{"id":"b9fb2fcc-b335-42b2-bd0f-9320f5ef06f2","arxiv_id":"2505.06798","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A pairwise autoregressive graphical model trained with first-order gradients matches or outperforms heavier neural-network quantum states for stoquastic spin Hamiltonians, especially with limited time and samples.","lead":"Quantum physicists often use neural networks to find the ground states of magnets and spin models. This paper shows that a much simpler linear model, made of pairwise interactions in an autoregressive network, can match or beat neural network states in short, resource-limited runs, and it converges faster on frustrated systems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pairwise sufficiency for 2D/frustrated systems is inferred from 7-spin 1D exact learning; an exact-learning check on small frustrated instances would settle it.","rationale":"The reader's conditional verdict already flags the transfer from 7-spin 1D exact learning to larger 2D/frustrated systems. I agree that this is the load-bearing step: every headline statement about pairwise sufficiency and about AGMs outperforming NNQS depends on the variational family containing a good approximation to the target ground state. Without that, faster energy decrease is merely faster underfitting. The proposed test directly probes Eq. (7) on target problem classes using the paper's own exact-learning tool, so it is feasible and decisive. I considered the fairness of the NN/SR comparison (fixed NN hyperparameters vs. AGM Hyperopt) as an alternative concern; it is real but secondary because the authors disclose it and it does not bear on the representational claim. I also considered the lack of error bars, which weakens the quantitative comparisons but does not single out the central construction. The proposed check would either rescue the pairwise Ansatz or force the paper's claims to be narrowed to resource-limited optimization speed.","tokens_in":15311,"tokens_out":5720,"duration_ms":62528,"concrete_test":"Perform exact learning for a small frustrated instance, e.g., a 4x4 or 16-spin D-TIM with random +/- J couplings at g=1.0 or the ANNNI model at alpha=1/3, g=0.25: compute the exact ground state by full diagonalization, then solve the interaction-screening objective in Eq. (9) with the complete polynomial including all interaction orders. Report ||theta*_k||_1 per order for several variable orderings (row-major, random, and one chosen to minimize the pairwise residual). If pairwise terms are not dominant in these 2D or frustrated conditionals, the pairwise Ansatz is not representationally sufficient and the central claim should be weakened to a resource-limited optimization-speed claim; if pairwise terms remain dominant across orderings, the extrapolation is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The pairwise truncation in Eq. (7) is justified by exact interaction screening on 7-spin 1D chains (Sec. III.A, Fig. 1a,b), and that evidence is then used to justify the same pairwise Ansatz for 10x10 2D TIM/XXZ and for frustrated D-TIM/ANNNI models (Secs. III.B-III.C), yet no exact-learning check is performed for 2D or frustrated systems, and the variable ordering used in the autoregressive factorization is never specified. The conditional P(sigma_i|sigma_>i) is order-dependent: a pairwise-logistic form can exactly capture an Ising distribution under one ordering while requiring high-order terms under another, and frustrated or longer-range models can have irreducible multi-spin conditionals. If the true conditionals in the 2D or frustrated cases are not pairwise-dominated, the observed AGM advantage is not evidence that the Ansatz represents the ground state; it is evidence only that an underparameterized model reaches a (possibly poor) variational energy faster under ADAM. The title/abstract claim that these models efficiently find ground state representations would then rest on an untested extrapolation. The paper itself lists a principled variable-ordering choice as future work in the Conclusions, which is an admission that the representational question is unresolved.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Autoregressive Graphical Models (AGMs) whose conditionals have pairwise energy functions, as a variational Ansatz for stoquastic spin Hamiltonians. For small 1D systems (7 spins), exact learning via Interaction Screening shows that pairwise terms dominate the autoregressive conditionals; this motivates using the pairwise AGM for larger systems. The authors compare AGMs trained with first-order stochastic gradient descent against NetKet neural-network quantum states (NNQS) trained with stochastic reconfiguration, and against tensor-network benchmarks, on 10×10 2D transverse-field Ising (TIM) and XXZ models, and on frustrated disordered TIM and ANNNI models. They find that AGMs achieve comparable or better variational