{"id":"3a78753d-bc29-4aa2-a8e3-6af81e56af8e","arxiv_id":"2502.02016","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"CrysBFN adapts Bayesian Flow Networks to periodic crystal coordinates via von Mises distributions and entropy conditioning, achieving SOTA generation and 100x faster sampling.","lead":"CrysBFN is a new generative model that creates crystal structures by treating atomic positions as points on a circle and using a Bayesian flow with an entropy-based control signal. It reports state-of-the-art quality on standard benchmarks and a roughly 100-fold sampling speedup over the leading diffusion baseline.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The linear-entropy schedule in §4.1/Appendix A.5 matches expected c to a scalar c(t) rather than expected H(c), so the claimed linear-entropy objective is not actually enforced; a Monte Carlo check would settle whether this matters.","rationale":"The reader's verdict of CONDITIONAL is appropriate. The strongest claim is empirical SOTA plus 100x speedup; the strongest theoretical novelty is the periodic Bayesian flow with entropy conditioning. The load-bearing assumption is that the schedule actually delivers the entropy behaviour the method relies on. The paper's own Appendix A.5 defines the objective as linear expected entropy but solves a different, mean-field problem (matching a scalar H(c(t_i)) and a one-step expected concentration). Since H is concave in c, Jensen gives E[H(c)] ≤ H(E[c]), so the implemented schedule systematically makes the receiver more confident than the stated linear schedule, at least if the mean-field recursion is accurate. The 'arbitrarily selected x' is not by itself fatal because of rotational symmetry when m_{i-1}=x, but it is indicative that the derivation is not tracking the true random flow. This is a real gap in the theoretical argument, but it does not invalidate the method: the empirical results, released code, and the ablation showing the schedule matters provide independent support. The secondary issues (overclaim of SOTA on COV-R in Table 1, notational off-by-one in Eq. (18)/Algorithm 1, missing error bars on ab initio metrics) reinforce the conditional verdict but are not the primary attack. The Monte Carlo test above would settle whether the entropy-schedule gap is large enough to matter.","tokens_in":30382,"tokens_out":20007,"duration_ms":192614,"concrete_test":"Using the released code/schedule (e.g., MP-20, β1=1000, n=1000), Monte Carlo simulate c_i from Eq. (16) for i=1..n at a fixed x (and a few x values), and compute E[H(c_i)] and H(E[c_i]). Compare both to the linear target (1−t_i)H(0)+t_iH(1). If the max relative deviation of E[H(c_i)] from the linear target exceeds ~10% of the entropy range, the schedule does not realize the claimed linear entropy. Then re-run the Table 3 ablation replacing the published schedule with one that exactly linearizes E[H(c_i)] (found by Monte Carlo binary search); if the match rate does not improve over 'w/o approx. sch.', the concern is not performance-critical; if it improves substantially, the current schedule's approximation is the bottleneck.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central novelty is entropy conditioning for the periodic Bayesian flow, and the load-bearing step is the 'linear entropy sender accuracy schedule' in §4.1 (detailed in Appendix A.5). The stated objective is H(t_i) = E_{pF(c_i|x;α_1..α_i)}[H(vM(·|c_i))] = (1−t_i)H(0)+t_i H(1). The numerical procedure, however, first solves H(c(t_i)) = linear for a scalar c(t_i), and then chooses α_i by matching E_{y∼vM(x,α_i)}[√(α_i^2+c_{i-1}^2+2α_i c_{i-1} cos(y−m_{i-1}))] = c(t_i), for 'arbitrarily selected x'. This does not implement the stated objective: H is nonlinear in c (H(c)≈ln 2π − c²/4 for small c), so E[H(c_i)] ≠ H(E[c_i]); matching expected c does not linearize expected entropy. Moreover, the recursion treats c_{i-1} and m_{i-1} as deterministic scalars, whereas in the actual flow they are random (m_{i-1} depends on all previous y's). The x-dependence partially cancels by rotational symmetry when m_{i-1}=x, but the randomness of m_{i-1} and c_{i-1} is ignored. Hence the schedule is a mean-field approximation, not the linear-entropy schedule claimed. Table 3 shows the searched schedule matters empirically (w/o approx. sch. drops to 49.76), so if this gap is large, the theoretical rationale for entropy conditioning is weakened, though the empirical gains could persist from a different mechanism. This concern is about the argument, not the experimental results.