{"id":"8f525114-b0ae-496a-86de-63a565d7228a","arxiv_id":"2502.02582","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"By tuning stochastic interpolants for periodic crystals and coupling them with discrete flow matching, OMatG sets new state-of-the-art results on crystal structure prediction and de novo materials generation.","lead":"OMatG is a new open-source toolkit for generating inorganic crystal structures using stochastic interpolants, a flexible framework that includes diffusion and flow matching as special cases. The authors report new state-of-the-art match rates and stability rates on standard materials benchmarks, and show that tuning the interpolation path matters substantially.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is load-bearing on single-run, best-of-HPO results, and no single OMatG configuration dominates all reported DNG/CSP metrics; the 2.18-point S.U.N. margin over MatterGen-MP has no error bars.","rationale":"I examined the periodic-boundary SI extension (Sec. III.B.1, App. B.5). The construction is not formally proven, but the conditional-expectation velocity for wrapped interpolants is periodic because the lift ambiguity cancels in the sum over periodic images, so the reader's theoretical concern is plausible but less likely to invalidate the central empirical claim. The load-bearing condition for the SOTA claim is instead that the reported single-run, best-of-HPO numbers are reliable and that the claim is interpreted consistently with the metric-by-metric results. The paper's own tables show different OMatG configurations winning different metrics, so 'outperforming' must be scoped to a pre-specified primary metric per task. This does not change the reader's conditional verdict but sharpens the required revision: pre-register primary endpoints and report seed-resolved uncertainty for the small DNG margins.","tokens_in":32842,"tokens_out":17777,"duration_ms":182149,"concrete_test":"Run the reported best DNG configuration (Linear SDE w/γ) with 5 independent training seeds and run MatterGen-MP with 5 sampling seeds, computing S.U.N. rate with 95% bootstrap CIs. If the CI lower bound of OMatG does not exceed the CI upper bound of MatterGen-MP (or if the gap is within noise), the Section I.B DNG-SOTA claim is not supported. As a secondary check, repeat the CSP MP-20 Linear-ODE run with 5 seeds to confirm the 69.83% match rate is stable; if it varies by more than ~2 points, the 'across all datasets' claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The Section I.B claim requires that OMatG outperforms DiffCSP, FlowMM, and MatterGen on the evaluated tasks. The evidence does not support the strongest reading of this claim: no single OMatG configuration wins all metrics. In Table III, the best S.U.N. rate is Linear SDE (22.48%), the best stability rate is Enc-Dec (60.04%), and the best novelty is CFP+CSP (79.08%); MatterGen-MP has the best RMSD (0.1038 Å vs. best OMatG 0.4187 Å), and in Table II DiffCSP has higher structural validity (99.91% vs. best OMatG 99.04%). The SOTA claim is thus true only if 'outperform' is read as 'some configuration wins the headline metric for that task.' More importantly, every reported OMatG number is a single run selected from 27-32 Bayesian HPO trials per interpolant (Appendix C.3). The margins that carry the DNG-SOTA claim are small: S.U.N. 22.48 vs. MatterGen-MP 20.30, a 2.18-point gap with no seed variance reported. If the HPO selection procedure overfits to the validation metric or the single seed is favorable, the central claim is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces OMatG, a generative model for inorganic crystal structures built on stochastic interpolants (SI), jointly modeling lattice vectors, fractional coordinates, and atomic species (via discrete flow matching). It extends SI to periodic boundary conditions by an unwrap–interpolate–wrap procedure, and evaluates the method on crystal structure prediction (CSP) on perov-5, MP-20, MPTS-52, and Alex-MP-20, and on de novo generation (DNG) on MP-20, claiming a new state of the art against DiffCSP, FlowMM, and MatterGen-MP. The paper also releases code, recomputes baseline metrics with SMACT 3.0, and adds a new average-coordination-number metric and a MatterSim/DFT stability pipeline.","tokens_in":33146,"tokens_out":6603,"duration_ms":62945,"significance":"The paper makes a valuable contribution to the materials-generation literature by demonstrating the practical flexibility of the stochastic-interpolants