{"id":"2e78d365-2deb-4c13-956d-9f7c33006a6f","arxiv_id":"2508.15895","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An attention-based neural network, trained only on measurement data, finds the measurement-induced phase transition as a learnability boundary and gives an empirical upper bound on the transition from a single initial state.","lead":"A machine-learning model called QuAN locates the measurement-induced phase transition directly from quantum measurement records, avoiding the usual need to post-select on measurement outcomes or classically simulate the full quantum state. If it holds up, this yields a practical, noise-tolerant probe for seeing this transition on near-term quantum hardware.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Phase-recognition classifier may detect measurement-rate cues rather than the physical MIPT signal.","rationale":"The reader's weakest_assumption identifies exactly this concern: the phase-recognition classifier may exploit trivial dataset differences such as measurement rate or outcome bias. My stress-test agrees that this is the most load-bearing threat because the abstract's promise of a scalable, post-selection-free MIPT detector depends on the classifier extracting a physical signal, not on a dataset shortcut. The full text is not available, but the abstract gives no evidence of a control for measurement-rate imbalance, no derivation of the upper bound, and no noise-model specification. These are not internal inconsistencies but unverified external validity claims. A concrete control experiment—matching measurement rates across classes and testing label-shuffled data—would settle whether the classifier's boundary is physical. The verdict remains CONDITIONAL because the full text might contain the missing controls, but the current abstract does not establish them. No ad hominem is intended; the critique is on the argument. I agree with the reader's assessment, so no change to the verdict is proposed.","tokens_in":1009,"tokens_out":1366,"duration_ms":17612,"concrete_test":"Run the phase-recognition experiment on a dataset where the measurement rate is exactly matched across the weak- and strong-monitoring classes—for example, fix the number of measurement events per trajectory and vary only the measurement strength (or outcome probabilities) while keeping trajectory length constant. If QuAN's accuracy still peaks at the known MIPT location (from exact simulation for small systems), the result is not a rate artifact. Additionally, train the same architecture on label-shuffled trajectories where the measurement rate distribution is preserved but the physical phase labels are scrambled; chance-level accuracy would indicate the classifier is not relying on trivial dataset statistics.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The abstract's central claim—that QuAN yields an upper bound on the MIPT from measurement data alone—rests on the phase-recognition task, where a classifier distinguishes weak- from strong-monitoring trajectories generated from a single initial state. The burden is to show the classifier's decision boundary tracks the entanglement-based MIPT rather than trivial statistical differences in the measurement records. In the setup described, the monitoring strength likely changes both the frequency of measurements and the distribution of Born probabilities per trajectory. A classifier could exploit the measurement rate (e.g., the number of measurement outcomes per trajectory, or the variance of outcome counts) as a shortcut, achieving high accuracy away from the true MIPT. The abstract provides no control experiment to rule this out. The claimed 'upper bound' also lacks a derivation in the abstract; without a formal argument that classifier accuracy or attention scores bound a known order parameter, the bound is only an empirical correlation for the specific Haar circuits tested. The interpretation that attention focuses on the early-time tail of Born probabilities is presented as 'reassuring,' but this is a post-hoc observation unless the model's decisions are causally shown to depend on that tail rather than on other record features.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a data-centric machine-learning approach, QuAN, to detect measurement-induced phase transitions (MIPTs) from classical measurement records, avoiding post-selection and classical simulation of the full state. Two tasks are studied: a “learnability” transition, where the network distinguishes two different initial states from measurement records, and a “phase recognition” task, where the network classifies weak- versus strong-monitoring data from a single initial state. The authors report that the learnability transition pinpoints a phase boundary consistent with exact results for Haar random circuits, and that the phase-recognition task yields an efficient and noise-tolerant upper bound on the MIPT. They also report an attention analysis suggesting that the network focuses on the early-time tail of Born probabilities, which they interpret as