{"id":"dd772e9a-a970-4493-b9a3-14618a04ea63","arxiv_id":"2506.10532","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Equivariant Neural Diffusion (END) combines a learnable, equivariant forward process with diffusion for 3D molecule generation, improving conditional controllability on QM9 and GEOM-Drugs.","lead":"This paper introduces a diffusion model for 3D molecules where the noise corruption process is learned rather than fixed, while staying symmetric under rotations, translations, and reflections. It reports competitive or better generation quality on standard benchmarks, with large gains in conditional molecule generation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Invertibility of the learned affine map U_phi(x,t) is assumed but never enforced or verified; a singular block makes the Jacobian and score computations in Eqs. (14)-(18) and Eq. (5) undefined, so the invariance guarantee holds only under an unverified condition.","rationale":"I read the paper as a synthesis of NFDM and EDM with a learnable affine forward process, and the central claim is that this yields an invariant generative distribution and better conditional generation. The invariance argument itself is structurally sound given the stated assumptions: equivariant F_phi, equivariant \\hat x_theta, and an invariant prior imply an equivariant reverse drift and hence an invariant marginal. The empirical study is also internally coherent, with a controlled EDM* ablation and large conditional gains. The most load-bearing weakness is not the missing noise schedule, the absent code URL, or the mixed unconditional metrics, but the unverified invertibility of U_phi. The entire score and Jacobian machinery in Eqs. (5), (14)-(18), and Algorithm 2 depends on inverting U_phi and on the auxiliary matrices \\bar U_m and V. The parameterization in Eq. (9) leaves \\bar U_phi unconstrained in the interior of [0,1], so the matrices can in principle be singular, and the paper reports no check of this during training or sampling. Because the reader already identified exactly this assumption and conditioned the verdict on it, my stress-test pass does not move the verdict; it reinforces the condition and specifies a concrete, cheap verification that would settle the issue.","tokens_in":24888,"tokens_out":24700,"duration_ms":300953,"concrete_test":"Load the released END checkpoint (or, absent code, retrain a small reference implementation) and, for a grid of test molecules x and times t in [0,1], compute the minimum singular value and condition number of U_phi(x,t), of each block \\bar U_phi^{(m)}(x,t), and of V = (1/M) sum_m (\\bar U_phi^{(m)})^{-1}. Also instrument training to log these quantities before every inverse evaluation. If all minimum singular values stay above, say, 1e-6 for every checkpoint and sampled point, the invertibility premise is empirically satisfied and the concern is resolved; if any value crosses zero or the condition number exceeds about 1e6, the training and sampling equations are ill-defined in that region and the central claim needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central construction rests on F_phi(epsilon,t,x) = mu_phi(x,t) + U_phi(x,t) epsilon (Eq. 6) being invertible in epsilon for every x and t. NFDM theory in Section 2.3 requires this for the score and Jacobian computation, and Section A.5.1 explicitly uses \\bar U^{-1} and V^{-1} (Eqs. 14-18), where V = (1/M) sum_m \\bar U_m^{-1}. However, Eq. 9 parameterizes U_phi as a positive scalar multiple of I plus t(1-t) \\bar U_phi(x,t), with \\bar U_phi unconstrained for t in (0,1). Nothing in the architecture (Section A.6.2) restricts \\bar U_phi or V to be invertible, and no spectral check is reported. If any 3x3 block \\bar U_m or V is singular, or if U_phi itself is singular, the Matrix Determinant Lemma and Woodbury formulas used to compute log|J_{F_phi}| and F_phi^{-1} break down; the drift-matching loss (Eq. 5) and the sampling loop in Algorithm 2 then contain undefined terms. The invariance proof in Section 3.1 likewise inherits the invertibility assumption from Eq. (1). This is not a demonstrated failure in the current runs, but it is a load-bearing premise of the method that is neither enforced nor checked.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Equivariant Neural Diffusion (END), a diffusion model for 3D molecule generation whose forward process is a learnable, time- and data-dependent affine transformation F_phi(epsilon,t,x) = mu_phi(x,t) + U_phi(x,t) epsilon, in contrast to the fixed forward process of EDM. The construction builds on Neural Flow Diffusion Models and adds an E(3)-equivariant parameterization of the forward map and the data predictor. The authors prove, under equivariance and invertibility assumptions, that the learned marginal p_{theta,phi}(z0) is O(3)-invariant, and they report unconditional generation experiments on QM9 and GEOM-Drugs, including a controlled ablation against an EDM* baseline with matched architecture and parameter count, plus conditional generation experiments for composition and substructure conditions.","tokens_in":25175,"tokens_out":7868,"duration_ms":99265,"significance":"If