{"id":"9f86a837-8cd9-448c-8f10-5a6183449eca","arxiv_id":"2501.12765","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"AFQMC/NOCI self-refinement selects determinants from the AFQMC random walk to build compact multi-determinant trial states, achieving chemical accuracy for weakly correlated molecules.","lead":"This paper teaches an expensive quantum chemistry method, AFQMC, to improve its own starting guess by cherry-picking promising configurations from its random walk and folding them into a better trial wavefunction. On small molecules the trick cuts energy errors by up to tenfold, and on the HEAT benchmark set it reaches chemical accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The untested truncation of the selection walk to the first 100 Cholesky vectors may cap the determinant manifold, making the reported 100-200 determinant accuracy conditional on this choice.","rationale":"The reader identified the general assumption that the AFQMC random walk adequately samples the determinant manifold needed for the exact ground state. My concern is a concrete, testable instance of that assumption: the selection walk is explicitly restricted to the first 100 Cholesky vectors, and the paper provides no convergence evidence with respect to this truncation. This is more specific than the reader's statement and is not addressed in the reader's rationale, hence 'partial' agreement. The concern does not overturn the central claim for weakly correlated systems, because the benchmarks may still be valid for the tested molecules; however, it adds a clear condition: the reported success depends on the adequacy of the 100-vector manifold. The recommended verdict remains CONDITIONAL, so 'UNCHANGED' is appropriate. The proposed check is direct, computational, and would settle whether the truncation is benign or load-bearing. I credit the paper for reporting the N2 failure and for systematic parameter analysis, but the missing Cholesky convergence study is a genuine gap in the evidence for the headline claim.","tokens_in":14366,"tokens_out":5061,"duration_ms":55085,"concrete_test":"Repeat the AFQMC/NOCI selection for benzene at epsilon_min = 1e-6 and, as a control, for O2, using NChol = 100, 200, 500, and full Cholesky rank, with all other parameters fixed to Table I. If the final AFQMC correlation energy shifts by more than the 0.2 mEh statistical target, or if the selected NOCI expansion grows substantially beyond 214 determinants, the 100-vector truncation is load-bearing and the '100-200 determinants suffice' claim must be re-scoped. Additionally, compare the overlap between the selected determinant set and the full-rank walker distribution to quantify the missing manifold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Sec. III the authors restrict the Hubbard-Stratonovich propagator used during NOCI selection to the first 100 Cholesky vectors for all systems by setting L_g = 0 for g > 100 in Eq. (6). This truncation defines the space of Slater determinants reachable by the random walk, and therefore determines the candidate pool available to the three selection tests (Eqs. 18, 20, 23). No convergence study in the number of Cholesky vectors is reported. For benzene (30 electrons, 108 orbitals) and for several HEAT molecules, the full Cholesky rank at the stated 1e-6 threshold is likely much larger than 100; if the truncated propagator cannot generate determinants outside this 100-dimensional manifold, the resulting NOCI trial wave function is limited no matter how small epsilon_min is. The central claim that 100-200 non-orthogonal determinants achieve chemical accuracy is then conditional on an untested numerical restriction rather than on the self-refinement algorithm itself. The N2 failure is a related symptom: the random walk can miss an important component of the exact state, here the RHF-like part. The text asserts the 100-vector constraint is 'not essential' but supplies no comparison with larger or untruncated selections.