causally meaningful.","tokens_in":1219,"tokens_out":2642,"duration_ms":31715,"significance":"If the claims hold, this is a potentially valuable step toward experimentally accessible detection of MIPTs: the protocol uses only the classical measurement record, avoids post-selection, and does not require classical simulation. The external validation against exact results in the tested Haar-circuit family is a genuine strength, and the attention-based interpretation is a useful diagnostic. However, the central claim—that classifier accuracy or attention scores constitute an upper bound on the MIPT—is asserted rather than derived, and the phase-recognition setup is vulnerable to shortcut learning via trivial dataset statistics. The paper is therefore promising but needs substantial additional evidence and formalization before the main claim is established.","major_comments":[{"comment":"The central claim that QuAN provides an “upper bound” on the MIPT is not supported by a derivation. Classifier accuracy on a binary phase-classification task is by construction a measure of how separable the training records are under the chosen architecture; without a formal relation between that separability and the entanglement-based order parameter (e.g., a theorem, a scaling collapse, or an inequality linking accuracy to the entropic diagnostic), the statement is not a bound. The paper reports agreement with exact results in specific cases, but the word “bound” implies a guaranteed relation. I ask the authors to either provide a derivation or soften the claim to “empirical estimator” and support it with finite-size scaling and error bars.","section":"Abstract and phase-recognition section"},{"comment":"No control experiment rules out a trivial shortcut. In the described setup, weak versus strong monitoring changes the measurement rate and the distribution of Born probabilities per trajectory. A classifier could achieve high accuracy by counting measurements per trajectory, or by using the variance or tail of outcome counts, without learning anything about the MIPT. The authors should add (i) a baseline logistic regression or random forest on hand-crafted features such as record length, number of measurements, and outcome counts; (ii) an input ablation that removes measurement times or outcome counts while retaining Born probabilities; and (iii) label-shuffled controls to quantify the chance-level accuracy. Without these, the phase-recognition result is consistent with a dataset-separability artifact.","section":"Phase-recognition task description"},{"comment":"The statement that QuAN “paid special attention to the tail of the distribution of the Born probabilities at early times” is based on inspection of attention scores. This does not establish that the network’s decision depends causally on that tail. I recommend a quantitative test: mutate or permute the tail region of the Born-probability distribution and measure the change in classification accuracy; or train the network on records where the tail is removed and show the phase-recognition accuracy collapses. Without such a test, the attention interpretation should be described as a post-hoc correlation, not a mechanistic explanation.","section":"Attention analysis paragraph"},{"comment":"The phrase “minimal sample size” is used to motivate the phase-recognition task, but the paper does not define how this minimal sample size is chosen or how the results depend on it. If the minimal sample is determined by the network’s training curve, then the claim of sample efficiency is circular. Furthermore, the learnability-to-MIPT equivalence is only demonstrated for Haar random circuits with weak measurements; the paper should discuss and test whether this equivalence is expected to hold for other circuit ensembles (e.g., Clifford, Floquet, or circuits with conserved quantities), where the distinguishability of initial states from the record may transition at a different point or not at all. A concrete test on at least one additional circuit family would substantially increase confidence in the generality of the method.","section":"Minimal sample size and generality"}],"minor_comments":[{"comment":"The abstract states the method is “noise-tolerant” but no noise model or quantitative noise analysis is presented. Please specify the noise model (e.g., depolarizing, measurement errors) and show performance as a function of noise strength.","section":"Throughout"},{"comment":"The paper does not provide a complete table of hyperparameters (number of layers, heads, embedding dimension, learning rate, batch size, optimizer, training epochs, and the exact train/test split). Such a table is essential for reproducibility, especially because the architecture is central to the claim.","section":"Architecture and training details"},{"comment":"The terms “Born-distribution-level (inter-trajectory) attention” and “dynamical (temporal) attention” are introduced without precise definitions. Please define these mathematically, including how the attention weights are computed from the measurement