{"id":"eef1e3af-bfe0-43cd-b5c7-ee2134de1513","arxiv_id":"2504.12705","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A lead-free bismuth-halide memristor is reported as a physical reservoir, with claimed MNIST accuracy of 82.26%, but the classification equations omit the reservoir outputs.","lead":"This paper reports memristive switching in 7-methylquinolinium iodobismuthate and claims the material can act as a physical reservoir computer, classifying MNIST digits at 82.26% and a spoken digit at 82%. If the reservoir results hold up, the work adds a lead-free candidate material to neuromorphic hardware, though the paper's own equations do not show the reservoir outputs being used.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MNIST readout equations operate on raw pixels, not the 15 reservoir outputs; the 82.26% reservoir-classification claim is therefore not demonstrated.","rationale":"The reader's verdict of REJECT is correct, but the most load-bearing weakness is not the transfer of two-electrode memristor behavior to the 16-electrode chip. The deeper problem is that the MNIST readout, as written in Eqs. (16)–(17), never uses the 15 reservoir outputs at all. u is defined as the 784-element image vector, W is 784×10, and z = uW; this is a linear perceptron on raw pixels. No equation connects the device outputs to the classifier, so the abstract's claim that the device achieves 82.26% 'has been demonstrated' is unsupported by the presented mathematics. The reader's rationale did flag the raw-pixel readout as a load-bearing problem, but the formal 'weakest_assumption' field focused on device transfer; hence partial agreement. The Table 2 entry of 97% for a different Au/7-MeqBiI3/Au geometry is an additional internal inconsistency that independently undermines the report. The crystallography, optical characterization, and two-electrode memristor data (I-V hysteresis, retention, potentiation/depression, STDP) appear to be credible and are independently supported by CCDC deposits and standard characterization; those parts of the paper are not the problem. The rejection should stand because the headline computational claim—physical reservoir MNIST classification—is not demonstrated as written. The authors should either correct the readout equations to show the 15 reservoir outputs being used, provide the actual training pipeline and data, and reconcile the 97% entry, or reframe the paper as a materials and memristor-characterization study without the reservoir classification claim.","tokens_in":35763,"tokens_out":2823,"duration_ms":31694,"concrete_test":"Reconstruct the training pipeline and run two classifiers on the same 8000-image split: (A) softmax or ridge regression on the raw 784-pixel vectors exactly as in Eqs. (16)–(17), and (B) the same readout trained on the 15 measured reservoir outputs (O1–O15) after serialized images are fed to the device. If pipeline A alone reproduces the reported 82.26%, the reservoir contributes nothing and the central claim fails. Also verify whether any recorded reservoir output appears in the loss or weight update; if not, the experiment is a raw-pixel perceptron with a device attached. In addition, re-run the raw-pixel linear baseline on standard MNIST with the identical train/test split; if that baseline exceeds 82.26%, the reported 74% baseline is anomalous and the claimed reservoir-induced improvement is spurious.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that the 16-electrode 7-MeqBiI3 device classifies MNIST at 82.26% as a physical reservoir—is not supported by the paper's own readout equations. In Section 3.3, Eq. (16) defines z = u × W with u explicitly the 784-element serialized image vector and W of dimension 784×10; Eq. (17) sums u_k·W_{k,j}. No term involving the 15 reservoir outputs O1–O15 appears in the classifier. The surrounding text says the perceptron 'processes the 15 outputs,' but the mathematics defines a linear classifier on raw pixels only. The claimed improvement from 74% (raw perceptron) to 82.26% therefore cannot be attributed to the reservoir; as written, 82.26% is simply a softmax/linear classifier on pixel intensities. Moreover, the 74% raw-pixel baseline is far below typical softmax regression accuracy on MNIST (~92%), indicating an undocumented split, preprocessing, or training protocol. Table 2 adds a further inconsistency: it lists 'Au/7-MeqBiI3/Au 97 PRC This work'—a 97% accuracy, a different electrode stack (Au/Au versus Cu/ITO and the Al-on-SiO2/Si reservoir chip), and a geometry not described in the experimental section. Unless Eqs. (16)–(17) are corrected to show readout on the 15 reservoir outputs, and the accuracy numbers and electrode geometries are reconciled, the reservoir-computing claim is unverifiable. This concern is more load-bearing than the two-electrode-to-16-electrode transfer question, because if the readout never uses the reservoir states, the transfer issue is moot.