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CrysBFN, a Bayesian Flow Network for crystal generation, by constructing a periodic Bayesian flow on the hypertorus using von Mises distributions. Because the usual additive-accuracy property of Gaussian BFNs fails for von Mises distributions, the authors introduce a non-additive accuracy formulation, a fast simulation-free equivalent Bayesian flow distribution, and an entropy-conditioning mechanism in which the receiver's concentration parameter c, rather than time alone, is fed to the network. The method is evaluated on ab initio crystal generation (Perov-5, Carbon-24, MP-20) and crystal structure prediction (Perov-5, MP-20, MPTS-52), reporting strong property statistics, competitive coverage, and a large sampling-efficiency gain over diffusion baselines. Appendices contain derivations, invariance proofs, implementation details, and additional ablations.","tokens_in":30846,"tokens_out":8409,"duration_ms":80893,"significance":"If the theoretical claims are fully supported after revision, this is a substantial contribution: it extends Bayesian Flow Networks to non-Euclidean circular manifolds, identifies and addresses the non-additive accuracy problem, and demonstrates that entropy conditioning improves over time conditioning. The paper also provides useful engineering contributions: a code repository, a fast non-autoregressive sampling formulation, equivariance propositions with proofs, and ablations showing the importance of the searched sender schedule. The main limitation is that the central 'linear entropy' schedule is currently a mean-field approximation rather than the exact objective stated, and there is an apparent atan2 inconsistency in the equivalent-flow formulas that affects training-time sampling. These issues are correctable but must be addressed before the method's theoretical rationale and empirical claims can be fully accepted.","major_comments":[{"comment":"The stated objective is to make the receiver-belief entropy H(t) = E_{pF(c_i|x;α_1..α_i)}[H(vM(·|c_i))] linear in t, but the numerical procedure does not implement this objective. It first solves H(c(t_i)) = (1-t_i)H(0)+t_iH(1) for a scalar c(t_i), then chooses α_i by matching the expected concentration E_{y∼vM(x,α_i)}[sqrt(α_i^2+c_{i-1}^2+2α_i c_{i-1} cos(y−m_{i-1}))] to c(t_i), using an arbitrarily selected x. Since H(c) = −c I1(c)/I0(c) + ln(2π I0(c)) is nonlinear in c, matching the expected c does not linearize the expected entropy; moreover, the recursion treats c_{i-1} and m_{i-1} as deterministic scalars, whereas in the true flow they are random variables. The resulting schedule is therefore a mean-field approximation, not the exact 'linear entropy' schedule claimed. This matters because Table 3 shows that replacing the searched schedule with the hand-designed roughly linear schedule c(t) = t c(1) drops the MP-20 match rate from 64.35% to 49.76%. The authors should either provide a Monte Carlo check that E[H(c_i)] is close to linear under the proposed schedule, or explicitly reframe the schedule as an approximate/heuristic design and support its choice empirically.","section":"§4.1 / Appendix A.5 / Eq. (56)"},{"comment":"The equivalent Bayesian flow distribution and the fast training sampler use atan2(Σ α_j cos y_j, Σ α_j sin y_j). With the standard atan2(y,x) convention used in Eq. (8), these arguments are reversed; the correct mean direction of the vector [Σ α_j cos y_j, Σ α_j sin y_j] is atan2(Σ α_j sin y_j, Σ α_j cos y_j). As written, pF(m_i) is centered at π/2 − x rather than at x, so Proposition 4.1 and Algorithm 1 are inconsistent with the recursive Bayesian update in Eq. (8) and Algorithm 2. If a nonstandard atan2 