framework, with an open-source implementation and careful attention to evaluation (recomputed baselines, updated SMACT rules, DFT relaxation for stability metrics, and a new coordination-number metric). If the technical concerns below are resolved, the work would constitute a strong benchmark contribution. However, the headline state-of-the-art claim is currently not robustly supported: it rests on single-run, best-of-hyperparameter-optimization results with no seed variance, and no single OMatG configuration dominates across metrics. The torus extension of SI is also asserted without proof, leaving the soundness of the training objective unresolved for nonlinear and stochastic interpolants.","major_comments":[{"comment":"The central claim that OMatG outperforms DiffCSP, FlowMM, and MatterGen in DNG is not established as stated, because no single OMatG configuration dominates across the reported metrics: MatterGen-MP achieves a markedly lower RMSD (0.1038 Å versus the best OMatG value of 0.4187 Å in Table III), DiffCSP achieves higher structural validity (99.91% versus the best OMatG 99.04% in Table II), and the best OMatG S.U.N. rate (22.48%) exceeds MatterGen-MP (20.30%) by only 2.18 points without any seed variance. Since each configuration is selected from 27–32 Bayesian HPO trials per interpolant (Appendix C.3) and reported once, the observed margin could be favorable noise. Please report multiple seeds with standard errors for the headline metrics, or revise the claim to a per-metric best-configuration statement.","section":"Section I.B, Table III, Appendix C.3"},{"comment":"The unwrap–interpolate–wrap procedure for periodic boundary conditions is asserted to extend stochastic interpolants to the torus, including the latent term γ(t)z, but no proof is given that the wrapped stochastic process x_t = wrap(α(t)x0 + β(t)x1' + γ(t)z) satisfies the SI velocity or Fokker–Planck equations on the torus. For the linear interpolant with γ=0 this reduces to the FlowMM geodesic, but for nonlinear interpolants (trig, enc-dec, VP/VE SBD) and for SDE sampling with γ>0 the path crosses the boundary and the time derivative of the wrapped path is not simply the wrapped derivative; hence the conditional velocity in Eq. (2) may not be the correct one for the target density on the torus. Please provide a proof or direct numerical validation (e.g., exact density recovery on a known toroidal target); without this, the training objective is not shown to be sound.","section":"Section III.B.1, Appendix B.5, Eq. (2)"},{"comment":"Velocity annealing, which rescales the learned velocity as b_theta(t,x) → (1+st)b_theta(t,x), is applied with large coefficients (e.g., s=14 for positions in Table VII) and is acknowledged to lack theoretical justification. Because this ad-hoc modification changes the generative ODE/SDE dynamics, the reported SOTA results cannot be attributed solely to the SI framework, and the rescaling may invalidate the marginal density guaranteed by the SI construction. Please provide an ablation with s=0 or a demonstration that the annealed flow still samples the target distribution, or treat velocity annealing as a core component of the method and provide a theoretical or systematic empirical justification.","section":"Appendix B.7, Table VII"}],"minor_comments":[{"comment":"There is a formatting error in the DiffCSP row of the MPTS-52 columns: the text reads \"15.79 / 14.290.1533/ 0.1489\" without a separator between the match rate and RMSE; this should be corrected.","section":"Table I"},{"comment":"The manuscript repeatedly refers to the \"four-dimensional torus\" for fractional coordinates in three dimensions; please clarify whether a three-dimensional torus is intended or explain the extra dimension.","section":"Section III.B.1 and Appendix B.5"},{"comment":"The column header \"⟨E⟩/N(↓) RMSD\" is ambiguous; the energy above hull is reported in eV/atom and should be labeled clearly, with the RMSD unit (Å) separated.","section":"Table III"},{"comment":"The manuscript does not state explicitly how the single OMatG configuration shown in Tables II and III is chosen (e.g., best validation evalDNG, or best S.U.N. on a hold-out set), nor whether any test-set information was used in the selection; please make this selection rule explicit.","section":"Section V.C and Appendix C.3"},{"comment":"The phrase \"score based diffusion\" should be hyphenated as \"score-based diffusion\".","section":"Figure 6 caption"}],"recommendation":"major_revision","confidential_remarks":"The two load-bearing issues are the statistical support for the SOTA claim and the missing torus proof for the periodic SI extension. Both are addressable: multiple-seed reporting or a toned-down claim, and either a proof or a clear validation of the wrapped process. The velocity-annealing concern is also important but can be handled with ablations and a slightly more cautious framing. If these are fixed, the paper would be a solid contribution to the materials-generation literature with reproducible code and careful benchmarking."