the technical assumptions hold, the paper makes a useful and credible contribution: it demonstrates that a learnable equivariant forward process can improve molecule generation, and the conditional-generation gains are substantial (composition matching 91.5% vs. 76.2% for CEDM* at 500 steps; Tanimoto similarity 0.828 vs. 0.673 at 1000 steps). The controlled comparison to EDM* with matched architecture and parameter count is a genuine strength, as is the clear derivation of the invariance argument. The main reservation is that the method's validity depends on invertibility of U_phi, which is neither enforced nor verified; the current paper therefore establishes a conditional guarantee rather than an unconditional one. I did not find a circularity problem: the empirical gains are not an artifact of the equations, and the contribution relative to NFDM lies in the equivariant, data-dependent forward parameterization and the experimental demonstration.","major_comments":[{"comment":"The invertibility of U_phi is a load-bearing premise that is assumed but never enforced or checked. The framework requires F_phi to be invertible with respect to epsilon (Section 2.3, Eq. (1)), and Section A.5.1 computes the Jacobian determinant and the inverse using the Matrix Determinant Lemma and the Woodbury identity (Eqs. (14)-(18)); these formulas require every d x d block \\tilde U_m and the matrix V = (1/M) sum_m \\tilde U_m^{-1} to be invertible. However, in Eq. (9), U_phi is a positive scalar multiple of I plus an unconstrained t(1-t)\\bar U_phi(x,t) for t in (0,1), and the architecture description in Section A.6.2 imposes no constraint on \\bar U_phi or V. If any block or V is singular, the score computation in Eq. (5) and the sampling recursion in Algorithm 2 become undefined. This is not a demonstrated failure in the reported runs, but it is a central assumption of the method; the authors should either parameterize U_phi so that invertibility holds by construction, or provide a spectral verification during training and sampling, and state the resulting guarantee precisely.","section":"Section 3.2, Eq. (9); Section A.5.1, Eqs. (14)-(18)"}],"minor_comments":[{"comment":"The caption states that END 'compares favorably to the baseline across all metrics on both datasets', but on GEOM-Drugs the validity of END is lower than that of EDM* at every step count (e.g., 89.2% vs. 94.4% at 1000 steps). The main text later acknowledges this ('slightly subpar in validity'), so the caption should be rephrased to avoid overstatement.","section":"Table 1 caption"},{"comment":"The sampling loop is written as 'for t = 1, ..., 1/T do', which appears to be a typo; it should presumably be 'for t = 1, ..., T do' or 'for t = T, ..., 1' depending on the intended discretization.","section":"Algorithm 2"},{"comment":"The text says 'we release a public code repository with our implementation of END' but no URL is provided. Please include the repository link.","section":"Section A.6"},{"comment":"The phrase 'ab-initioQuantum' in the introduction is missing a space; it should read 'ab-initio Quantum'.","section":"Section 1"},{"comment":"The ablation discussion could be more precise: on QM9, END(mu only) sometimes matches or slightly exceeds the full END on validity, uniqueness, and total variation (e.g., V x U 93.5% vs. 92.6% at 1000 steps), while the full END is better on strain energy. The claim that the full model is uniformly better is not supported by the table.","section":"Section 4.1 and Table 6"}],"recommendation":"major_revision","confidential_remarks":"The main technical risk is the unverified invertibility of U_phi. If the authors address this by enforcing invertibility or by reporting a spectral check over training and sampling, I would be willing to support acceptance. The paper's contribution is sufficiently distinct from NFDM thanks to the equivariant affine forward parameterization and the strong conditional results; the citation to NFDM is appropriate. I see no novelty or circularity concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read on arXiv:2506.10532. The paper does something genuinely new: it takes NFDM's learnable forward process and makes it E(3)-equivariant for 3D molecule generation, with a specific affine parameterization F_phi(eps,t,x)=mu+U eps and a clean proof that equivariant F and x_theta yield an invariant marginal. The controlled ablation against EDM*—same architecture, same parameter count, continuous-time training—is the right way to isolate the benefit of the learnable forward. And the conditional results are impressive: composition matching jumps from 76.2% to 91.5%, substructure Tanimoto from 0.750 to 0.828. That's not noise.\n\nThe soft spots are real but manageable. The biggest one is the invertibility of U_phi. The whole NFDM machinery—the Jacobian determinant in Eqs. (14)-(18) and the inverse via Woodbury—requires each 3x3 block of \\bar U_phi and the averaged V to be invertible. The paper neither constrains the network to guarantee this nor reports any spectral check. At t=0 and t=1 the parameterization forces U to be delta*I and I, but in between it's an unconstrained perturbation of a positive scalar multiple of the identity. Singularity is unlikely in practice but possible, and the method would silently break. This is a theoretical gap, not a demonstrated failure, but the authors should either enforce invertibility (e.g., via spectral parameterization) or at least verify it on trained models.