convention is intended, it must be stated explicitly and reconciled with the other equations. This is load-bearing because the simulation-free training procedure relies on sampling m_i from the correct Bayesian flow distribution.","section":"Eq. (15), Eq. (52), Algorithm 1 line 13"},{"comment":"The abstract and introduction claim that CrysBFN 'consistently achieves new state-of-the-art on all benchmarks.' This is contradicted by Table 1: on Perov-5 the COV-R is 99.52 vs 99.74 for DiffCSP, and on MP-20 the COV-R is 99.09 vs 99.71 for DiffCSP and 99.49 for FlowMM. CrysBFN does improve several property statistics and COV-P, but it does not dominate all coverage metrics. The claims should be revised to 'competitive or state-of-the-art on most metrics' or should specify the exact metric set for which state-of-the-art is claimed.","section":"Abstract / §5.1 / Table 1"},{"comment":"The ab initio generation results in Table 1 are reported without error bars, number of seeds, or statistical significance tests. Appendix D provides error bars only for the crystal structure prediction task (Table 5). Given that several Table 1 comparisons are close (e.g., COV-P on Carbon-24 is 99.12 vs 97.27, and COV-R on MP-20 is 99.09 vs 99.49), the claim of consistent superiority needs at least a statement of whether the differences are stable across seeds or should be treated as single-run comparisons.","section":"§5.1 / Table 1 / Appendix D"}],"minor_comments":[{"comment":"The assignment lists ΨF twice; the third term should presumably be ΨA (the atom-type prediction), not ΨF again.","section":"Algorithm 1, line 18"},{"comment":"The text states '+4.34% compared to DiffCSP' for compositional validity on MP-20, but Table 1 gives 87.51 vs 83.25, a difference of 4.26 percentage points. Please verify the number.","section":"§5.1, text after Table 1"},{"comment":"The FlowMM entry '0.239 - 0.083' is visually ambiguous; it should indicate which cells are dρ and delem and that dE is not reported.","section":"Table 1, FlowMM row"},{"comment":"The abbreviation 'w/o approx. sch.' is not defined in the table caption; the text in §5.3 explains it, but the caption should be self-contained.","section":"Table 3 / §5.3"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"CrysBFN is a real advance, not a wrapped-up baseline. The authors build the first Bayesian flow on a hypertorus using von Mises distributions, correctly identify that the additive accuracy property of Gaussian BFN fails on the circle, and derive an equivalent non-autoregressive sampling distribution that avoids expensive sequential simulation. The entropy-conditioning mechanism is a genuine design insight: the receiver's concentration c is not a deterministic function of time, so conditioning the network on c rather than t makes sense. The empirical work is substantial: state-of-the-art or competitive results on Perov-5, Carbon-24, MP-20, and MPTS-52, plus a sampling speedup that is convincingly demonstrated in Fig. 4. I would take the \"consistently new SOTA on all benchmarks\" claim with a grain of salt—Table 1 shows CrysBFN is actually lower in COV-R than DiffCSP on Perov-5 and MP-20, and the ab initio metrics have no error bars. The CSP results do include error bars and are clearly strong.\n\nThe main theoretical soft spot is the linear-entropy schedule. The paper states the objective as making E[H(c_i)] linear in t, but the numerical procedure in Appendix A.5 actually solves H(c(t_i)) = linear for a scalar c(t_i) and then matches E[c_i] to that scalar. Since H is nonlinear, matching expected c does not match expected entropy. The recursion also treats c_{i-1} and m_{i-1} as deterministic in the expectation, ignoring their randomness, and uses an arbitrary x. So the schedule is a mean-field approximation, not the exact linear-entropy schedule claimed. The ablation shows the searched schedule matters (49.76 vs 64.35 match rate), so this is not purely cosmetic, but the paper does not flag the approximation. I suspect the empirical gains come partly from entropy conditioning, which is well supported, and partly from a schedule that happens to work well. This weakens the paper's theoretical narrative but not the core method.