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an algorithm for constructing non-orthogonal multi-Slater-determinant (NOCI) trial wave functions for AFQMC by selecting determinants directly from an AFQMC random walk. The selection uses three criteria: low local energy, small overlap with the current trial space, and a variational energy-lowering test. The trial wave functions are refined over several epochs with a decreasing energy threshold epsilon_min. The method is applied to O2 for calibration, then to second-row atoms, the HEAT set, benzene, and N2 dissociation. The authors report systematic convergence in epsilon_min, reductions of AFQMC error by up to a factor of 10 for second-row atoms, average HEAT-set errors within chemical accuracy, an 80% error reduction for benzene with 214 determinants, and a reduction in sampling variance. The N2 dissociation case is reported as a failure mode where the method is worse than AFQMC with a UHF trial.","tokens_in":14635,"tokens_out":7756,"duration_ms":80542,"significance":"If the central claims hold, this is a practically valuable contribution: it removes the need for an external CI method to generate multi-determinant trial states for weakly correlated AFQMC calculations. The paper has several strengths: the energy test uses a variational NOCI upper bound, so the selection is not circular in a damaging sense; the final AFQMC energies are independent Monte Carlo estimates benchmarked against external FCI and CCSDTQP values; and the N2 failure is honestly reported and analyzed. The main weakness is that the determinant manifold accessible to the selection walk is restricted by an untested truncation to 100 Cholesky vectors, and several selection parameters are calibrated on O2 alone. These issues do not necessarily invalidate the results, but they make the generality of the claims conditional and need to be addressed.","major_comments":[{"comment":"The random-walk propagator used during selection is restricted to the first 100 Cholesky vectors, and no convergence study with respect to this cutoff is reported. Because the auxiliary-field directions in Eq. (6) define the manifold of Slater determinants the walk can visit, this truncation directly limits the candidate pool available to the three selection tests in Eqs. (18), (20), and (23). The text states that the constraint is \"not essential\" but provides no comparison with larger cutoffs or with the untruncated propagator. For benzene (108 orbitals) and for the larger HEAT molecules, the full Cholesky rank at the stated 1e-6 threshold is likely much larger than 100. The authors should add a systematic study of this cutoff, for example on O2 and benzene, and either relax the truncation or justify it quantitatively.","section":"Sec. III, Eq. (6)"},{"comment":"The N2 dissociation results show that the algorithm can produce a trial state worse than the AFQMC/UHF result when the random walk does not explore the RHF-like component of the wave function. Since the selected determinants are drawn from the walk defined by Eqs. (4)-(5), the method can only refine the trial state within the manifold sampled under the current trial. This is a fundamental limitation of the self-refinement idea, not a numerical accident. The paper should state this limitation more prominently in the abstract and conclusion, and should discuss or test possible remedies, such as seeding the walk with multiple initial determinants or using a small ensemble of trial states during selection.","section":"Sec. IV C and Eqs. (12)-(13)"},{"comment":"The default selection parameters lambda=4.0, mu=0.6, N_w=6400, N_k=100, and tau=0.05 are calibrated on O2. The sensitivity analysis in Figs. 1-3 is performed only for O2. The paper claims that epsilon_min is the only remaining adjustable parameter that determines accuracy, but this is only demonstrated on a single system. The authors should show, for at least one other system such as benzene or a representative HEAT molecule, that the default parameters are not in a sensitive regime, before asserting that the method has a single convergence parameter.","section":"Table I and Sec. III"}],"minor_comments":[{"comment":"The abstract says a \"10-fold increase of the time to solution\" for the benzene calculation, but Sec. IV D and Fig. 10 state that the 214-determinant calculation is eight times slower than AFQMC/HF. Please reconcile these numbers.","section":"Abstract and Sec. IV D"},{"comment":"The conclusion states that \"100-200\" non-orthogonal Slater determinants achieve chemical accuracy for all weakly correlated systems, but Table III reports an average of 194 and a maximum of 324 determinants at epsilon_min=1e-6, and benzene uses 214 determinants. Please rephrase to \"on the order of 100-300 determinants\" or quote a more precise statistic.","section":"Table III and Conclusion"},{"comment":"The Data Availability section says the data are available \"within the article,\" but the text states that all AFQMC values are provided in the Supporting Information, which is not included here. Please include the raw energy tables or correct the data availability statement.","section":"Data Availability"},{"comment":"The quantity E used in Eq. (23) is not defined in Section II B 3; it should be explicitly identified as the current NOCI energy of |Phi_Nd> before adding the candidate determinant.","section":"Eq. (22)"},{"comment":"There are several typographical errors, including \"Inforamtion,\" \"reamins,\" and \"becuase\"; the manuscript should be proofread.