records.","section":"Notation"},{"comment":"The reported phase-boundary locations should include error bars over random seeds and, where possible, a finite-size scaling collapse (e.g., plotting the crossing point as a function of system size and extrapolating to the thermodynamic limit). Currently the abstract’s claim of consistency with exact results is not quantitatively supported by the presented metrics.","section":"Figures"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses an important problem and the external validation against exact MIPT results is encouraging, but the central 'upper bound' claim is currently overstated. The phase-recognition shortcut issue is the main technical risk: the classifier may simply be counting measurement outcomes. The revision should focus on adding a control experiment and a derivation or careful qualification of the bound. I would be willing to look at a revised version that includes these elements."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the idea is worth taking seriously, but the abstract overclaims. What is actually new is the pairing of a learnability transition with a single-initial-state phase-recognition task, using inter-trajectory attention to target the tail of Born probabilities, and calling the classifier accuracy an upper bound on the MIPT. That combination is new to me, and the external check against exact results for Haar circuits is a good sign—it gives the learnability claim some independent footing. The paper should get full credit for aiming at a protocol that avoids post-selection and classical simulation.\n\nSoft spots: the 'upper bound' is asserted, not derived. The abstract gives no error bars, no finite-size scaling, and no description of the noise model behind 'noise-tolerant.' More concretely, the phase-recognition classifier may be solving an easier problem: in the setup, weak versus strong monitoring changes the number of measurement outcomes per trajectory and the variance of Born probabilities, so the network could separate the classes by counting measurements rather than by learning the physical phase boundary. The stress-test note is right that this needs a control experiment. The attention-to-the-tail interpretation is also post-hoc unless the paper shows the model's decisions causally depend on that tail. These may be addressed in the full text, but from the abstract they are open.\n\nNet: this is a conditional accept for peer review. A referee should demand (1) a derivation or formal statement of the upper-bound claim, (2) controls that rule out measurement-rate shortcuts, and (3) scaling with system size and noise strength. If the full text delivers those, it could be a genuinely useful experimental protocol. Otherwise, it is an interesting correlation study. Either way, it deserves a serious referee.","headline":"A plausible protocol with an over-stated upper-bound claim; worth refereeing, but demand controls.","tokens_in":1764,"tokens_out":2268,"would_cite":false,"duration_ms":24665,"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 quantum attention network can locate the measurement-induced phase transition directly from classical measurement records, without post-selecting trajectories or simulating the full quantum state.","keywords":["measurement-induced phase transition","quantum attention network","monitored quantum circuits","Born probabilities","learnability transition","weak measurements","phase recognition","near-term quantum hardware"],"falsifier":"Apply the same QuAN phase-recognition pipeline to measurement records from a non-monitored or fully scrambled circuit that has no entanglement phase transition; if the classifier still produces a sharp boundary between 'weak' and 'strong' labels, the reported upper bound is an artifact of the training setup rather than a signature of MIPT.","tokens_in":848,"feed_emoji":"⚛️","tokens_out":5951,"duration_ms":61519,"temperature":0.7,"pith_summary":"The paper aims to show that measurement-induced phase transitions (MIPTs) can be detected from purely classical measurement records, using a neural network architecture called a Quantum Attention Network (QuAN). It first demonstrates that a \"learnability\" transition—the monitoring strength at which two initial states become distinguishable from the readout—matches the known entanglement transition for Haar-random circuits. It then shows that a simpler \"phase recognition\" task, classifying weak- vs strong-monitoring data from a single initial state, gives an efficient, noise-tolerant upper bound on the MIPT, and that the network's attention falls on the tail of the Born-probability distribution at early times. If correct, this removes the experimental bottleneck of post-selection and classical simulation, making MIPT observation feasible on near-term hardware.","feed_headline":"Neural net detects quantum phase transition without state simulation","feed_subtitle":"Measurement-only records reveal the entanglement transition, sidestepping post-selection and classical simulation.","key_machinery":"The central