{"id":"a56f8922-80a1-4e2c-8540-6d193cea059f","arxiv_id":"2607.11273","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":5.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A fixed-protocol amortized MPS estimator conditioned on informative local Paulis recovers high-fidelity states with few measurements, beating prior-only and shuffled controls, with conformal uncertainty.","lead":"The paper trains a fixed-protocol neural estimator that maps a carefully chosen set of local Pauli measurements to matrix-product-state cores, recovering high-fidelity quantum states with few shots. A prior-only control shows that measurement design, not just a generative prior, is what turns memorization into genuine tomography, with conformal intervals and a small IBM hardware check.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Abstract-only review leaves the decisive measurement-use claim (local Paulis vs prior-only/shuffled) unverifiable; no internal contradiction is visible, so the Reader's CONDITIONAL is already correctly calibrated.","rationale":"The Reader already isolates the load-bearing assumption (sufficient concentration on a low-χ MPS manifold so that a fixed local Pauli design plus a once-trained amortized map recovers the state) and correctly flags that the abstract's own k=0 control shows the claim collapses without that concentration. No stronger internal flaw is visible from the abstract alone: the gauge-invariant fidelity loss, the deliberate non-use of a set encoder, the conformal recalibration, and the hardware loop are all presented as supporting evidence rather than as hidden assumptions that break the argument. Because the full text is unavailable, any deeper critique of training details, hyper-parameters, or statistical significance would be speculative. The appropriate posture is therefore to leave the Reader's CONDITIONAL / LOW-confidence verdict untouched and to wait for the concrete control recomputation once artifacts exist.","tokens_in":2174,"tokens_out":494,"duration_ms":4480,"concrete_test":"When the full paper and code appear, recompute the main fidelity table under three fixed conditions: (i) the reported informative local Pauli design, (ii) the identical design with measurement outcomes randomly shuffled across states, and (iii) the pure k=0 prior. If the local-Pauli advantage over both (ii) and (iii) remains ≥+0.3 fidelity at the reported n/χ points, the measurement-use claim stands; otherwise the efficiency is largely memorization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"No significant objection identified beyond the Reader's own weakest_assumption and the abstract-only limitation. The paper's central claim is carefully hedged by its own k=0 prior-only and shuffled-measurement controls; those controls are the right ones for distinguishing genuine measurement efficiency from family memorization. Because the full text, methods, ablations, and artifacts are unavailable, the numerical claims (≈0.95 fidelity, +0.59 gain, ≈90% conformal coverage, IBM 0.97) cannot be stress-tested further. The abstract itself does not contain an internal inconsistency that would overturn the claim if the reported numbers hold.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript studies sample-efficient tomography of states concentrated on a low-bond-dimension MPS manifold. It contrasts a generative prior with measurement-guided posterior inference (Approach A) against a fixed-protocol amortized MPS estimator (Approach B, the main proposal) trained once with a gauge-invariant fidelity loss on a shared MPS-core parameterization. Conditioning on an informative local Pauli set rather than random strings is reported to raise fidelity to ≈0.95 (up to +0.59 over a k=0 prior-only baseline) and to pass a shuffled-measurement control. A dropout ensemble with conformal recalibration is claimed to yield ≈90% coverage intervals, including for unmeasured observables; quality is reported to hold at n=10 and χ=4, with a closed hardware loop on IBM (five states at fidelity 0.97).","tokens_in":2371,"tokens_out":1058,"duration_ms":24517,"significance":"If the reported controls and numbers hold, the work would supply a practical once-trained amortized route to MPS tomography that explicitly separates measurement use from family memorization—the right experimental standard for this literature. Creditable elements include the gauge-invariant fidelity loss, the deliberately simple (non-set-encoder) architecture, conformal predictive intervals for unmeasured observables, polynomial native contraction, and a hardware closing loop. The abstract itself hedges the scope to concentrated low-χ families via the k=0 control, which is a strength of the framing rather than a hidden caveat.","major_comments":[{"comment":"The load-bearing claim that the estimator genuinely uses measurements (rather than memorizing the training family) rests on the k=0 prior-only gain (up to +0.59) and the shuffled-measurement control. The full manuscript must report these ablations with error bars, seed counts, train/test splits, and an explicit shuffle protocol (which labels are permuted; whether local structure is preserved). Without those tables the decisive measurement-efficiency claim cannot be assessed.","section":"Abstract (k=0 and shuffled controls)"},{"comment":"The abstract states that on concentrated families a prior is already near-optimal and that local RDMs determine a χ-MPS. The efficiency claim therefore applies only under concentration on a low-χ manifold. The manuscript must quantify that concentration (core distribution, bond-dimension