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the scope of the journal and the core idea is attractive, but the untested Cholesky truncation is a load-bearing gap that affects all reported results. The N2 failure is honestly reported but the abstract and conclusion overstate the generality of the method. If the authors can provide the requested convergence study, demonstrate that the O2-calibrated defaults transfer to other systems, and temper the claims accordingly, I would support publication. The raw AFQMC energy tables should also be made available."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know: this is a genuinely new and well-executed method for building multi-determinant AFQMC trial states by mining the AFQMC random walk itself, and the results on weakly correlated systems are credible — the systematic epsilon_min convergence, the five-seed reproducibility, and the honest N2 failure all suggest the authors understand what their method can and can't do. But the paper has one untested restriction, the 100-Cholesky-vector cap on the selection walk, that could limit the method's generality, and no code/data is shipped, so independent checking is hard.\n\nThe core idea is simple and good: during short AFQMC walks, pick walkers with low local energy, small overlap with the current trial, and meaningful variational energy gain in a two-step NOCI test. This removes the need for an external CI method, which is the practical bottleneck the paper claims to solve. On second-row atoms, the HEAT set, and benzene, the NOCI trial states with 100-200 determinants systematically reduce AFQMC errors to chemical accuracy or near it, with variance reduction offsetting much of the cost. That's a real advance.\n\nThe soft spots are real but proportionate. The 100-Cholesky truncation is asserted to be 'not essential' but no convergence study in the number of Cholesky vectors is reported. Since this truncation defines the determinant manifold the walk can explore, it is load-bearing for the claim that the self-refinement algorithm itself, rather than a lucky prior restriction, achieves the accuracy. The N2 dissociation result is a related warning: the walk can miss an important part of the space. This should be tested directly, and the authors should either remove the cap for a few systems or show that the cap is inactive. The hand-tuned parameters are calibrated on O2, though the sensitivity plots show they are not delicate. No code or raw energies are shipped, which is a minor reproducibility gap for a methods paper. The text's claim that 100-200 determinants achieve chemical accuracy is accurate for the specific systems shown, but the abstract's framing is broader than the evidence.\n\nBottom line: this deserves a serious referee. The central idea is sound for weakly correlated systems, the benchmarks are honest, and the failure mode is reported rather than hidden. I'd send it to review with a request for a Cholesky-convergence study, raw energies, and ideally code. I wouldn't cite it as a proven general method until the truncation question is settled.","headline":"A genuinely new method that mines the AFQMC walk for multi-determinant trial states, with credible results on weakly correlated systems, but an untested 100-Cholesky-vector cap on the selection walk and missing code/data leave the generality open.","tokens_in":15182,"tokens_out":2229,"would_cite":false,"duration_ms":22612,"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":"AFQMC builds its own trial wavefunction, cutting errors tenfold.","keywords":["auxiliary-field quantum Monte Carlo","non-orthogonal configuration interaction","trial wavefunction refinement","multi-Slater determinant","phaseless approximation","chemical accuracy","HEAT set","static correlation"],"falsifier":"Run AFQMC/NOCI on stretched N2 (R ≈ 4.2 a0) twice: once with a spin-restricted initial random walk that is forced to explore the RHF-like manifold, and once with the standard spin-contaminated UHF walk. If the restricted walk reproduces the hand-built RHF-UHF accuracy while the standard walk does not, the walker-covering premise is confirmed; if neither reaches FCI accuracy, then the selection criteria themselves are insufficient and the claim that the walk is a complete basis for the ground state fails.","tokens_in":14156,"feed_emoji":"⚛️","tokens_out":6117,"duration_ms":53281,"temperature":0.7,"pith_summary":"Phaseless