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports the synthesis and characterization of three 7-methylquinolinium halobismuthates (I, Br, Cl), including crystal structures, optical band gaps, and DFT calculations. It then describes ITO/7-MeqBiI3/Cu memristive devices and characterizes their I-V hysteresis, retention, ON/OFF ratio, temperature-dependent conductivity, potentiation/depression, and STDP. To assess physical reservoir computing, the authors built a 16-electrode device with one input and 15 outputs and report benchmark tasks including waveform generation, NARMA-2, memory capacity, noise analysis, MNIST digit classification claimed at 82.26% accuracy, and speaker-dependent voice recognition at 82%. The central claim is that this material functions as a physical reservoir for pattern classification.","tokens_in":36126,"tokens_out":8024,"duration_ms":80721,"significance":"The materials-science portion is careful: the crystallographic work covers four structures with deposited CCDC numbers, the Tauc gap/DFT comparison is consistent, and the two-electrode memristive data (I-V, retention, STDP) are plausible and generally support the authors' interpretation of interfacial switching. If the reservoir-computing claim were demonstrated, the paper would be a notable contribution to lead-free physical reservoir computing. However, the central classification claim is not supported by the manuscript's own readout equations, and the reported benchmark numbers contain internal inconsistencies. The structural and memristive portions could be valuable, but the reservoir-computing advance as presented is not established.","major_comments":[{"comment":"The readout is defined as z = u × W with u the 784-element serialized image vector and W of dimension 784×10; the 15 reservoir outputs O1–O15 never enter the classifier. The sentence stating that the perceptron processes the 15 outputs is therefore contradicted by the mathematics, and the reported 82.26% accuracy cannot be attributed to the reservoir. The authors must reformulate the readout as a function of the reservoir state vector (e.g., z = W_O O) and recompute the accuracy and the comparison with the unprocessed baseline.","section":"Section 3.3, Eqs. (16)-(17)"},{"comment":"Table 2 includes the row 'Au/7-MeqBiI3/Au 97 PRC This work', reporting 97% accuracy with a device stack that appears nowhere in the experimental description; the abstract and text report 82.26% on a 16-electrode chip described either as gold or aluminum electrodes on SiO2/Si. This internal inconsistency cannot be resolved by a minor edit; the authors must provide a single consistent set of device geometry, electrode metal, and accuracy values for all reported benchmark results.","section":"Section 3.3 and Table 2"},{"comment":"The memristive mechanism and nonlinear dynamics are characterized only on the two-electrode Cu/7-MeqBiI3/ITO sandwich, while the reservoir experiments use a different device: SiO2/Si substrate, stated Al or Au electrodes, drop-cast material across a 100 µm gap. No I-V, retention, or plasticity data are supplied for the reservoir chip, so the key assumption that the reservoir inherits the sandwich-device dynamics is unverified. Provide basic electrical characterization of the actual 16-electrode device or justify explicitly why the two geometries share the same mechanism.","section":"Sections 3.2 vs. 3.3"},{"comment":"The 74% baseline accuracy for a linear perceptron trained directly on the 784-dimensional pixel vectors is far below reported softmax regression on MNIST (typically ~92% on the full training set). Because the reported 82.26% result is itself defined as a linear classifier on the same raw pixels, the claimed 'ca. 20% increase' is not meaningful unless the preprocessing, data split, and training protocol are fully disclosed and the baseline is recomputed on the identical split and serialization.","section":"Section 3.3, baseline comparison"}],"minor_comments":[{"comment":"The softmax definition appears garbled: the numerator should read exp(z_j) and the denominator should be the sum over k of exp(z_k); the printed form seems to omit the exponentials.","section":"Equation (19)"},{"comment":"The memory capacity formula has a misplaced division dot; the denominator should be the product of the variances (or standard deviations) of the actual and predicted inputs, not the covariance squared divided by a dot-separated variance expression.","section":"Equation (15)"},{"comment":"The abstract states that the 16-electrode device uses gold electrodes, while Section 2.2 states that the substrate is SiO2/Si with Al electrodes; please specify the correct metallization and ensure consistency throughout the