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nYou should know two things about this paper. It is the first serious application of stochastic interpolants to crystal generation, and it ships with unusually thorough evaluation: recomputed baselines under SMACT 3.0, new metrics like average coordination number, and MatterSim+DFT stability checks. The central empirical finding—that the choice of interpolant and sampling scheme matters a lot for CSP/DNG benchmarks—is real and useful. But the headline SOTA claim is too strong for what the tables show.\n\nWhat's actually new: OMatG unifies diffusion and flow matching in a single tunable framework, extends SI to periodic fractional coordinates via unwrap-interpolate-wrap, couples continuous flows with discrete flow matching for species, and adds a learned equivariant encoder. The appendix documentation is careful, including hyperparameter sweeps and ablation tables. The code is public. This is the kind of work the field needs more of.\n\nWhere it gets soft. First, no single OMatG configuration wins all metrics. On CSP, OMatG beats match rates on perov-5, MP-20, MPTS-52, but often at the cost of much higher RMSE (e.g., perov-5: 0.083 vs 0.077 for DiffCSP; MP-20: 0.074 vs 0.063). On DNG, MatterGen-MP has lower RMSD (0.104 Å vs best OMatG 0.419) and DiffCSP higher structural validity (99.91 vs 99.04). So the phrase “outperforming ... under existing, revised, and new performance measures” is simply not true as stated; it should be qualified to “better match/S.U.N. rates on most datasets, with some fidelity trade-offs.”\n\nSecond, every reported number is a single run selected from 27–32 Bayesian HPO trials per interpolant. No seed variance is given. The S.U.N. margin over MatterGen-MP is only 2.18 points (22.48 vs 20.30, Table III). That's within plausible run-to-run noise. This needs error bars or at least multiple seeds for the headline configurations.\n\nThird, the periodic-boundary SI extension is asserted without proof. The unwrap-interpolate-wrap recipe is reasonable, and for linear interpolants it reduces to the geodesic used by FlowMM. But for nonlinear and stochastic interpolants, the paper does not show that the wrapped process has the correct marginal density or that the learned velocity generates the intended distribution on the torus. A proof sketch or a citation to a theorem covering this case would settle it.\n\nThe velocity annealing is acknowledged to lack theoretical justification, which is honest, but it remains a strong free parameter that could inflate apparent gains.\n\nWho this is for: anyone working on generative models for crystal structures. The ablation study of interpolants is genuinely informative. The paper deserves a serious referee and likely publication after revision, provided the authors qualify the SOTA claim, add uncertainty estimates, and address the torus question. I'd send it; I'd also tell the authors to cool the abstract.","headline":"A careful, well-verified framework for crystal generation whose genuine contribution—interpolation choice matters—is undercut by an overstated SOTA claim and missing error bars.","tokens_in":33737,"tokens_out":3632,"would_cite":true,"duration_ms":35316,"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":"OMatG applies stochastic interpolants to inorganic crystal generation and claims state-of-the-art results over diffusion and flow-matching baselines.","keywords":["stochastic interpolants","generative modeling","crystal structure prediction","de novo materials generation","equivariant graph neural networks","periodic boundary conditions","discrete flow matching"],"falsifier":"On a one-dimensional torus, take a uniform base $\\rho_0$, a target density whose exact form is known, and a nonlinear or stochastic interpolant with $\\gamma(t) > 0$; numerically integrate the learned ODE/SDE and compare the empirical endpoint density with the exact wrapped target marginal. A mismatch would mean the wrapped process is not sampling