spectrum, or distance to the training family) and state failure modes when it is violated, so the scope is falsifiable.","section":"Abstract (scope / concentration assumption)"},{"comment":"The ≈90% conformal coverage claim, including for never-measured observables, is central to the uncertainty contribution. The full text must specify the conformal recipe (split vs. CV+, nonconformity score, dropout rate, recalibration set), report coverage versus n and χ, and compare to any available shot-based or bootstrap baseline. Absent that, the coverage figure remains an unchecked numerical assertion.","section":"Abstract (conformal intervals)"},{"comment":"The IBM result (five states at 0.97 from hardware-measured Paulis) closes the loop but is a small sample. The manuscript must document device, shot budget, mitigation, relation of the five states to the training distribution, and whether the same fixed local Pauli protocol was used without re-optimization; confidence intervals or a larger set are needed to support a hardware-generalization claim.","section":"Abstract (IBM hardware loop)"}],"minor_comments":[{"comment":"Notation k=0 for the prior-only control is used without a parenthetical definition; a brief gloss (no measurements / prior mean) would help abstract readers.","section":"Abstract"},{"comment":"The claim that 'a plain MLP matches' a permutation-invariant set encoder is intriguing but underspecified; the full text should name the architectures and report the head-to-head numbers.","section":"Abstract"},{"comment":"Point estimates (0.95, 0.90 at n=10, 0.88 at χ=4, 0.97 on IBM) should be accompanied by standard errors or IQRs in the results tables.","section":"Abstract (reported fidelities)"}],"recommendation":"major_revision","confidential_remarks":"Abstract-only assessment: full text, methods, ablations, and artifacts were unavailable. The abstract is carefully hedged and the control design (k=0, shuffle) is the correct one for distinguishing measurement efficiency from memorization, so I do not recommend reject. major_revision is appropriate because the central claim is internally coherent and the required work is documentation and verification of the stated controls, not a redesign. If the full paper delivers those ablations as advertised, the bar for minor_revision or accept would be met. Scope fit for quant-ph / QST venues looks appropriate."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing to know: this abstract is unusually careful about the real failure mode of amortized tomography—family memorization—and builds the right controls for it. A k=0 prior-only baseline and a shuffled-measurement control are exactly what you want when someone claims few-shot recovery of MPS states. Their main claim is that a once-trained fixed-protocol estimator, conditioned on an informative local Pauli set rather than random strings, beats those baselines (≈0.95 fidelity, up to +0.59 over prior-only), keeps quality as n and χ grow, ships conformal ≈90% coverage (including for unmeasured observables), and closes a small IBM loop at 0.97.\n\nWhat is actually new is not amortized tomography or MPS cores or conformal prediction in isolation. It is the combination under a fixed protocol, the deliberate refusal to lean on a fancy set encoder (plain MLP matches), the gauge-invariant fidelity loss, and the insistence that local reduced densities determine a χ-MPS so the measurement design is the lever. That design story is coherent on its face. The hardware check and the claim that gain grows with n are the pieces that would make this useful for near-term characterization if the full paper backs them.\n\nSoft spots are almost entirely about missing text. We have no methods, ablations, data splits, error bars, or code. The free parameters (network weights, conformal level, choice of local Paulis) are ordinary; nothing invented. The weakest assumption is the one they themselves surface: states must sit on a concentrated low-χ MPS manifold, otherwise the prior-only control already wins and you are not doing tomography. That is a real scope limit, not a hidden flaw. Circularity burden looks low because the decisive controls sit outside the training loss.\n\nWho it is for: people who actually run few-shot state estimation or device benchmarking on MPS-like states and care about calibrated uncertainty. Not a theory reorganizer. It deserves a serious referee if the full paper and artifacts exist; desk-rejecting on abstract alone would be wrong. I would send it to review and ask hard for the control tables, the Pauli design details, and the IBM protocol. Bring to reading group only after the PDF is up; cite only after numbers check out.","headline":"Abstract-only: fixed-protocol amortized MPS tomography with the right controls (prior-only + shuffled) and conformal uncertainty; numbers look useful if they hold.","tokens_in":2966,"tokens_out":558,"would_cite":false,"duration_ms":5705,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["03.65.Wj","03.67.