text.","section":"Abstract and Section 2.2"},{"comment":"Figure 15 shows accuracy versus epoch for the voice-recognition task, and the text reports '~84%' after 2000 epochs while the abstract and conclusion cite 82%; specify whether these are training or test accuracies and provide the final held-out test accuracy.","section":"Section 3.3 and Figure 15"},{"comment":"The NMSE is defined twice with different variable names for the target and predicted signals; unify the definitions or reference the first equation to avoid confusion.","section":"Equations (7) and (13)"},{"comment":"The phrase 'Cauchy products' should read 'matrix-vector product' or 'dot product'; the current wording is not standard terminology for the operation in Eq. (17).","section":"Section 3.3, text near Eq. (17)"}],"recommendation":"reject","confidential_remarks":"The manuscript contains a strong materials-science component, but the central reservoir-computing claim is not supported by the readout equations: Eqs. (16)-(17) describe a linear/softmax classifier on raw pixels, not on the 15 reservoir outputs. The additional inconsistency in Table 2 (97% Au/7-MeqBiI3/Au versus the 82.26% result) and the discrepancy between abstract and experimental text regarding electrode metal suggest that the benchmark section needs fundamental reworking. If the authors can supply a corrected readout based on the actual reservoir states and consistent accuracy numbers, a resubmission could be considered; as written, the paper is not acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you need to know: this paper's materials work is credible, but its headline claim—82.26% MNIST via a physical reservoir—is not supported as written. The readout equations (16)-(17) define z = u × W using the raw 784-pixel vector u, so the 15 reservoir outputs never enter the classifier. The text says the perceptron processes the 15 outputs, but the math doesn't. Also, Table 2 lists a 97% accuracy for \"Au/7-MeqBiI3/Au,\" which doesn't match the Cu/ITO or Al-on-SiO2 devices described anywhere in the experimental section. These are load-bearing inconsistencies.\n\nWhat's genuinely new: four crystal structures deposited in CCDC for 7-methylquinolinium halobismuthates (I, Br, Cl, and a high-temperature iodide phase), with careful Hirshfeld analysis, Tauc gaps, and DFT band structures that follow the halide trend. The Cu/7-MeqBiI3/ITO memristor characterization is solid: I-V hysteresis, retention over 11 hours, 1000-cycle endurance, low ON/OFF ratio consistent with interface-type switching, activation energies around 0.3 eV, and STDP plus potentiation/depression data. That part deserves publication somewhere.\n\nThe soft spots are concentrated in the reservoir-computing section. Beyond the equations, the 16-electrode chip is a different device from the characterized one: SiO2/Si substrate, drop-cast film, and electrode metals that are inconsistent (gold in the abstract, aluminum in the text, copper/ITO elsewhere). No I-V, retention, or plasticity data are shown for that chip. The baseline \"74% raw perceptron\" is suspiciously low—softmax regression on raw MNIST typically gets ~92%—so the training protocol is either poorly described or the comparison is misreported. The voice classification at 82% is plausible but not independently verified here.\n\nI think the correct read is: a materials-and-memristor paper that overreached on a computational claim. The fix is straightforward: either rewrite the readout to use the 15 reservoir outputs with a proper 15×10 weight matrix and give actual training details, or drop the classification claim and keep the waveform/NARMA/memory benchmarks as demonstrations of nonlinearity. If the classification is corrected, the accuracy is still mediocre compared to other physical reservoirs (90-98% in Table 2), but the materials contribution is real.\n\nThis paper is for researchers working on lead-free halide memristors and physical reservoir computing. It deserves a serious referee: the crystallography and memristor data should be reviewed, and the reservoir claim can be fixed or honestly reframed. I would accept it for peer review.","headline":"Good crystallography and memristor characterization undermined by a reservoir-computing section whose equations never use the reservoir outputs.","tokens_in":36713,"tokens_out":1675,"would_cite":true,"duration_ms":17128,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single-input iodobismuthate memristor chip classifies handwritten digits with 82.26% accuracy and a spoken digit with 82%, acting as a physical reservoir.","keywords":["iodobismuthate","7-methylquinolinium","memristor","physical reservoir computing","neuromorphic electronics","MNIST digit classification","voice classification","synaptic plasticity"],"falsifier":"Run current–voltage cycling, retention, and