\n\nThere is also a small inconsistency about FlowMM baseline coverage: the paper says FlowMM only reports MP-20, which is fine, but the related work section calls FlowMM's quality worse without showing numbers on other datasets.\n\nVerdict: worth a careful read and a serious referee. The formulation is novel, the code is released, and the empirical claims are mostly reproducible. A revision should either fix the schedule description or temper the \"linear entropy\" claim. I would cite this if working on crystal generation or Bayesian flows.","headline":"A novel periodic Bayesian flow that mostly delivers on its empirical promises, but the headlining linear-entropy schedule is a mean-field approximation the paper doesn't acknowledge.","tokens_in":31336,"tokens_out":3300,"would_cite":true,"duration_ms":30701,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","62H11","62F15"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims crystal generation can be cast as a periodic Bayesian flow on a hyper-torus, where the non-additive accuracy of von Mises updates makes entropy, not time, the right conditioning signal—yielding state-of-the-art results…","keywords":["crystal generation","Bayesian flow networks","von Mises distribution","periodic equivariance","non-additive accuracy","entropy conditioning","fractional coordinates","sampling efficiency"],"falsifier":"Retrain CrysBFN on MP-20 with the sender-accuracy schedule re-solved for two very different fixed values of $x$ (for example $x=0$ and $x\\approx 3$) and compare match rates: a large gap would show the reported numbers depend on the arbitrary choice. In addition, record the realized receiver entropy $H(c_i|x)$ for each test crystal at each step and compare the average trajectory with the scheduled line $H(t)=(1-t)H(0)+tH(1)$; a systematic deviation that leaves match rate unchanged would show the schedule is not performing the entropy linearization the theory claims.","tokens_in":30204,"feed_emoji":"💎","tokens_out":18095,"duration_ms":149080,"temperature":0.7,"pith_summary":"Generative models for crystals must respect that atomic coordinates are periodic: a coordinate shifted by a lattice vector is the same coordinate. This paper argues that Bayesian Flow Networks, which generate data by accumulating noisy evidence into a confidence parameter, can be built directly on that circular geometry by replacing Gaussian noise with the von Mises distribution, and that this is what makes high-quality crystal generation affordable. The obstacle it identifies and removes is the non-additive accuracy of von Mises updates: unlike the Gaussian case, the receiver's accumulated confidence after several updates is not the sum of the sender accuracies and is not a deterministic function of time. The resolution is to condition the generating network on the actual entropy of its current belief, which the paper shows is worth roughly a twelve-point jump in structure-prediction accuracy, and to rewrite the flow distribution in a fast, single-shot form that avoids autoregressive simulation. The payoff is state-of-the-art results on standard crystal benchmarks with about a hundred-fold reduction in the number of network evaluations needed at sampling.","feed_headline":"100x faster: periodic Bayesian flow tops crystal benchmarks","feed_subtitle":"CrysBFN conditions on entropy instead of time and reaches 64% match rate in just 10 steps.","key_machinery":"The load-bearing mechanism is the von Mises Bayesian update in polar-vector form. The belief state is $\\theta=\\{m,c\\}$, with concentration $c$ playing the role of $1/\\sigma^2$; observing a sender sample $y$ at accuracy $\\alpha$ updates the belief by weighted vector addition: $\\dot m_i c_i = \\alpha\\dot y + c_{i-1}\\dot m_{i-1}$, which unpacks to $m_i = \\operatorname{atan2}(\\alpha\\sin y + c_{i-1}\\sin m_{i-1}, \\alpha\\cos y + c_{i-1}\\cos m_{i-1})$ and $c_i = \\|\\alpha\\dot y + c_{i-1}\\dot m_{i-1}\\|_2$. Two consequences follow. First, after $i$ steps the accumulated accuracy is $c_i = \\|\\sum_{j=1}^i \\alpha_j \\dot y_j\\|_2$, a random variable in time, so entropy dynamics are non-monotonic and the network must be told its own confidence—hence the entropy-conditioning mechanism that distinguishes this work from time-conditioned diffusion models. Second, the same vector-sum identity makes the Bayesian flow distribution simulable in one shot from $i$ independent von Mises draws, bypassing iterative simulation (Proposition 4.1). The sender accuracies $\\alpha_i$ are pre-computed numerically so that receiver entropy decreases linearly in $t$, and a periodic-translation-equivariant network head carries the per-dimension updates while the original Gaussian and categorical BFN machinery handles the lattice and atom-type parts.","core_discovery":"On the paper's own terms, the discovery is that a periodic Bayesian flow on the hyper-torus $T^{3\\times N}$—the space of crystal fractional coordinates—can be defined from scratch, despite the fact that the von Mises distribution lacks the additive-accuracy property that makes the original Gaussian-based BFN tractable. The receiver's belief is a von Mises distribution with mean direction $m$ and concentration $c$; after observing sender samples $y_1,\\dots,y_i$ at accuracies $\\alpha_1,\\dots,\\alpha_i$, the belief update is equivalent to adding weighted unit vectors, giving $c_i = \\|\\sum_{j=1}^i \\alpha_j \\dot y_j\\|_2$ and $m_i = \\operatorname{atan2}(\\sum_j \\alpha_j \\sin y_j, \\sum_j \\alpha_j \\cos y_j)$. Because $c_i$ is a random variable over the step index rather than a bijective function, the network is fed $\\log c_i$ as an entropy condition, Proposition 4.1 proves the single-shot flow distribution of Eqs. (15)–(16) equals the iterated one, and Propositions 4.2–4.3 establish periodic-translation and O(3) invariance of the generated crystal distribution. The paper reports new state-of-the-art on all four benchmarks, including 64.35% match rate on MP-20 structure prediction, 99.12% coverage-precision on Carbon-24, and 60.02% match rate with only 10 network forwards.","pith_inferences":["I expect the entropy-conditioning fix to transfer to other non-Euclidean Bayesian flows—spherical or directional data, molecular torsions, discrete flows—wherever the same additivity failure appears; the paper gestures at this but does not test it.","Because the accuracy schedule is solved for an arbitrarily chosen $x\\in[-\\pi,\\pi)$, I infer the linear-entropy guarantee is approximate per sample; re-solving the schedule on the empirical distribution of coordinates, or making accuracies adaptive, is a direct sensitivity test.","The paper states the continuous-time loss is intractable under non-additive accuracy, so training is discrete-step only; a continuous formulation would need a different theoretical handle, which currently limits export to step-count-free samplers.","I infer that a 100x sampling speedup at matched quality could change the economics of crystal structure prediction, since candidate structures for downstream DFT or machine-learned-potential validation could be screened at far lower sampling cost—a use case the paper does not run."],"forward_implications":["CrysBFN claims new state-of-the-art on Perov-5, Carbon-24, MP-20, and MPTS-52 for both ab initio generation and structure prediction, with 64.35% match rate on MP-20 and 99.12% coverage-precision on Carbon-24.","Entropy conditioning is claimed to be decisive: replacing it with time conditioning drops MP-20 match rate from 64.35% to 52.16%, and swapping the torus flow for plain Gaussian BFN collapses it to 6.17%.","Sampling requires about two orders of magnitude fewer network forwards, with 60.02% match rate at 10 steps surpassing DiffCSP's 51.49% at 2000 steps.","The single-shot flow distribution removes autoregressive simulation during training, yielding roughly a 4x per-batch speed-up in the paper's measurement.","The periodic Bayesian flow transfers to other data on the hyper-torus, which the paper explicitly offers as a route beyond crystal generation."],"supporting_citations":[{"why":"Supplies the BFN framework this paper extends: sender/receiver distributions, the