{"id":"30ac51f7-5f7b-4fe1-85c3-0cbc37204ef0","arxiv_id":"2608.13457","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"SbCD is a crystal diffusion model that learns space-group transitions with a Markovian jump process and generates complete crystallographic specifications from a P1 prior.","lead":"This paper introduces SbCD, a diffusion model that generates crystals by learning transitions between space groups instead of sampling them from a fixed empirical list. It reports improved composition validity and predicted stability on two materials datasets compared with its symmetry-preserving baseline.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Generated 'complete specifications' are never checked for internal consistency: F is not constrained to the declared Wyckoff positions, so the central claim may be vacuous.","rationale":"The strongest claim is not merely good generative quality; it is that SbCD produces complete crystallographic specifications. The single most load-bearing condition is therefore that each output triple (G,S,F) is self-consistent. The paper's own Appendix E concedes that no mechanism enforces Wyckoff constraints on F. The theoretical bound in Proposition 4.1 does not repair this: even if the reverse-bridge factorization in Eq. (10) were correct, a factorized Pθ_F would not couple F to S unless implemented, and the implementation is exactly an unconstrained DiffCSP coordinate head with no special-position loss. I agree with the reader that the reverse-bridge factorization is an unsupported assumption and deserves testing, but the empirical consistency audit is more decisive for the central 'complete specifications' claim. If the audit passes, the factorization concern is downgraded to a theory-repair issue; if it fails, the headline is unsupported regardless of the ELBO. The paper has real strengths: the masked continuous kernel, the interpolated discrete prior, the analytical jump kernels, and the positive ablations. I therefore do not recommend rejection; I recommend keeping the conditional verdict and adding the consistency test as an explicit condition.","tokens_in":26876,"tokens_out":11822,"duration_ms":140881,"concrete_test":"Run a consistency audit on the 10,000 MP-20 and 10,000 MPTS-52 SbCD samples. For each sample, reconstruct the unit cell exactly as Algorithm 1 returns it, using the declared (G0, S0, F0, A0, k0); then run a symmetry finder (spglib or PyXtal's get_symmetry) with no knowledge of G0/S0. Report: (1) fraction of cells whose detected space group equals the declared G0; (2) fraction of atoms whose detected site symmetry equals the declared S0; (3) mean/median fractional-coordinate distance (with PBC) from F0 to the assigned Wyckoff manifold returned by search_closest_wp; (4) fraction of cells whose atomic multiplicities are consistent with the declared Wyckoff positions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"SbCD's headline claim is that it outputs complete crystallographic specifications: a space group G, site symmetries S, and coordinates F consistent with them. The training objective (Algorithm 2) has no term forcing F0 to lie on the fixed-point subspace of the declared site-symmetry group S, and the sampling pipeline (Algorithm 1, line 9) generates F with DiffCSP's unconstrained wrapped-Gaussian diffusion before assigning a Wyckoff position only at the end via PyXtal's search_closest_wp (Appendix B.3). The authors themselves state in Appendix E that enforcing 'dynamically varying symmetry constraints on fractional coordinates' is a non-trivial open problem left for future work. Consequently, a generated (G,S,F) triple can be internally inconsistent: a representative coordinate off the Wyckoff manifold has trivial stabilizer, so the orbit multiplicity and declared site symmetry are wrong even though the replicated unit cell may still be G-periodic. None of the reported metrics (structure/composition validity, coverage, dsg, S.U.N., Ehull) checks per-sample agreement between F0 and the declared G0/S0. The claim of producing complete crystallographic structure specifications therefore outruns the evidence; whether the model has learned to place atoms on special positions is an empirical question the paper does not test.