STDP measurements directly on the 16-electrode SiO2/Si reservoir chip: the central claim collapses if its outputs show no pinched hysteresis, no voltage-dependent harmonic generation, and no fading memory, since then the reported digit and voice accuracies cannot be attributed to the material's memristive dynamics.","tokens_in":35560,"feed_emoji":"🧠","tokens_out":11475,"duration_ms":104204,"temperature":0.7,"pith_summary":"This paper sets out to show that 7-methylquinolinium iodobismuthate (7-MeqBiI3), a lead-free hybrid halobismuthate, can serve as the active material of a physical reservoir computer—a nonlinear physical system whose dynamics map inputs into high-dimensional outputs read by a simple trained layer. In two-terminal Cu/7-MeqBiI3/ITO devices the authors find memristive behavior (pinched hysteresis in current–voltage scans), a low ON/OFF ratio, temperature-activated mixed ionic-electronic conduction, potentiation/depression, and spike-timing-dependent plasticity. They then assemble a 16-electrode chip with one input and fifteen outputs and report that it classifies handwritten digits with 82.26% accuracy and the spoken digit 2 from six speakers with 82% accuracy, outperforming the same linear readout trained without the reservoir (74%). Supporting benchmarks include waveform generation, a NARMA-2 normalized mean squared error of 0.097, and a memory capacity of about 35 time steps. If correct, this means a single-input physical reservoir with only a few hundred trained weights can approach neural-network-level image classification.","feed_headline":"Iodobismuthate reservoir chip hits 82% on handwritten digits","feed_subtitle":"A one-input, 15-output physical reservoir classifies image and voice data through material nonlinearity, not a trained network.","key_machinery":"The load-bearing mechanism is Schottky barrier modulation: at the Cu/7-MeqBiI3 interface, trapping and detrapping of electrons in metal-induced gap states changes the barrier height, giving pinched hysteresis, a rectification factor near 2, an ON/OFF ratio below 2, and switching that is nearly independent of scan rate. The junction's equivalent circuit is a bulk resistor in series with a parallel diode, capacitor, and two-state memristive resistor, so each electrode pair acts as a nonlinear, partly capacitive element. This makes the 16-electrode chip a network of such junctions whose outputs are diverse, nonlinear transforms of the single input, and the readout is trained once on the 15 outputs.","core_discovery":"The central claim is that 7-MeqBiI3 memristive junctions switch by modulating the Schottky barrier at the copper contact through charge trapping and detrapping at metal-induced gap states, not by filament formation, and that this interface mechanism, together with a partly capacitive response and a small ionic contribution, creates the nonlinear, voltage-dependent dynamics needed for reservoir computing. The paper reports that the 16-electrode reservoir—one input, fifteen output electrodes on silicon with a drop-cast layer of the compound—maps serialized 784-pixel images and MFCC voice features into higher-dimensional output, from which a trained linear perceptron achieves 82.26% digit accuracy and 82% classification of digit 2 across six speakers. Supporting benchmarks include waveform reconstruction, a NARMA-2 normalized mean squared error of 0.097, a total memory capacity of about 35 time steps, and $1/f^{\\beta}$ noise with $\\beta$ rising to about 1.5 under bias.","pith_inferences":["A decisive next test would be to characterize the 16-electrode chip directly; if the drop-cast layer does not reproduce the two-terminal device's hysteresis and STDP, the classification gains might come from electrode contacts or readout bias rather than from the material's memristive dynamics.","The observed phase reversal and capacitive transients suggest the chip behaves like a stochastic network of memristors and capacitors, so temporal tasks beyond static images—such as continuous speech or video frames—may be within reach if memory depth can be increased.","Varying layer thickness, grain size, or electrode gap could trade memory capacity (about 35 steps here) against speed, which would determine whether serialized higher-resolution images can be processed with one input.","If the single-input serialization idea scales, energy consumption may be lower than parallel neural hardware because most computation happens in the material itself, with minimal trained weights."],"forward_implications":["The same device can classify both image and voice data from a single serialized input, so physical reservoirs of this type do not need a crossbar array or per-input synaptic training.","Because the reservoir performs the nonlinear feature extraction, the trainable readout needs on the order of 100–200 