energies in resource-limited settings, especially for frustrated systems where exact sampling avoids MCMC mixing problems. The paper claims that simple pairwise AGMs can efficiently represent ground states of a variety of stoquastic models.","tokens_in":15643,"tokens_out":8598,"duration_ms":81276,"significance":"If the results hold, the paper offers a simple, interpretable, and exactly samplable variational Ansatz that is competitive with more complex neural-network states in resource-limited settings. The use of exact learning to guide Ansatz selection is a valuable methodological idea, and the observation that exactly samplable models help in frustrated systems is practically relevant. However, the evidence has important gaps: the representational justification rests on small 1D systems, and the performance comparisons lack statistical robustness. With additional validation, the paper could make a useful contribution to the variational quantum Monte Carlo toolbox.","major_comments":[{"comment":"The pairwise truncation in Eq. (7) is justified only by exact learning on 7-spin 1D chains (Fig. 1a,b). No exact-learning check is performed for 2D or frustrated systems, and the variable ordering in the autoregressive factorization is never specified. Because the conditional P(σ_i|σ_{>i}) is order-dependent, the pairwise Ansatz may not be able to represent the ground state for the 10×10 TIM/XXZ, D-TIM, or ANNNI models even approximately; the observed energy advantage then reflects only the optimization dynamics of an underparameterized model. The paper itself lists a principled ordering choice as future work (§IV), acknowledging that this representational question is unresolved. I recommend adding exact-learning diagnostics on small 2D and frustrated instances (e.g., 4×4 TIM and small frustrated systems) or, if infeasible, restricting the representational claim to the tested models and framing the results as an empirical resource-limited comparison only.","section":"III.A, III.B, III.C"},{"comment":"The central performance claim that AGM 'outperforms' NNQS is supported only by single-run energy-vs-time curves (Figs. 2-4) and a disorder-averaged difference (Fig. 3b) that is shown without error bars or the number of disorder realizations. For a statement about variational optimization, run-to-run fluctuations from weight initialization and stochastic sampling are substantial; without multiple seeds and standard errors, the claimed advantage could be within noise. The disorder-averaged experiment in Fig. 3b should report the mean and standard error over at least tens of disorder realizations.","section":"III.B, III.C (Figs. 2-4)"},{"comment":"The NNQS baseline is trained with a fixed set of hyperparameters (§A.2), while the AGM receives a low-cost hyperparameter optimization (§A.3). The authors state that the NN performance 'can presumably be increased further by using more intense hyperoptimization' (§III.B). This asymmetry biases the comparison in favor of AGM. To make the 'outperforms' claim credible, the NN baseline should either receive comparable hyperparameter tuning or the paper should report the sensitivity of the NN results to hyperparameter choices.","section":"III.B, III.C (and Appendix A.2)"},{"comment":"For the 2D and frustrated systems, the paper reports only variational energies relative to NNQS and a non-variational PEPS benchmark. It never reports the exact ground state energy (e.g., from exact diagonalization on small 2D systems or from extrapolated DMRG on quasi-1D systems) or the overlap/fidelity of the AGM state with the true ground state. Consequently, the title's claim that the method 'efficiently find[s] ground state representations' is not quantitatively supported: lower energy than a competing variational method does not imply that the state is a good representation of the ground state. Adding an accuracy benchmark (e.g., exact ground state energies for 4×4 and 6×6 systems) would substantially strengthen the paper.","section":"III.B, III.C"}],"minor_comments":[{"comment":"'Given an-qubit HamiltonianH' should read 'Given an n-qubit Hamiltonian H'; there are also broken sentences such as 'and this condition is satisfied for all local Hamiltonains'.","section":"Section II"},{"comment":"The text says 'we look at the sum of absolute values in the solution of (9) for every order (||θ*_k||_1 for order k)', but the caption of Fig. 1 says 'The magnitude of the largest term at each order'. These are different quantities; please clarify.","section":"Section III.A"},{"comment":"The description of the ANNNI model says 'anti-ferro magnetic interactions are only present in the horizontal directions', but the n.n.n. interactions are along the y-axis; this appears to be a typo for 'vertical directions'.","section":"Section III.C"},{"comment":"The variable ordering used in the autoregressive factorization (e.g., raster-scan order) is never specified