-a","07.05.Mh"],"model":"grok-4.5","headline":"A once-trained amortized MPS map, conditioned on fixed local Pauli measurements, recovers high-fidelity quantum states with calibrated uncertainty, and uses the data rather than memorizing the family.","keywords":["quantum state tomography","matrix product states","amortized inference","local Pauli measurements","gauge-invariant fidelity","conformal prediction","bond dimension","hardware validation"],"falsifier":"On a concentrated MPS family, replace the informative local Pauli inputs with a shuffled or random measurement set of equal size; if the amortized estimator no longer produces a large fidelity gain over the k=0 prior-only baseline and fails the control, the claim that the protocol genuinely uses measurements is false.","tokens_in":3068,"feed_emoji":"⚛️","tokens_out":932,"duration_ms":6892,"temperature":0.7,"pith_summary":"Quantum state tomography is starved of samples, yet the states experimenters actually prepare usually live on a narrow, learnable manifold such as low-bond-dimension matrix-product states. The paper shows that a pure prior already looks near-optimal on such concentrated families, so claims of \"high fidelity from few measurements\" can be family memorization rather than genuine tomography. Its main proposal is a fixed-protocol amortized estimator: a single network is trained once to map a carefully chosen set of local Pauli measurements into MPS cores, using a gauge-invariant fidelity loss. Conditioning on an informative local Pauli design (motivated by the fact that local reduced density matrices determine a χ-MPS) turns a modest, memorization-prone map into a high-fidelity reconstructor (≈0.95, gains of up to +0.59 over prior-only) that decisively beats a shuffled-measurement control. A dropout ensemble, conformally recalibrated, supplies ≈90% coverage intervals even for observables never measured. Quality holds as system size and bond dimension grow, the representation contracts natively to 20 qubits, and the loop closes on IBM hardware at average fidelity 0.97.","feed_headline":"Fixed local Paulis turn one trained MPS map into high-fidelity tomography","feed_subtitle":"Gains up to +0.59 over prior-only, passes shuffle control, ~90% conformal intervals, 0.97 on IBM","key_machinery":"The fixed-protocol amortized MPS estimator: a once-trained network that maps a fixed informative local Pauli design into MPS cores under a gauge-invariant fidelity loss; the local design is the decisive lever because local reduced density matrices determine a χ-MPS.","core_discovery":"A fixed-protocol amortized MPS estimator trained once with a gauge-invariant fidelity loss, when conditioned on an informative local Pauli set rather than random strings, recovers high-fidelity states (≈0.95, up to +0.59 over prior-only) and decisively passes a shuffled-measurement control, with quality holding as n and χ grow and with conformal ≈90% coverage intervals.","pith_inferences":["The same local-RDM motivation may extend the fixed-protocol idea to other tensor-network families (PEPS, tree tensor networks) whose local reduced densities also fix the global state.","Because the design is fixed and non-adaptive, the method is a natural candidate for simultaneous multi-state characterization or continuous monitoring of a device whose states remain on the trained manifold.","If the concentration assumption is only partially satisfied, hybrid schemes that fall back to per-instance posterior refinement when the amortized residual is large become a concrete next experiment."],"forward_implications":["Once trained, the amortized map reconstructs new states from a fixed local Pauli protocol without per-instance optimization, scaling to n=10 at fidelity 0.90 with the gain growing in n.","Conformal recalibration of a dropout ensemble yields ≈90% coverage intervals for both reconstructed states and never-measured observables.","Native MPS contraction remains polynomial, supporting reconstruction up to 20 qubits.","The same trained estimator applied to hardware-measured local Paulis recovers five IBM states at average fidelity 0.97."],"fun_headline_variants":["Fixed local Paulis lift amortized MPS tomography to ≈0.95 fidelity","Local Pauli set turns fixed-protocol MPS map into high-fidelity estimator","Amortized MPS recovers 0.95 fidelity on informative local Paulis not random","One trained MPS estimator gains +0.59 over prior via fixed local Paulis","Fixed-protocol MPS tomography holds 0.90 fidelity at n=10 with local Paulis"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The states of interest must be sufficiently concentrated on a low-bond-dimension MPS manifold so that a fixed local Pauli design plus a once-trained amortized map can recover them; without that concentration the measurement-efficiency claim collapses to family memorization.","fun_headline_variants_meta":{"raw":{"variants":["Fixed local Paulis lift amortized MPS tomography to ≈0.95 fidelity","Local Pauli set turns fixed-protocol MPS map into high-fidelity estimator","Amortized MPS recovers 0.95 fidelity on informative local Paulis not random","One trained MPS estimator gains +0.59 over prior via fixed local Paulis","Fixed-protocol MPS tomography holds 0.90 fidelity at n=10 with local Paulis"]},"model":"grok-4.5","effort":"low","cost_usd":0.005092,"raw_usage":{"total_tokens":1502,"prompt_tokens":932,"num_sources_used":0,"completion_tokens":109,"cost_in_usd_ticks":50920000,"prompt_tokens_details":{"text_tokens":932,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":461,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":932,"tokens_out":109,"duration_ms":4039,"temperature":1.0,"reasoning_tokens":461,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T02:19:46.476239+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a concentrated MPS family, replace the informative local Pauli inputs with a shuffled or random measurement set of equal size; if the amortized estimator no longer produces a large fidelity gain over the k=0 prior-only baseline and fails the control, the claim that the protocol genuinely uses measurements is false.","supporting_citations":[],"review_version":1}