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SbCD, a generative model for crystals that jointly models the space group, site symmetries, lattice, atom types, and fractional coordinates of an asymmetric unit. The key idea is a Markovian jump-diffusion framework in which a forward process corrupts a crystal by repeatedly transitioning its space group to the lowest-symmetry group P1, while the reverse process learns to restore symmetry by generating space groups, site symmetries, and the associated structural components. The authors derive a variational upper bound on the negative log-likelihood, split into continuous and discrete diffusion terms, and provide analytical forward/reverse kernels for the lattice and site-symmetry processes under symmetry breaking. On MP-20 and MPTS-52, SbCD outperforms its symmetry-preserving counterpart and several baselines in stable-structure generation and space-group distribution matching. The paper's central claim is that SbCD is the first framework to generate complete crystallographic structure specifications, including space groups and Wyckoff positions.","tokens_in":1734,"tokens_out":1877,"duration_ms":101105,"significance":"If the central claims are borne out, SbCD is a useful step: it replaces empirical space-group sampling with a learned distribution, unifies continuous and discrete diffusion objectives, and demonstrates consistent gains over a symmetry-preserving baseline. The paper is also careful to provide analytical forms and explicit training and sampling algorithms, and the ablation against the no-symmetry-breaking variant supports the contribution of the symmetry-breaking mechanism. However, the central claim about complete crystallographic specifications is not verified: the evaluation never checks whether generated coordinates actually respect the declared site symmetry and space group. The theoretical bound also rests on an unverified factorization assumption. These issues do not invalidate the empirical method, but they need to be addressed before the paper's claims can be accepted.","major_comments":[{"comment":"The paper's central claim that SbCD produces complete crystallographic structure specifications is not supported by the evaluation. None of the metrics in Tables 2 and 3 checks whether a generated fractional coordinate F0 lies on the Wyckoff manifold of the declared site-symmetry group S0, or whether the replicated structure actually has the declared space group G0. Algorithm 1 line 9 generates F with the unconstrained DiffCSP diffusion, and Appendix B.3 assigns a Wyckoff position only a posteriori via search_closest_wp; Appendix E explicitly states that enforcing dynamic symmetry constraints on fractional coordinates is left to future work. Consequently, a generated (G,S,F) triple can be internally inconsistent, and the 'complete specification' claim outruns the evidence. I request either a per-sample consistency check (e.g., using PyXtal or Spglib to determine the actual space group of the generated structure and comparing it with the declared G, or checking that F0 lies on the appropriate Wyckoff subspace) or a softened claim that does not assert full crystallographic consistency.","section":"Section 5 / Appendix B.3 / Appendix E"},{"comment":"The variational bound in Proposition 4.1 assumes that the exact reverse bridge measure factorizes as Q_G times Q_{k|G} times Q_{S|G} times Q_A times Q_F. For crystalline asymmetric units, this factorization is questionable because the fractional coordinates F are constrained by the site symmetry S and the space group G: a Wyckoff position determines the support and symmetry constraints on F. If Eq. (10) fails, the KL decomposition in Eq. (11) omits non-negative mutual-information terms, so the claimed upper bound on the negative log-likelihood is not established. The proof simply states 'We assume...' without justification. This is load-bearing for the theoretical contribution; the authors should either prove the bound under a weaker assumption, add the missing terms, or reframe the objective as a heuristic approximation supported by empirical validation.","section":"Appendix D.1, Eqs. (10)-(11)"},{"comment":"The reverse sampling step for the lattice parameters is inconsistent between Algorithm 1 and Proposition 4.4. Algorithm 1 line 7 gives k_{t-1} = m⊙( sqrt(alpha_bar_{t-1})/(1-alpha_bar_t)(1-alpha_t) times the predicted k_0 + m⊙ sqrt(alpha_t)/(1-alpha_bar_t)(1-alpha_bar_{t-1}) k_t + m_b + sigma_t epsilon ) + m_b, whereas Proposition 4.4 states the same expression but with the mask and bias of the target higher-symmetry space group G_{t-1} (denoted em and em_b) instead of the current group G_t. Using m in place of em and adding m_b instead of em_b means the sampled lattice is projected onto the wrong crystal family (the one being left, not the one being entered). If the implementation follows Algorithm 1, the method differs from the described model; if it follows Proposition 4.4, the pseudocode is incorrect. This must be fixed for reproducibility.","section":"Algorithm 1 line 7 vs. Proposition 4.4"}],"minor_comments":[{"comment":"The notation in Figure 1 is confusing: the right panel uses 'g' and 'g' (presumably 'g' with a tilde) to denote the marginal site-symmetry priors of P2/m and P1, but the typography makes the distinction nearly invisible; please use a clearer notation such as g0 and g1.","section":"Figure 1"},{"comment":"In the statement of Proposition 4.4, the definition