in the main text or the appendix; please state it explicitly.","section":"Sections II and III"},{"comment":"In Fig. 2, the TN result is shown as a line on the energy-vs-time plot, but the TN algorithm does not have a time-dependent energy curve; please clarify that this is a reference value.","section":"Figure 2 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a potentially useful variational Ansatz and an interesting application of exact learning to guide architecture choice. The main weaknesses are statistical (no error bars, single runs) and the lack of validation of the pairwise truncation for the actual 2D/frustrated systems studied. These are fixable within the scope of the manuscript, so I recommend major revision rather than rejection. The authors should also be careful to separate the empirical resource-limited comparison from the stronger representational claim in the title."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper introduces autoregressive graphical models (AGMs) with pairwise energy conditionals as a variational ansatz for stoquastic Hamiltonians. That is genuinely new and useful: the ansatz is simple, exactly samplable, costs O(n^2) per energy estimate, and the numerical experiments show it can reach competitive energies quickly on 2D TIM, XXZ, and frustrated D-TIM/ANNNI models. I appreciate that they do exact learning on small systems to justify the pairwise truncation, even if that evidence is limited. The comparison against NetKet NNQS and PEPS/DMRG benchmarks is a step beyond many papers in this area, and the appendix gives enough implementation detail to reproduce the main curves.\n\nThe soft spots are real. First, there are no error bars or multiple-seed statistics anywhere in the energy-vs-time plots or the disorder-averaged difference. Given that the core claim is about 'often outperform,' a few independent runs and standard deviations would substantially strengthen it. Second, the pairwise-dominance observation comes from two 7-spin 1D chains, and the stress-test note is right: no exact-learning check is done for 2D or frustrated systems, and the variable ordering in the autoregressive factorization is never specified. The paper itself admits in the conclusions that a principled ordering is future work, which undercuts the 'sufficient to capture ground state properties' phrasing in the introduction. Third, the NN baseline is given fixed hyperparameters while the AGM gets hyperopt; the authors acknowledge this, but it makes the comparative claim less clean. The PEPS results are also selected post hoc across update rules and bond dimensions, which is fine for a reference point but not a fair variational comparison.\n\nDespite these gaps, the central idea is sound. A pairwise autoregressive model is a legitimate variational family, exact sampling is a real advantage in the frustrated regime, and the empirical results, while not definitive, are plausible. This is an honest empirical methods paper, not a breakthrough, but it is useful for practitioners who want a cheap, interpretable, exactly samplable ansatz.\n\nI would send this to peer review. A serious referee should ask for error bars, a small-system exact-learning check for 2D/frustrated models, and a fairer NN hyperparameter treatment. With those additions, the claims would be on solid ground.","headline":"A useful, simple exactly-samplable ansatz with a resource-limited performance claim that is supported only qualitatively; the pairwise-sufficiency motivation is thin, and the benchmark is uneven.","tokens_in":16129,"tokens_out":2017,"would_cite":true,"duration_ms":24242,"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":"Simple pairwise model beats neural nets at stoquastic ground states","keywords":["autoregressive graphical models","stoquastic Hamiltonians","variational Monte Carlo","pairwise energy Ansatz","exact sampling","frustrated spin systems","ground-state representation","energy-based models"],"falsifier":"Diagonalize a small frustrated 2D cluster exactly (for example an 8-spin ANNNI or $\\pm J$ transverse Ising plaquette), compute its full autoregressive representation by interaction screening in the infinite-sample limit, and compare the $\\ell^1$ norms of pairwise versus fourth-order parameters; if any higher-order norm exceeds the pairwise norm, the modeling premise fails for that case. A complementary check is to test whether the pairwise AGM's variational-energy error, measured against the exact ground state, decreases with $N_s$ in the frustrated regime or plateaus.","tokens_in":15141,"feed_emoji":"⚛️","tokens_out":7534,"duration_ms":67344,"temperature":0.7,"pith_summary":"The paper introduces Autoregressive Graphical Models (AGMs): variational states built by factoring a stoquastic ground-state wavefunction into conditional probabilities, each modeled by a logistic function of an energy containing at most pairwise spin