weights, far fewer than a conventional network for the same task, at the cost of slower serial processing.","The reported accuracy gain over the linear perceptron (82.26% versus 74%) implies that the material's nonlinearity and memory, not the readout alone, carry the classification.","The measured memory capacity of about 35 time steps and fading memory define the usable input length: tasks needing longer context would require a deeper or slower reservoir.","Halide substitution tunes the optical band gap from 2.09 eV (iodide) to 3.19 eV (chloride), so related compounds could extend reservoir operation into different spectral ranges."],"supporting_citations":[{"why":"Supplies the handwritten-digit database and the 28×28 input format used for the classification task.","marker":"[97]"},{"why":"Provides the Free Spoken Digit Dataset of six speakers' utterances used for voice classification.","marker":"[111]"},{"why":"Gives the MFCC feature-extraction and reservoir-readout procedure for voice recognition.","marker":"[112]"},{"why":"Introduces the echo-state/reservoir training principle that the readout perceptron relies on.","marker":"[37]"},{"why":"Demonstrates a single-memristor physical reservoir and is the experimental baseline for amplitude-sensitive reservoir behavior.","marker":"[90]"},{"why":"Extends the single-memristor reservoir to a multivalued SnI4 memristor, the direct precedent for this work's architecture.","marker":"[91]"},{"why":"Shows bismuth-based memristors emulating leaky integrate-and-fire and STDP, grounding the plasticity claims.","marker":"[31]"},{"why":"Characterizes the structural and electronic properties of the pyridinium iodobismuthate family that this material extends.","marker":"[21]"},{"why":"Provides the practical reservoir-computing protocol used for waveform generation, NARMA-2, and memory-capacity tests.","marker":"[2]"}],"fun_headline_variants":["Iodobismuthate reservoir hits 82% on digits","Interface switching, not filaments, drives memristor reservoir","Single-input, 15-output reservoir classifies digits and voice","Charge trapping at Schottky barrier gives reservoir nonlinearity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 16-electrode reservoir chip is assumed to share the memristive and plastic behavior measured only on the two-electrode Cu/7-MeqBiI3/ITO device, even though the chip uses a different substrate, a different stated electrode metal (gold in the abstract, aluminium in the methods), a different deposition method, and no device-level electrical characterization of its own.","fun_headline_variants_meta":{"raw":{"variants":["Iodobismuthate reservoir hits 82% on digits","Interface switching, not filaments, drives memristor reservoir","Single-input, 15-output reservoir classifies digits and voice","Charge trapping at Schottky barrier gives reservoir nonlinearity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000253,"raw_usage":{"total_tokens":1632,"prompt_tokens":1080,"completion_tokens":552,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":696,"completion_tokens_details":{"reasoning_tokens":482}},"tokens_in":696,"tokens_out":552,"duration_ms":5931,"temperature":1.0,"reasoning_tokens":482,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:24:40.383076+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run current–voltage cycling, retention, and STDP measurements directly on the 16-electrode SiO2/Si reservoir chip: the central claim collapses if its outputs show no pinched hysteresis, no voltage-dependent harmonic generation, and no fading memory, since then the reported digit and voice accuracies cannot be attributed to the material's memristive dynamics.","supporting_citations":[{"cited_title":"Gradient-based learning applied to document recogni�on","cited_arxiv_id":null,"evidence_quote":"Supplies the handwritten-digit database and the 28×28 input format used for the classification task."},{"cited_title":"Mel- frequency cepstral coefficients feature extracted voice recogni�on task using atomic switch Ag/Ag<sub>2</sub>S device -based �me -delayed reservoir compu�ng","cited_arxiv_id":null,"evidence_quote":"Gives the MFCC feature-extraction and reservoir-readout procedure for voice recognition."},{"cited_title":"Memristor in a Reservoir System - Experimental Evidence for High-Level Compu�ng and Neuromorphic Behavior of PbI2","cited_arxiv_id":null,"evidence_quote":"Demonstrates a single-memristor physical reservoir and is the experimental baseline for amplitude-sensitive reservoir behavior."},{"cited_title":"Neuromorphic Applica�ons of a Mul�valued [SnI4{(C6H5)2SO}2] Memristor Incorporated in the Echo State Machine","cited_arxiv_id":null,"evidence_quote":"Extends the single-memristor reservoir to a multivalued SnI4 memristor, the direct precedent for this work's architecture."}],"review_version":1}