of sigma_t contains mixed notation: it is written as sqrt((1-alpha_bar_{t-1})(1-alpha_t)/(1-alpha_bar_t)) with both alpha_t and alpha_bar_t in the same expression; please clarify the noise schedule notation consistently with the rest of the paper.","section":"Section 4.3 / Proposition 4.4"},{"comment":"The sampling of the space-group trajectory in step 4 is ambiguous: 'G_t ∼ q_{t|0}(G_t|G_0) → G_{t+1} ∼ Π_t(G_t, G_{t+1})' suggests a two-stage draw, but it is not clear how this corresponds to the continuous-time jump-diffusion loss in Lemma 2.1. Please specify the exact distribution used for the pairs (G_t, G_{t+1}).","section":"Algorithm 2, step 4"},{"comment":"The column header 'ASG' is not defined in the text; it appears to mean asymmetric-unit space-group representation, but this should be stated explicitly.","section":"Table 1"},{"comment":"The list of oriented site-symmetry symbols is printed as a dense dictionary with integer values; presenting it as a table with the symbol names would aid readability.","section":"Appendix B.1"}],"recommendation":"major_revision","confidential_remarks":"The main concern is that the 'complete crystallographic specification' claim is not tested, and the theoretical bound relies on an unverified factorization. Both are fixable: add a per-sample consistency analysis, address the factorization assumption (or reframe the theory), and correct the Algorithm 1 typo. If the empirical implementation actually used the algorithm as written, the reported results may not reflect the intended symmetry-breaking process, so the authors should clarify which formula was used."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take: this is a real methodological contribution—applying continuous-time Markov jump diffusion to space-group transitions in crystal generation, with analytically derived symmetry-breaking kernels for lattice and site-symmetry priors. The ablation against the no-symmetry-breaking variant shows a clear effect, and the paper is honest about its main limitation. But the abstract and Table 1 overstate what is actually delivered. The model does not generate internally consistent (G,S,F) triples, and no metric checks that.\n\nWhat is good: the variational bound in Prop 4.1, the diagonalized rate matrix in Prop 4.2, the posterior holding-time sampling in Cor 4.3, and the interpolated site-symmetry prior in Prop 4.5 are all careful and look correct. The one-hot site-symmetry representation over 81 oriented symbols is a nice simplification. The model clearly helps: SbCD beats its symmetry-preserving counterpart on composition validity and stability, and the low-NFE results are practically useful.\n\nWhere it falls short: the central claim—”first framework capable of producing complete crystallographic structure specifications, including space groups and Wyckoff positions”—outruns the evidence. In Algorithm 1, F is generated by DiffCSP’s unconstrained diffusion, then a Wyckoff position is attached at the end via PyXtal’s search_closest_wp. Nothing forces F0 to lie on the fixed-point subspace of the declared S0. The authors themselves say in Appendix E that enforcing dynamically varying symmetry constraints on fractional coordinates is an unsolved problem. So a generated (G,S,F) triple can be internally inconsistent, and none of the reported metrics (validity, coverage, Ehull, S.U.N.) checks that. The factorization of the reverse bridge in Appendix D.1 Eq (10) is also an assumption; if the posterior couples components, the KL bound is only approximate. That is common in this literature, but worth flagging.\n\nAdditionally, baselines in Tables 2 and 3 have no error bars; only SbCD has repeated runs. No code or artifacts are released, which makes the “complete specification” claim even harder to verify.\n\nBottom line: this is a credible, careful step toward learning space-group distributions rather than sampling them empirically. The method is solid enough to warrant serious refereeing, but the paper needs either (a) a consistency check on generated triples—measure how often the assigned Wyckoff position actually contains the sampled fractional coordinate—or (b) a tempered claim that stops short of “complete specifications”. I would send it to review, but I would tell the authors the claim needs to be backed or softened.