interactions. Working on small systems where the exact autoregressive representation can be computed, the authors find that pairwise terms dominate higher-order terms across the parameter range studied, motivating the same pairwise Ansatz for larger systems. They then show that this linear pairwise AGM, trained with the first-order ADAM optimizer, matches or beats a neural-network quantum state of equal parameter count trained with stochastic reconfiguration within the same wall-clock time on $10\\times10$ transverse-field Ising and XXZ models, and converges faster on frustrated disordered and ANNNI models when compute is limited. The practical upshot is that a transparent pairwise model with exact sampling can be a better use of limited compute than a larger nonlinear black-box Ansatz for stoquastic spin problems.","feed_headline":"Simple pairwise model beats neural nets at stoquastic ground states","feed_subtitle":"It samples exactly, and in frustrated spin systems it converges faster than larger neural-nets within a fixed time budget.","key_machinery":"The machinery is the pairwise autoregressive conditional $P(\\sigma_i \\mid \\sigma_{>i};\\theta^{(i)}) = 1/(1+\\exp(-2\\sigma_i(\\theta^{(i)}_i + \\sum_{j>i} \\theta^{(i)}_j \\sigma_j)))$ for a fixed ordering of the $n$ spins. Multiplying these conditionals gives a normalized distribution from which independent samples are drawn in one forward pass, so variational Monte Carlo energy and gradient estimates need no Markov chain; this is the Autoregressive Graphical Model. The pairwise truncation is the load-bearing simplification: it keeps the model linear in the couplings, gives $O(n^2)$ evaluation cost, and matches the dominant order found by exact learning on small systems. Higher-order polynomial terms can be added to the same conditional form, so the pairwise choice is a deliberate modeling decision rather than an architectural ceiling.","core_discovery":"The central claim is that a simple pairwise-energy AGM is sufficient to capture the ground-state properties of a variety of stoquastic Hamiltonians, including frustrated ones, and that this transparent model often outperforms more complex nonlinear models when both are compared under realistic wall-clock and sample budgets. The claim rests on exact learning of the full autoregressive representation for 7-spin 1D chains: in both the transverse-field Ising and XXZ ground states, the norm of the order-2 parameters dominates order-4 and other terms, indicating that pairwise interactions carry the essential structure. The paper argues this dominance justifies truncating each conditional in Eq. (7) at pairwise order, which yields an $O(n^2)$ Ansatz that is exactly samplable, and that the resulting model is especially effective in the quantum-dominated regime of non-frustrated models and in the short-time regime of frustrated models, where Markov-chain sampling is slow to mix.","pith_inferences":["If pairwise dominance holds beyond the tested models, the AGM makes variational ground-state search for stoquastic systems nearly parameter-light: its structure is fixed by the Hamiltonian graph and a chosen variable order, leaving only a small set of coupling weights to learn.","The paper does not optimize the autoregressive variable ordering; comparing exact-learning norm profiles across random spin orderings is a direct way to test whether the pairwise advantage survives in stronger-frustration or larger systems.","A wall-clock comparison conflates per-sample cost with statistical efficiency; measuring gradient variance at equal sample counts could separate the benefit of exact i.i.d. sampling from the benefit of the linear parameterization.","The pairwise dominance found in the conditionals echoes the Jastrow wavefunction structure the paper mentions, so AGMs may offer a generative-model route to explaining when Jastrow-type Ansätze work and when higher-order corrections matter."],"forward_implications":["On $10\\times10$ ferromagnetic transverse-field Ising and antiferromagnetic XXZ models, the pairwise AGM reaches variational energies comparable to a neural-network state with the same parameter count, and outperforms it in the quantum-dominated regime, within the same compute-time budget.","On frustrated systems (disordered transverse Ising model and the 2D ANNNI model at small transverse field), the AGM converges to lower variational energy faster; the neural-network baseline catches up only after orders of magnitude more time.","Raising the number of samples $N_s$ systematically lowers the AGM's variational energy, while imposing known spin-flip and lattice-reflection symmetries helps mainly at low sample counts, suggesting the model learns the symmetries from data as samples increase.","Because the AGM is exactly samplable, it avoids slow-mixing Markov chains in frustrated regimes, making it a candidate warm-start or cheap first pass before more