training objective, additive accuracy, and the Gaussian/categorical flows reused for lattices and atom types.","marker":"Graves et al. (2023)"},{"why":"Defines DiffCSP, the principal diffusion baseline, and the crystal representation, periodic E(3) equivariance, dataset splits, and CSPNet backbone that CrysBFN inherits with modifications.","marker":"Jiao et al. (2023)"},{"why":"Establishes the variance-reduced parameter-space advantage of BFN for 3D molecule geometry, the Euclidean template this paper moves to the hyper-torus.","marker":"Song et al. (2023)"},{"why":"Provides the von Mises distribution, its differential entropy formula, and the directional-statistics foundation for the circular Bayesian update.","marker":"Mardia & Jupp (2009)"},{"why":"FlowMM, the flow-matching baseline whose sampling-efficiency improvement over diffusion CrysBFN claims to surpass while keeping quality.","marker":"Miller et al. (2024)"},{"why":"CDVAE; supplies the evaluation metrics, datasets, and an early diffusion-VAE baseline for ab initio crystal generation.","marker":"Xie et al. (2021)"},{"why":"Lemma for proving a Markov chain marginal is G-invariant from an invariant prior and equivariant transitions, used for Propositions 4.2 and 4.3.","marker":"Xu et al. (2021)"},{"why":"Derivation of the von Mises posterior update that yields the closed-form belief-update equations (8)-(9).","marker":"Mardia & El-Atoum (1976)"},{"why":"Gives the KL divergence between von Mises distributions used in the discrete-time training loss for the circular flow.","marker":"Kitagawa & Rowley (2022)"}],"fun_headline_variants":["Periodic Bayesian flow cuts crystal generation to 10 steps","CrysBFN: entropy-conditioned flow tops all crystal benchmarks","100x faster material generation with periodic Bayesian flow","Entropy beats time: new flow for crystals hits SOTA","From 2000 to 10 steps: periodic flow for crystals"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The numerical schedule that sets the noise levels at each step is computed from one arbitrarily chosen coordinate value, so the theory assumes that schedule remains valid for every real crystal; if real samples' confidence dynamics stray far from that curve, the principled rationale for entropy conditioning weakens, even if the model still works in practice.","fun_headline_variants_meta":{"raw":{"variants":["Periodic Bayesian flow cuts crystal generation to 10 steps","CrysBFN: entropy-conditioned flow tops all crystal benchmarks","100x faster material generation with periodic Bayesian flow","Entropy beats time: new flow for crystals hits SOTA","From 2000 to 10 steps: periodic flow for crystals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000839,"raw_usage":{"total_tokens":3744,"prompt_tokens":1118,"completion_tokens":2626,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":734,"completion_tokens_details":{"reasoning_tokens":2542}},"tokens_in":734,"tokens_out":2626,"duration_ms":17858,"temperature":1.0,"reasoning_tokens":2542,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T13:38:52.553445+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain CrysBFN on MP-20 with the sender-accuracy schedule re-solved for two very different fixed values of $x$ (for example $x=0$ and $x\\approx 3$) and compare match rates: a large gap would show the reported numbers depend on the arbitrary choice. In addition, record the realized receiver entropy $H(c_i|x)$ for each test crystal at each step and compare the average trajectory with the scheduled line $H(t)=(1-t)H(0)+tH(1)$; a systematic deviation that leaves match rate unchanged would show the schedule is not performing the entropy linearization the theory claims.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"FlowMM, the flow-matching baseline whose sampling-efficiency improvement over diffusion CrysBFN claims to surpass while keeping quality."},{"cited_title":"Bayesian inference for the von mises-fisher distribution","cited_arxiv_id":null,"evidence_quote":"Derivation of the von Mises posterior update that yields the closed-form belief-update equations (8)-(9)."}],"review_version":1}