\n\nTwo other small things: g_phi(t), the diffusion coefficient in the SDE, is never specified in the paper—it appears in Algorithm 2 and the loss but there's no description of how it's parameterized or learned. And the appendix says a code repository will be released, but there's no URL or commit hash. Both are easy fixes.\n\nThe unconditional results are more mixed than the abstract suggests: on GEOM-Drugs, validity is lower than EDM* though connectivity and strain energy are better. The paper is honest about this, and the connectivity metric is arguably the more meaningful one. So I don't see it as a load-bearing flaw.\n\nOverall: this is a competent, honest paper with a real contribution and a fair comparison. It deserves a serious referee. I'd send it to review and ask the authors to address the invertibility question and provide code and the missing g_phi.","headline":"A solid, referee-worthy synthesis of NFDM and EDM; the learnable forward is a real step, but the unverified invertibility of U_phi is a gap the authors should close.","tokens_in":25756,"tokens_out":2531,"would_cite":true,"duration_ms":28771,"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":"By replacing the fixed forward corruption of an equivariant diffusion model with a learnable, time- and data-dependent E(3)-equivariant affine map, END generates molecules competitively on QM9 and GEOM-Drugs and lifts…","keywords":["equivariant diffusion","molecule generation","learnable forward process","E(3) equivariance","conditional generation","geometric graphs","composition conditioning","diffusion models"],"falsifier":"Take a trained END checkpoint and scan $t\\in(0,1)$ and molecules from the training set, computing the smallest singular value of each $3\\times3$ block of $U_\\varphi(x,t)$; a single exactly singular block, or one so small that the determinant and Woodbury-based inverse formulas blow up, would show that the construction fails on the very objects it samples.","tokens_in":24679,"feed_emoji":"🧪","tokens_out":11129,"duration_ms":112775,"temperature":0.7,"pith_summary":"The paper claims that the forward process of a diffusion model for 3D molecules — the fixed schedule that gradually corrupts atom positions and features into noise — is a bottleneck, and that replacing it with a learnable, E(3)-equivariant affine transformation $F_\\varphi(\\varepsilon,t,x)=\\mu_\\varphi(x,t)+U_\\varphi(x,t)\\varepsilon$ improves generation. On the QM9 and GEOM-Drugs benchmarks, the resulting model, Equivariant Neural Diffusion (END), matches or exceeds fixed-forward baselines for unconditional generation while needing far fewer sampling steps. The largest gains appear under conditions: END matches a requested chemical composition in 91.5% of samples versus 76.2% for its fixed-forward counterpart, and its substructure-conditioned samples reach higher Tanimoto similarity than a guided baseline. The paper also proves the sampled distribution stays invariant under rotations, reflections, and translations whenever the learnable map and the data predictor are equivariant and the prior and noise are invariant. If correct, this implies that the noise route itself — not just the denoiser — is worth learning in symmetry-aware generative models.","feed_headline":"Molecule diffusion that learns its noise hits 91% composition control","feed_subtitle":"Replacing the fixed corruption schedule with a learnable, equivariant one improves both sampling and conditioning.","key_machinery":"The carrying object is the learnable, E(3)-equivariant affine forward transformation $F_\\varphi(\\varepsilon,t,x)=\\mu_\\varphi(x,t)+U_\\varphi(x,t)\\varepsilon$, with $U_\\varphi$ a block-diagonal matrix holding one $3\\times3$ block per atom. It turns injected noise $\\varepsilon$ into the latent $z_t$ through a data- and time-dependent mean and covariance, interpolating from a low-variance Gaussian centered on the data at $t=0$ to a unit Gaussian at $t=1$. This one function defines the conditional marginal, provides the reverse-process drift through its time-derivative and Jacobian determinant, and — being equivariant — carries the invariance of the learned distribution; an equivariant data-point predictor and an invariant prior complete the construction.","core_discovery":"The paper's central claim is that the corruption mechanism of a diffusion model can itself be learned rather than pre-specified, without breaking the symmetries of the generated molecules. END defines the latent variable at time $t$ as $z_t=F_\\varphi(\\varepsilon,t,x)=\\mu_\\varphi(x,t)+U_\\varphi(x,t)\\varepsilon$, where the mean and the per-atom block-diagonal matrix $U_\\varphi$ are outputs of an equivariant network, so the conditional marginal $q_\\varphi(z_t|x)$ is a Gaussian whose mean and covariance depend on both the data point and the time. Because $F_\\varphi$, its inverse, and the