auxiliary-field quantum Monte Carlo (AFQMC) is only as accurate as its trial wavefunction, and multi-determinant trial states are normally obtained from expensive external configuration-interaction codes. This paper argues that AFQMC can refine its own trial state: short random walks generate candidate non-orthogonal Slater determinants, and a three-step selection keeps only those that lower the NOCI energy. With 100-200 selected determinants, the method reduces AFQMC errors by up to a factor of ten for second-row atoms, brings the HEAT-set average within chemical accuracy, and cuts the benzene error by 80% while also lowering sampling variance. The remaining failures occur precisely where the random walk misses the relevant determinant manifold, as in N2 dissociation with a spin-contaminated unrestricted walk. If the claim holds, multi-determinant AFQMC becomes practical for weakly correlated molecules without needing an external CI method.","feed_headline":"AFQMC builds its own trial wavefunction, cutting errors tenfold","feed_subtitle":"Self-selected 100–200 Slater determinants bring HEAT-set molecules within chemical accuracy.","key_machinery":"The central object is the NOCI expansion in non-orthogonal Slater determinants harvested from the AFQMC walk, $|\\Phi_{N_d}\\rangle = \\sum_\\alpha c_\\alpha \\hat{B}(x_\\alpha)|\\Psi_0\\rangle$, where $\\hat{B}(x)$ is the imaginary-time propagator (6) applied to a reference determinant. The selection machinery is the triple test: an energy preselection threshold $\\lambda$, a metric test using the Hermitian, idempotent projector $\\hat{Q} = 1 - \\sum_{\\alpha\\beta}|\\Psi_\\alpha\\rangle S^{-1}_{\\alpha\\beta}\\langle\\Psi_\\beta|$ to reject determinants nearly parallel to the current space, and an energy test solving a two-configuration variational problem (22) to keep only determinants that change the energy by more than $\\varepsilon$. The single parameter $\\varepsilon_{\\min}$ controls the compactness-versus-accuracy trade-off, with the trial updated epoch by epoch. The claim is that these criteria yield the most compact NOCI expansion: roughly 8 determinants per electron across atoms, CO2, and benzene, versus millions for a CISD expansion.","core_discovery":"The paper's central claim is that the AFQMC random walk itself contains the information needed to build an accurate trial state. Treating each walker, at each time step, as a non-orthogonal Slater determinant $|\\Psi_\\alpha\\rangle = \\hat{B}(x_\\alpha)|\\Psi_0\\rangle$ (Eqs. 12-13), the algorithm selects determinants that (1) have local energies below $\\bar{E}_L - \\lambda \\sigma_{E_L}$, (2) have small overlap with the current NOCI space through the projector $\\hat{Q}$ with metric threshold $\\mu$, and (3) lower the variational energy by more than $\\varepsilon$ in a two-determinant test. Solving the NOCI equation (14) then fixes the coefficients. Iterating over epochs with decreasing $\\varepsilon$ converges the AFQMC/NOCI energy smoothly to the reference; for O2 the error drops to 0.7 mEh with 119 determinants at $\\varepsilon_{\\min}=10^{-6}$. Across second-row atoms the RMSD falls from 2.7 mEh (HF trial) to 0.2 mEh at $\\varepsilon_{\\min}=10^{-7}$, the HEAT-set RMSD falls to 1.1 mEh with an average of 194 determinants, and benzene reaches 0.7 mEh error with 214 determinants, an 80% reduction. The same variance reduction makes the added cost sub-linear when targeting fixed statistical errors.","pith_inferences":["If the walker-exploration premise holds in general, the same self-refinement could be ported to periodic and solid-state AFQMC codes, where external CI trial states are unavailable; the paper hints at this in its conclusion but does not test it.","A testable extension is to seed the random walk with both RHF and UHF initial determinants (or multiple symmetry-broken references) so that the selection can explore disconnected manifolds; the N2 failure suggests this could fix spin-contamination errors without hand-built RHF-UHF trials.","The observed roughly 8 determinants per electron across systems hints at a system-size scaling that, if confirmed, would make the method competitive with PHMSD approaches that need thousands to millions of determinants; this is the authors' own future-work suggestion, not a demonstrated result."],"forward_implications":["AFQMC/NOCI with 100-200 determinants is within chemical accuracy for weakly correlated molecules including the HEAT set, so multi-determinant AFQMC no longer requires an external CI code.","For second-row atoms the RMSD improves by a factor of 10 over AFQMC/HF (2.7 to 0.2 mEh at $\\varepsilon_{\\min}=10^{-7}$).","Sampling