expensive variational simulations are run."],"supporting_citations":[{"why":"Supplies the energy-based representation of quantum many-body states that the AGM extends to conditional distributions.","marker":"[20]"},{"why":"Introduces neural-network quantum states and variational Monte Carlo, the baseline the AGM is compared against.","marker":"[5]"},{"why":"Establishes deep autoregressive models as efficient variational Ansätze for quantum systems.","marker":"[42]"},{"why":"Shows autoregressive variational models give exact sampling, the advantage AGM exploits in frustrated regimes.","marker":"[52]"},{"why":"Provides the interaction-screening estimator used for exact learning of small-system autoregressive representations.","marker":"[17]"},{"why":"Establishes consistency of interaction-screening learning, guaranteeing the learned conditionals reproduce the ground-state distribution.","marker":"[47]"},{"why":"The first-order ADAM optimizer whose cheap updates are central to the claim that expensive second-order methods are unnecessary.","marker":"[23]"},{"why":"Provides the neural-network baseline with stochastic reconfiguration used in the head-to-head timing comparisons.","marker":"[45]"},{"why":"Supplies near-exact 1D reference energies that benchmark the accuracy of the learned AGM ground states.","marker":"[50]"}],"fun_headline_variants":["Pairwise-only model wins on stoquastic ground states","Simple autoregressive model beats deep nets for stoquastic spins","Pairwise AGM outperforms complex nets for frustrated quantum systems","Less is more: pairwise model beats neural nets for stoquastic Hamiltonians","Pairwise beats neural nets for stoquastic ground states"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that pairwise dominance measured by exact learning on 7-spin 1D chains transfers to larger 2D systems and to frustrated Hamiltonians under a fixed, unoptimized spin ordering, so the pairwise truncation in the autoregressive conditionals remains accurate there.","fun_headline_variants_meta":{"raw":{"variants":["Pairwise-only model wins on stoquastic ground states","Simple autoregressive model beats deep nets for stoquastic spins","Pairwise AGM outperforms complex nets for frustrated quantum systems","Less is more: pairwise model beats neural nets for stoquastic Hamiltonians","Pairwise beats neural nets for stoquastic ground states"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000616,"raw_usage":{"total_tokens":2809,"prompt_tokens":844,"completion_tokens":1965,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":460,"completion_tokens_details":{"reasoning_tokens":1878}},"tokens_in":460,"tokens_out":1965,"duration_ms":15043,"temperature":1.0,"reasoning_tokens":1878,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:32:56.426971+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Diagonalize a small frustrated 2D cluster exactly (for example an 8-spin ANNNI or $\\pm J$ transverse Ising plaquette), compute its full autoregressive representation by interaction screening in the infinite-sample limit, and compare the $\\ell^1$ norms of pairwise versus fourth-order parameters; if any higher-order norm exceeds the pairwise norm, the modeling premise fails for that case. A complementary check is to test whether the pairwise AGM's variational-energy error, measured against the exact ground state, decreases with $N_s$ in the frustrated regime or plateaus.","supporting_citations":[{"cited_title":"Learning energy-based representations of quan- tum many-body states","cited_arxiv_id":null,"evidence_quote":"Supplies the energy-based representation of quantum many-body states that the AGM extends to conditional distributions."},{"cited_title":"Restricted boltzmann machines for quantum states with non-abelian or anyonic symmetries","cited_arxiv_id":null,"evidence_quote":"Shows autoregressive variational models give exact sampling, the advantage AGM exploits in frustrated regimes."},{"cited_title":"Markov random fields in statistics","cited_arxiv_id":null,"evidence_quote":"Provides the interaction-screening estimator used for exact learning of small-system autoregressive representations."},{"cited_title":"Deep autoregressive models for the efficient variational simulation of many-body quantum systems","cited_arxiv_id":null,"evidence_quote":"Establishes consistency of interaction-screening learning, guaranteeing the learned conditionals reproduce the ground-state distribution."},{"cited_title":"Many-body problem with strong forces","cited_arxiv_id":null,"evidence_quote":"The first-order ADAM optimizer whose cheap updates are central to the claim that expensive second-order methods are unnecessary."},{"cited_title":"Density-matrix algorithms for quantum renormalization groups","cited_arxiv_id":null,"evidence_quote":"Supplies near-exact 1D reference energies that benchmark the accuracy of the learned AGM ground states."}],"review_version":1}