object is the Quantum Attention Network (QuAN), an attention-based machine learning model that processes measurement records along two axes: inter-trajectory (across many stochastic measurement outcomes) and temporal (along the circuit time steps). The \"learnability\" transition is used as a proxy: at the MIPT, the measurement record becomes information-theoretically sufficient to distinguish two initial states. In the phase-recognition task, the model's attention weights on Born probability distributions supply the decision boundary.","core_discovery":"The central discovery is that the entanglement-based measurement-induced phase transition in random monitored circuits leaves a classical fingerprint in the measurement record itself, and an attention-based network can read it. In the phase-recognition setting, QuAN is trained on single-shot measurement outcomes from a single initial state under both weak and strong monitoring; its classification boundary tracks the MIPT and provides an upper bound consistent with exact results. The network's inter-trajectory attention scores reveal that it concentrates on the early-time tail of the Born probability distribution, indicating that the distinguishing signal is the rarity of high- or low-probabi","pith_inferences":["If the coincidence between the learnability transition and the entanglement MIPT holds beyond Haar random circuits, \"learnability\" could become a general operational definition of measurement-induced phase structure, applicable where entanglement measures are inaccessible.","The emphasis on the tail of Born probabilities suggests the transition may be governed by rare measurement outcomes; a natural testable extension is to check whether importance sampling or large-deviation statistics of the record sharpens the upper bound.","The upper-bound nature implies the phase-recognition classifier may detect precursor signatures even where the true MIPT is absent; comparing its boundary with direct entanglement computation on non-Haar circuits would clarify how tight the bound is.","The temporal attention component likely encodes memory effects; one could test whether truncating the record length changes the estimated critical point, revealing the relevant time scale of the transition."],"forward_implications":["MIPTs can be observed on near-term quantum hardware by recording measurement outcomes and feeding them to QuAN, bypassing post-selection and classical simulation of the full state.","The classifier's decision boundary and attention scores can be used to estimate the critical measurement rate, in agreement with exact entanglement calculations for the tested Haar random circuits.","The method is sample-efficient and noise-tolerant, meaning small datasets and imperfect measurement records may still yield a reliable upper bound on the transition.","Attention statistics give an interpretable indicator: the model focuses on the tail of early-time Born probability distributions, suggesting where in the measurement record the physical signal lives.","The same data-centric approach may extend to other phase transitions in monitored quantum dynamics."],"supporting_citations":[],"fun_headline_variants":["Attention network spots quantum phase transition from measurements alone","Quantum phase transition revealed by neural net on measurement records","No simulation needed: attention model finds quantum phase boundary","Measurement-only neural net detects entanglement transition","QuAN reads measurement history to pinpoint quantum phase transition"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The central assumption is that for the tested circuit family, the monitoring strength at which measurement records distinguish two initial states (and the signal the classifier exploits) is the same as the entanglement-based transition, rather than a coincidental or trivial correlate.","fun_headline_variants_meta":{"raw":{"variants":["Attention network spots quantum phase transition from measurements alone","Quantum phase transition revealed by neural net on measurement records","No simulation needed: attention model finds quantum phase boundary","Measurement-only neural net detects entanglement transition","QuAN reads measurement history to pinpoint quantum phase transition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000395,"raw_usage":{"total_tokens":1929,"prompt_tokens":787,"completion_tokens":1142,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":1070}},"tokens_in":531,"tokens_out":1142,"duration_ms":10153,"temperature":1.0,"reasoning_tokens":1070,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T17:41:18.066873+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply the same QuAN phase-recognition pipeline to measurement records from a non-monitored or fully scrambled circuit that has no entanglement phase transition; if the classifier still produces a sharp boundary between 'weak' and 'strong' labels, the reported upper bound is an artifact of the training setup rather than a signature of MIPT.","supporting_citations":[],"review_version":1}