variance drops enough that, at fixed statistical error, the wall-time scaling with determinant count is only $N_d^{1/4}$ (benzene: 8x slower with 214 determinants).","The remaining errors concentrate in strongly spin-contaminated cases; the N2 curve shows AFQMC/NOCI can be worse than AFQMC/UHF unless the walk explores the relevant (RHF-like) manifold.","Benzene's error decreases roughly as $N_d^{-0.6}$, suggesting systematic convergence with more determinants."],"supporting_citations":[{"why":"Supplies the AFQMC/HF baseline and the QMCFort implementation showing poor HF-trial performance for O2, HEAT molecules, and benzene.","marker":"[23]"},{"why":"Provides the AFQMC/CISD comparison and the large multi-determinant baseline that this paper's compact NOCI results are measured against.","marker":"[26]"},{"why":"Supplies the efficient local-energy evaluation algorithm for multi-Slater trial wavefunctions that makes many-determinant AFQMC practical.","marker":"[31]"},{"why":"Prior AFQMC calculations with non-orthogonal multi-Slater determinants that this work extends with a self-contained selection strategy.","marker":"[36]"},{"why":"The large-time-step propagator used in the random walk, which reduces time-step error and is the specific form of $\\hat{B}(x)$ used here.","marker":"[51]"},{"why":"Non-orthogonal Wick's theorem, the formalism used to compute local energies and NOCI Hamiltonian/overlap matrix elements.","marker":"[52]"},{"why":"Motivates constraining the random walk to the first 100 Cholesky vectors, which accelerates selection and produces more compact NOCI wavefunctions.","marker":"[55]"},{"why":"Provides the exact DMRG reference energies for the N2 dissociation curve used to expose the method's static-correlation failure.","marker":"[62]"}],"fun_headline_variants":["AFQMC self-refines via its own random walk","AFQMC selects its own trial determinants from the walk","Random walk refines AFQMC trial states, cutting errors 10x","AFQMC auto-tunes trial state for chemical accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the AFQMC random walk, driven by the current trial wavefunction, actually explores every region of determinant space that the exact ground state needs; if the walkers miss a region, the three selection tests can never recover it, which is exactly what happens in the N2 dissociation test.","fun_headline_variants_meta":{"raw":{"variants":["AFQMC self-refines via its own random walk","AFQMC selects its own trial determinants from the walk","Random walk refines AFQMC trial states, cutting errors 10x","AFQMC auto-tunes trial state for chemical accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000818,"raw_usage":{"total_tokens":3623,"prompt_tokens":1028,"completion_tokens":2595,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":2524}},"tokens_in":644,"tokens_out":2595,"duration_ms":18776,"temperature":1.0,"reasoning_tokens":2524,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:48:44.197262+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run AFQMC/NOCI on stretched N2 (R ≈ 4.2 a0) twice: once with a spin-restricted initial random walk that is forced to explore the RHF-like manifold, and once with the standard spin-contaminated UHF walk. If the restricted walk reproduces the hand-built RHF-UHF accuracy while the standard walk does not, the walker-covering premise is confirmed; if neither reaches FCI accuracy, then the selection criteria themselves are insufficient and the claim that the walk is a complete basis for the ground state fails.","supporting_citations":[{"cited_title":"Wei , author S","cited_arxiv_id":null,"evidence_quote":"Provides the AFQMC/CISD comparison and the large multi-determinant baseline that this paper's compact NOCI results are measured against."},{"cited_title":"Mahajan \\ and\\ author S","cited_arxiv_id":null,"evidence_quote":"Supplies the efficient local-energy evaluation algorithm for multi-Slater trial wavefunctions that makes many-determinant AFQMC practical."},{"cited_title":"Shavitt \\ and\\ author R","cited_arxiv_id":null,"evidence_quote":"Prior AFQMC calculations with non-orthogonal multi-Slater determinants that this work extends with a self-contained selection strategy."},{"cited_title":"Sukurma , author M","cited_arxiv_id":null,"evidence_quote":"Non-orthogonal Wick's theorem, the formalism used to compute local energies and NOCI Hamiltonian/overlap matrix elements."},{"cited_title":"Aquilante , author L","cited_arxiv_id":null,"evidence_quote":"Provides the exact DMRG reference energies for the N2 dissociation curve used to expose the method's static-correlation failure."}],"review_version":1}