\n\nHope that is useful.","headline":"Novel jump-diffusion machinery for space-group transitions, but the headline claim of complete crystallographic specifications is not actually validated; Appendix E concedes the coordinate constraint is open.","tokens_in":27602,"tokens_out":2436,"would_cite":true,"duration_ms":25600,"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":"The paper claims that a diffusion model can output a crystal's complete symmetry specification—space group and Wyckoff positions included—by reversing learned symmetry-breaking jumps from the lowest-symmetry group P1, and that this…","keywords":["crystal generation","diffusion models","space groups","Wyckoff positions","symmetry breaking","Markovian jump diffusion","site symmetry","materials discovery"],"falsifier":"Estimate the true log-likelihood of held-out crystals by importance sampling over the complete reverse path and compare it with the variational bound used in training; a systematic gap that grows with sample size would show the assumed factorization is not met. A cheaper check is to condition the atom-type predictor on the predicted space group and see whether validation likelihood improves.","tokens_in":26698,"feed_emoji":"💎","tokens_out":9150,"duration_ms":90484,"temperature":0.7,"pith_summary":"The paper proposes SbCD, a diffusion framework that generates crystals complete with their symmetry specifications: the space group and each atom's Wyckoff position are produced by the model rather than sampled from the training distribution. Its central move is to run the forward process as a symmetry-breaking chain that sends every crystal's space group through random jumps down to P1, the lowest-symmetry group, and to learn the reverse process that climbs back up. Because the space group changes during diffusion, the lattice and site-symmetry constraints change with it; the paper derives closed-form forward and reverse kernels for both continuous and discrete components. On de novo generation over MP-20 and MPTS-52, the model beats its symmetry-preserving counterpart on key metrics including thermodynamic stability, composition validity, and space-group distribution distance. If correct, this is the first generative framework whose output includes the full crystallographic specification, which matters because downstream property prediction and synthesis planning depend on symmetry.","feed_headline":"Diffusion model learns crystal space groups, not copies them","feed_subtitle":"It starts from minimal symmetry and climbs back up, matching or beating symmetry-preserving models on stability.","key_machinery":"The central object is a Markovian jump diffusion—a continuous-time Markov process that holds a state for a random duration and then jumps—run over the 230 crystallographic space groups, with a forward rate matrix that sends every group to the absorbing lowest-symmetry group P1. Around that jump chain, two adaptive kernels carry the argument: a masked Gaussian diffusion over the six-dimensional lattice representation (mask–bias pairs per crystal family switch when the space group jumps), and a categorical diffusion over site symmetry whose one-hot labels are the 81 distinct oriented site-symmetry symbols, with a prior that interpolates linearly between the old and new space group's site-symmetry priors. The reverse process starts at P1 with trivial site symmetry and samples space-group jumps via τ-leaping, then uses the sampled space group to choose the lattice mask and site-symmetry prior for the next step.","core_discovery":"On its own terms, the paper claims that generating a complete crystal specification—not just atoms in a cell but the space group and the Wyckoff positions of its sites—can be learned, without an empirical space-group prior, by training a reverse Markovian jump-diffusion process that starts at the lowest-symmetry space group P1. The forward process jumps any space group down to P1 at a constant rate; the reverse process learns the jump rates back up, and the space group's evolution changes the constraints applied to the lattice and site symmetries. Closed-form kernels are derived for the discrete space-group chain, the continuous lattice under mask–bias constraints, and a categorical chain over 81 site-symmetry symbols. On the MP-20 and MPTS-52 benchmarks, the claimed result is that SbCD variants outperform their symmetry-preserving counterpart across structural and compositional validity, property statistics, and thermodynamic stability, matching or approaching models that condition on empirical space-group distributions.","pith_inferences":["Beyond the paper: the variational bound's assumed factorization could be stress-tested by importance-sampling the true likelihood; a large gap would point to coupling between the space group and the atom-type or coordinate components.","Beyond the paper: because every forward jump goes directly to P1, the model only learns one-step symmetry breaking; intermediate transitions through subgroups are a concrete extension that could make generated paths physically more realistic.","Beyond the paper: the 81-symbol site-symmetry encoding is separable from the jump-diffusion machinery and could be reused by any asymmetric-unit crystal generator.","Beyond the paper: the fixed-window ablation suggests the symmetry-breaking schedule is a tunable trade-off between structural validity and crystal