I) fit is poor (reduced chi-squared = 10.68) and returns non-physical parameters, and notes long-time deviations in Fig. 3. That transparency is to its credit.\n\nThe soft spot is real. The application circuits—Metropolis-Hastings stepper and weighted random sampler—assume a fixed rate parameter lambda per device. The paper's own drift data contradicts that: the AP-state probability wanders over 205 s with spread about three times the statistical error, and the tau(I) fit fails by a lot. For the weighted sampler, the selection probability is not lambda_j / sum(lambda_k) but E[lambda_j / sum(lambda_k)] over the drift distribution. That difference is first-order in the drift variance, not a small correction. For the MH stepper, the acceptance probability becomes E[exp(-beta Delta E * lambda / lambda_0)] and detailed balance against the target distribution is not guaranteed. The paper does not bound the drift, offer a compensation scheme, or test whether the applications tolerate it. The stress-test note is correct on this.\n\nBut these are proposal circuits, explicitly described as future work, not demonstrated hardware. The measurement claim stands. This is not a fatal flaw; it is an unaddressed requirement for the applications. A referee should ask the authors to either characterize drift-induced bias quantitatively or add the caveat that the sampler circuits assume drift-free (or drift-compensated) devices.\n\nWho is this for: people working on probabilistic computing with SMTJs, temporal/race logic, and hardware samplers. I'd bring it to a reading group. I would not cite it in the next year unless I was doing temporal sampling myself, but it's a legitimately useful reference for the exponential first-switch result. Serious thinker: yes—clear, honest, and the core claim is measured, not fitted to a target. Recommendation: send to peer review, with the drift issue as a required revision.","headline":"Solid measurement of exponential first-switch times with a clean circuit; the proposed samplers need a drift bound before their claims hold.","tokens_in":20106,"tokens_out":2750,"would_cite":true,"duration_ms":26100,"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":"Measuring the time until a magnetic tunnel junction switches gives one exponential random sample.","keywords":["superparamagnetic tunnel junction","temporal encoding","probabilistic delay cell","exponential distribution","Metropolis-Hastings stepper","weighted random sampling","exponential clocks","time-to-digital conversion"],"falsifier":"Run a long stationary-current experiment, split the pooled switching times into early and late halves, and compare the two best-fit rates; if the rates differ by more than their statistical uncertainties, the aggregate is a mixture of exponentials rather than one exponential. The paper's Appendix A protocol already provides a template: over 205 s the per-bin mean dwell times spread by 0.042 s against an average within-bin uncertainty of 0.015 s, so the same comparison on the fitted $\\lambda$ values would settle whether drift breaks the model.","tokens_in":19009,"feed_emoji":"🧲","tokens_out":9090,"duration_ms":75738,"temperature":0.7,"pith_summary":"Superparamagnetic magnetic tunnel junctions already switch randomly between resistance states, and their dwell times are known to be exponential. This paper turns that known behavior into a usable sampling primitive by measuring, with on-board electronics, the exact time between a current step and the first switching event. The resulting circuit element, called a probabilistic delay cell, produces one tunable exponential random value per switching event, and repeated experiments of about 10,000 events per setting show that the measured times follow the cumulative distribution $F(t) = 1 - e^{-\\lambda t}$. The authors argue this matters because exponential variates are the native ingredient of Metropolis-Hastings Monte Carlo and of weighted random sampling, where conventional approaches must approximate an exponential by combining many Bernoulli random bits. If the claim holds, each switching event replaces many bit-draws, potentially saving energy and latency in probabilistic and temporal computing.","feed_headline":"One magnetic flip now yields one exponential random sample","feed_subtitle":"By timing each magnetic flip, the circuit turns every switch into a tunable exponential sample.","key_machinery":"The load-bearing object is the probabilistic delay cell (PDC), a circuit that applies a current step to an SMTJ and converts the first switching time into a digital count. Its statistical identity is the exponential cumulative distribution $F(t) = 1 - e^{-\\lambda t}$, with the rate $\\lambda$ set by the current amplitude through $\\tau_P(I) = \\tau_0 \\exp[(\\Delta E/kT)(1 + I/I_c)^\\alpha]$. The companion identity, drawn from the exponential-clocks method, is that when several PDCs run concurrently, the probability that the $j$-th one fires first is $\\lambda_j / \\sum_i \\lambda_i$, which is what turns a race of timers into a weighted random sample. The hardware that carries these identities consists of the transconductance, hysteresis, latching, and counter stages shown in the paper's Fig. 2(a), with a 625 ns path mismatch that produces a known one-bit offset in the counts.","core_discovery":"The paper's central claim is that the delay until a superparamagnetic magnetic tunnel junction (SMTJ), initialized in the parallel state, switches to the antiparallel state under a current step is an exponential random variable whose rate $\\lambda$ is controlled by the step amplitude. It establishes this by building a probabilistic delay cell: a transconductance stage applies the current step, a programmable hysteresis stage cleans the switching edge, a set-reset latch captures the first rising transition, and a 16-bit counter measures the elapsed time in units of a 500 ns clock. Plots of the cumulative distribution at three current levels (918, 924, and 930 $\\mu$A) fit Eq. (1), $F(t) = 1 - e^{-\\lambda t}$, with reduced $\\chi^2 = 1.92$ for the main dataset, and the mean switching time $\\tau = 1/\\lambda$ varies exponentially with current as $\\tau_P(I) = \\tau_0 \\exp[(\\Delta E/kT)(1 + I/I_c)^\\alpha]$, giving nearly two orders of magnitude of tunability. On this basis, the paper proposes two temporal applications: a Metropolis-Hastings stepper in which a probabilistic delay cell racing a deterministic delay cell produces an acceptance bit with probability $e^{-\\Theta(\\beta\\Delta E_i)}$, and a weighted random sampler in which $n$ probabilistic delay cells acting as exponential clocks select index $j$ with probability $\\lambda_j / \\sum_i \\lambda_i$. The applications are proposed and analyzed, not yet built as demonstrated systems.","pith_inferences":["Beyond the paper, the same temporal-encoding scheme should transfer to any Poissonian switching device, such as single-photon avalanche diodes or oxide-memristor random telegraph noise sources, which the paper itself mentions; the PDC would then be a generic analog-to-exponential interface rather than an MTJ-specific circuit.","Beyond the paper, the measured drift implies that a single long sampling run should be treated as a mixture of exponentials; a direct test is to split a long stationary-current run in half and compare the two fitted rates, and if they differ, circuit-level calibration or magnetic-field stabilization (noted in Appendix A for newer devices) becomes necessary before the Metropolis-Hastings acceptance","Beyond the paper, the bit-depth of the temporal sample is set by clock speed and counter width, so increasing the counter resolution or shortening the clock period directly increases random bits per switching event without changing the device, a design lever that does not exist for state-encoded Bernoulli sampling."],"forward_implications":["Each switching event yields one sample from an exponential distribution, so algorithms that need exponential variates can get them without assembling many Bernoulli trials.","The rate parameter $\\lambda$ is tunable by the applied current, allowing post-fabrication adjustment of the sampling distribution over nearly two orders of magnitude in mean delay.","A Metropolis-Hastings acceptance step reduces to comparing the arrival times of a probabilistic and a deterministic delay cell, producing the acceptance probability $e^{-\\Theta(\\beta\\Delta E_i)}$.","A set of $n$ probabilistic delay cells with an OR gate implements the exponential-clocks method, drawing the index $j$ with probability $\\lambda_j / \\sum_i \\lambda_i$ in one racing step.","The timing measurement is quantized by the clock period (500 ns here), so the number of usable bits per switching event depends on how finely the operating window is divided."],"supporting_citations":[{"why":"This work establishes superparamagnetic tunnel junctions as low-energy stochastic devices with exponential dwell-time statistics.","marker":"[1]"},{"why":"This work documents nanosecond-scale random telegraph noise and the exponential switching statistics that the probabilistic delay cell samples.","marker":"[2]"},{"why":"This work supplies the exponential-clocks method that the weighted random sampler circuit is built on.","marker":"[8]"},{"why":"This work previously reported instability in the same device stack, supporting the paper's drift interpretation.","marker":"[14]"},{"why":"This work provides the macrospin Fokker-Planck treatment used to analyze short-time corrections to the exponential distribution.","marker":"[35]"},{"why":"This work gives micromagnetic simulations of the transient response to a current step, used to bound the time scales of non-exponential behavior.","marker":"[36]"},{"why":"This work introduces race-logic temporal computing primitives, including the inhibit gate used in the proposed sampler circuits.","marker":"[43]"},{"why":"This work provides the Metropolis-Hastings and simulated-annealing context that motivates accepting transitions with exponential probability.","marker":"[7]"}],"fun_headline_variants":["Exponential randomness from timed magnetic flips","Magnetic flip timing yields exponential random draws","One magnetic switch becomes a tunable exponential sample","Time to flip gives exponential random numbers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim stands on the switching rate $\\lambda$ staying constant while samples are being collected, and the paper's own Appendix A shows the device's mean dwell time drifting over a 205 s run by almost three times the statistical uncertainty, so a long run mixes several exponentials instead of drawing from one.","fun_headline_variants_meta":{"raw":{"variants":["Exponential randomness from timed magnetic flips","Magnetic flip timing yields exponential random draws","One magnetic switch becomes a tunable exponential sample","Time to flip gives exponential random numbers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00067,"raw_usage":{"total_tokens":3082,"prompt_tokens":1002,"completion_tokens":2080,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":2037}},"tokens_in":618,"tokens_out":2080,"duration_ms":13580,"temperature":1.0,"reasoning_tokens":2037,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:57:56.668469+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a long stationary-current experiment, split the pooled switching times into early and late halves, and compare the two best-fit rates; if the rates differ by more than their statistical uncertainties, the aggregate is a mixture of exponentials rather than one exponential. The paper's Appendix A protocol already provides a template: over 205 s the per-bin mean dwell times spread by 0.042 s against an average within-bin uncertainty of 0.015 s, so the same comparison on the fitted $\\lambda$ values would settle whether drift breaks the model.","supporting_citations":[{"cited_title":"exponential clocks","cited_arxiv_id":null,"evidence_quote":"This work establishes superparamagnetic tunnel junctions as low-energy stochastic devices with exponential dwell-time statistics."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This work documents nanosecond-scale random telegraph noise and the exponential switching statistics that the probabilistic delay cell samples."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This work previously reported instability in the same device stack, supporting the paper's drift interpretation."},{"cited_title":"Keane and G","cited_arxiv_id":null,"evidence_quote":"This work provides the macrospin Fokker-Planck treatment used to analyze short-time corrections to the exponential distribution."},{"cited_title":"Dughmi, J","cited_arxiv_id":null,"evidence_quote":"This work gives micromagnetic simulations of the transient response to a current step, used to bound the time scales of non-exponential behavior."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This work introduces race-logic temporal computing primitives, including the inhibit gate used in the proposed sampler circuits."},{"cited_title":"Niazi, S","cited_arxiv_id":null,"evidence_quote":"This work provides the Metropolis-Hastings and simulated-annealing context that motivates accepting transitions with exponential probability."}],"review_version":1}