the target distribution, so the training objective itself is unsound.","tokens_in":32608,"feed_emoji":"🧪","tokens_out":9624,"duration_ms":87768,"temperature":0.7,"pith_summary":"This paper introduces OMatG, a generative framework for inorganic crystals built on stochastic interpolants, and argues that the extra design freedom of interpolants is what lets it beat existing diffusion and flow-matching models. The central idea is to bridge a random base structure and a target crystal through a tunable path $x_t=\\alpha(t)x_0+\\beta(t)x_1+\\gamma(t)z$, with periodic boundary conditions handled by interpolating along the shortest geodesic in unwrapped coordinates and wrapping back, and atomic species handled by discrete flow matching. The paper claims that this framework achieves a new state of the art on crystal structure prediction (CSP) and de novo generation (DNG), outperforming both DiffCSP and FlowMM on CSP and DNG, and MatterGen on DNG, across the four benchmark datasets considered. If true, the practical payoff is that materials discovery pipelines can treat the generative process itself as tunable rather than fixed to one diffusion or flow scheme.","feed_headline":"Tunable interpolants beat diffusion and flow models on crystals","feed_subtitle":"A single stochastic-interpolant framework claims new best match and stability rates for crystal structure prediction and de novo generation.","key_machinery":"The load-bearing object is the stochastic interpolant $x_t = \\alpha(t)x_0 + \\beta(t)x_1 + \\gamma(t)z$, with loss functions that learn the velocity $b_\\theta(t,x)$ and, for SDE sampling, the denoiser $z_\\theta(t,x)$. The interpolation coefficients are flexible except for endpoint constraints, and the paper tests linear, trigonometric, encoder-decoder, and score-based-diffusion (VP/VE) interpolants, with ODE or SDE sampling and a tunable latent-noise amplitude $\\gamma(t)$. For periodic fractional coordinates, the path is defined by unwrapping $x_1$ to its closest periodic image, interpolating in Euclidean space along the shortest geodesic, and wrapping the result back onto the torus; atomic species are generated by discrete flow matching through a continuous-time Markov chain. This mechanism is what lets the same code reproduce conditional flow matching (linear interpolant, ODE, $\\gamma=0$) and score-based diffusion (VP/VE interpolants with Gaussian base) while also exploring new in-between schemes.","core_discovery":"OMatG is described as the first implementation of stochastic interpolants for inorganic crystal generation, and as such it unifies conditional flow matching and score-based diffusion as special cases inside a single, tunable framework. The paper's central empirical claim is that by optimizing the interpolation functions for the fractional coordinates and lattice vectors separately, and by coupling the continuous flow with discrete flow matching for atomic species, OMatG reaches a new state of the art: higher structural match rates than DiffCSP and FlowMM on perov-5, MP-20, and MPTS-52, the first CSP baseline on Alex-MP-20, and higher S.U.N. (stable, unique, novel) rates than MatterGen-MP in de novo generation. The best OMatG models also generate structures with lower average energy above the convex hull and closer match to reference coordination-number and crystal-system distributions. The authors interpret the results as evidence that the choice of interpolant, latent noise, sampling scheme, and per-degree-of-freedom tuning is a decisive axis of model design for materials generation.","pith_inferences":["The periodic-boundary construction, based on the same unwrap-interpolate-wrap recipe, could carry over to other periodic or compact manifolds such as angle variables or SO(3) orientations, although the paper does not develop that generalization.","If the match-rate/RMSE trade-off observed on perov-5 is generic, then match rate alone may over-reward models that find the right chemical environment but miss exact symmetric sites; a combined metric penalizing both errors would sharpen future comparisons.","The framework's arbitrary base distributions suggest a direct way to inject prior knowledge, such as LLM-suggested compositions or machine-learned-potential relaxed motifs, as the base $\\rho_0$, which the paper only begins to explore with LLM-generated structures."],"forward_implications":["If the SOTA claim holds, any new generative model for crystals should be compared against OMatG under the same refined metrics, including validity