density, which could be exploited for density-targeted screening."],"forward_implications":["Generated crystals can carry space groups and Wyckoff positions that are rare or absent in the training distribution, because generation no longer requires an empirical space-group prior.","The learned inter-space-group transitions make few-step generation efficient: validity at 50 function evaluations is over 65 percent for SbCD against about 25 percent for its symmetry-preserving counterpart.","Site-symmetry learning over 81 symbols removes the need to project generated symmetries onto the nearest valid point group, which stabilizes training and downstream Wyckoff assignment.","The posterior holding-time formula gives an adaptive schedule for when symmetry breaking occurs during generation; fixed windows produce higher structural validity but poorer density statistics.","Lattices and site symmetries are constrained exactly at every generation step by the mask–bias pair of the current space-group family, so sampled crystals respect crystallographic constraints without external templates."],"supporting_citations":[{"why":"Provides the rate-matrix and Kolmogorov-equation foundations for continuous-time Markov chains used in the space-group jump process.","marker":"[12]"},{"why":"Supplies the continuous-time variational bound and reverse-rate estimation used to learn the space-group jump diffusion.","marker":"[5]"},{"why":"Provides the categorical diffusion framework with marginal priors that the site-symmetry kernels extend.","marker":"[2]"},{"why":"Provides the noise schedule and continuous-state diffusion basis for the lattice component.","marker":"[17]"},{"why":"Supplies the polar-decomposition lattice representation and mask–bias constraints that SbCD adapts to symmetry-breaking transitions.","marker":"[20]"},{"why":"The symmetry-preserving asymmetric-unit baseline whose site-symmetry representation and results SbCD directly compares against and replaces.","marker":"[23]"},{"why":"Provides the equivariant fractional-coordinate diffusion and denoiser architecture used for the coordinate component.","marker":"[19]"},{"why":"Provides the DDIM reverse-sampling parameterization used in the continuous-state lattice kernel.","marker":"[35]"}],"fun_headline_variants":["Diffusion model learns space groups, not copies them","Symmetry-breaking diffusion outperforms symmetry-preserving in crystals","Generating crystals from P1 up: no space-group priors needed","SbCD: reverse Markovian jumps generate full crystal symmetry"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training loss is a valid upper bound on the model's log-likelihood only if the reverse generative process factorizes into independent components (space group, lattice and site symmetry given the space group, atom types, coordinates); if those components actually interact during reversal, the model may be optimizing something other than the data likelihood.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model learns space groups, not copies them","Symmetry-breaking diffusion outperforms symmetry-preserving in crystals","Generating crystals from P1 up: no space-group priors needed","SbCD: reverse Markovian jumps generate full crystal symmetry"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000146,"raw_usage":{"total_tokens":1173,"prompt_tokens":930,"completion_tokens":243,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":174}},"tokens_in":546,"tokens_out":243,"duration_ms":3635,"temperature":1.0,"reasoning_tokens":174,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:35:37.366344+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Estimate the true log-likelihood of held-out crystals by importance sampling over the complete reverse path and compare it with the variational bound used in training; a systematic gap that grows with sample size would show the assumed factorization is not met. A cheaper check is to condition the atom-type predictor on the predicted space group and see whether validation likelihood improves.","supporting_citations":[{"cited_title":"On the theory of stochastic processes, with particular reference to applications","cited_arxiv_id":null,"evidence_quote":"Provides the rate-matrix and Kolmogorov-equation foundations for continuous-time Markov chains used in the space-group jump process."},{"cited_title":"A continuous time framework for discrete denoising models","cited_arxiv_id":null,"evidence_quote":"Supplies the continuous-time variational bound and reverse-rate estimation used to learn the space-group jump diffusion."},{"cited_title":"SymmCD: Symmetry-preserving crystal generation with diffusion models","cited_arxiv_id":null,"evidence_quote":"The symmetry-preserving asymmetric-unit baseline whose site-symmetry representation and results SbCD directly compares against and replaces."}],"review_version":1}