filters and the coordination-number benchmark, before claiming improvement.","The reported CSP gains on MP-20 and MPTS-52 mean that specified-composition structure prediction can be run at generative-model speed with better starting candidates for downstream DFT relaxation.","The DNG improvements in stability and S.U.N. rate mean that a larger fraction of generated structures should survive DFT screening, cutting the cost of de novo materials discovery pipelines.","Because OMatG's linear/ODE setting reproduces FlowMM and its SBD interpolants reproduce diffusion models, the framework gives a single code base for fair comparisons and ablations across previously separate methods."],"supporting_citations":[{"why":"Supplies the stochastic-interpolant framework: the interpolation formula, the velocity/denoiser loss functions, and the ODE/SDE sampling schemes that OMatG extends to crystals.","marker":"[33]"},{"why":"Provides the DiffCSP baseline, the equivariant graph encoder used in OMatG, and the periodic diffusion treatment for fractional coordinates.","marker":"[30]"},{"why":"Provides the FlowMM baseline, the shortest-geodesic periodic interpolation convention, the log-normal lattice base distribution, and velocity annealing.","marker":"[32]"},{"why":"Provides the MatterGen DNG baseline and the S.U.N. evaluation protocol used to measure stability, uniqueness, and novelty.","marker":"[31]"},{"why":"Supplies the discrete flow matching machinery, including the conditional rate matrix, for generating atomic species as a continuous-time Markov chain.","marker":"[44]"},{"why":"Supplies the Riemannian flow-matching geodesic formulation that motivates the shortest-path interpolation on the flat torus.","marker":"[42]"},{"why":"Defines the original crystal generation benchmarks and evaluation metrics, including match rate and coverage, that OMatG refines.","marker":"[29]"},{"why":"Supplies the composition-rules checker used to define compositional validity in the revised metrics.","marker":"[45]"},{"why":"Supplies the machine-learned potential used for the first structural relaxation before DFT stability calculations.","marker":"[54]"},{"why":"Supplies the StructureMatcher routine used to compute match rates and RMSE for the CSP benchmarks.","marker":"[52]"}],"fun_headline_variants":["Stochastic interpolants unify diffusion and flow for crystal generation","OMatG: one interpolant framework rules materials generation","Stochastic interpolants set new bar for crystal generation","OMatG's tunable stochastic interpolants outperform dedicated models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that interpolating fractional coordinates in unwrapped Euclidean space along the shortest periodic image and then wrapping back onto the torus gives the correct marginal distribution on the torus for nonlinear and stochastic interpolants with $\\gamma(t)>0$; the paper uses this construction for all interpolants but proves it only for the linear geodesic case.","fun_headline_variants_meta":{"raw":{"variants":["Stochastic interpolants unify diffusion and flow for crystal generation","OMatG: one interpolant framework rules materials generation","Stochastic interpolants set new bar for crystal generation","OMatG's tunable stochastic interpolants outperform dedicated models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000539,"raw_usage":{"total_tokens":2620,"prompt_tokens":1015,"completion_tokens":1605,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":631,"completion_tokens_details":{"reasoning_tokens":1537}},"tokens_in":631,"tokens_out":1605,"duration_ms":11187,"temperature":1.0,"reasoning_tokens":1537,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T11:41:31.591551+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a one-dimensional torus, take a uniform base $\\rho_0$, a target density whose exact form is known, and a nonlinear or stochastic interpolant with $\\gamma(t) > 0$; numerically integrate the learned ODE/SDE and compare the empirical endpoint density with the exact wrapped target marginal. A mismatch would mean the wrapped process is not sampling the target distribution, so the training objective itself is unsound.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the DiffCSP baseline, the equivariant graph encoder used in OMatG, and the periodic diffusion treatment for fractional coordinates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the composition-rules checker used to define compositional validity in the revised metrics."}],"review_version":1}