data-point predictor $\\hat{x}_\\theta$ are all equivariant while the noise and prior distributions are invariant, the reverse-time drift is equivariant and the learned marginal $p_{\\theta,\\varphi}(z_0)$ is invariant under the Euclidean group. Empirically, this learnable forward process yields unconditional generation on par with current fixed-forward models while matching the training distributions better (lower total variation and strain energy), and larger controllability gains once the forward process is also conditioned: 91.5% versus 76.2% composition matching and 0.825 versus 0.669 Tanimoto similarity on substructure-conditioned generation.","pith_inferences":["A testable extension the paper leaves implicit: slicing the conditional gains by feeding the condition to the forward map only, to the predictor only, or to both would isolate whether conditioning the corruption trajectory is what drives the controllability jump.","The invertibility assumption on $U_\\varphi(x,t)$ could be converted into a guarantee by parameterizing each block as a positive-definite factor (for example a Cholesky or spectral-normalized form), which would make the Jacobian determinant and inverse map well-defined by construction.","Because only equivariance of the forward map and the predictor is required, the same construction should transfer to other E(3)-symmetric point-cloud tasks such as protein backbone or crystal structure generation, where the conditioning gains are likely to matter at least as much as on small organic molecules.","The measured ~2.5x training and ~3x per-step sampling overhead are not intrinsic to the idea: a directly learned reverse drift that does not evaluate $F_\\varphi$'s time-derivative would keep the invariance argument while removing most of the cost."],"forward_implications":["END reaches near-peak QM9 validity and stability with as few as 100 integration steps, where the fixed-forward baseline with the same architecture needs 1000 steps to comparable quality, which the paper reports as a 3x cut in sampling time on GEOM-Drugs.","Composition-conditioned generation becomes nearly fully controllable: 91.5% of END samples match the requested formula at 500 steps, and the rate stays at 89.2% when sampling is cut to 50 steps.","Substructure-conditioned generation beats a guided baseline without training any auxiliary property predictor: Tanimoto similarity 0.825 versus 0.669 for the fixed-forward conditional model and 0.750 for the guided one.","Ablations attribute the gains to the learnable forward process: the identical architecture with a fixed schedule lags on every metric, and learning only the mean recovers part but not all of the improvement.","The forward process becomes condition-dependent, so the condition shapes the entire corruption trajectory rather than only the denoiser, a capability that fixed-schedule models cannot express."],"supporting_citations":[{"why":"Supplies the learnable-forward diffusion construction (conditional ODE/SDE, drift-matching objective) that END makes equivariant.","marker":"Bartosh et al. (2024)"},{"why":"Defines the baseline equivariant diffusion model and the bond-inference lookup table; END generalizes its fixed forward process.","marker":"Hoogeboom et al. (2022)"},{"why":"GEOLDM is a main comparison baseline and is also shown to be a particular case of END's parameterization.","marker":"Xu et al. (2023)"},{"why":"GEOBFN is a geometric Bayesian-flow baseline compared on both benchmarks.","marker":"Song et al. (2024)"},{"why":"EDM-BRIDGE provides a physics-informed prior bridge that END generalizes by learning the drift instead.","marker":"Wu et al. (2022)"},{"why":"Sets up the fingerprint-conditioned substructure task and supplies the guided EEGSDE baseline that CEND outperforms.","marker":"Bao et al. (2023)"},{"why":"Provides the zero-center-of-mass subspace and the ambient-space Jacobian computation used to implement the invertible forward map on full coordinates.","marker":"Garcia Satorras et al. (2021)"}],"fun_headline_variants":["Learnable diffusion noise sharpens molecule generation control","Equivariant diffusion learns its forward process for molecules","Molecule diffusion that learns noise boosts conditioning to 91%","Learnable equivariant corruption improves molecule diffusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The construction assumes the learned per-atom matrix $U_\\varphi(x,t)$ is invertible at every data point and time, and the paper neither constrains the network to guarantee this nor checks it after training, so a single singular block would make the score computation and the sampling recursion undefined.","fun_headline_variants_meta":{"raw":{"variants":["Learnable diffusion noise sharpens molecule generation control","Equivariant diffusion learns its forward process for molecules","Molecule diffusion that learns noise boosts conditioning to 91%","Learnable equivariant corruption improves molecule diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000407,"raw_usage":{"total_tokens":2085,"prompt_tokens":889,"completion_tokens":1196,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":1143}},"tokens_in":505,"tokens_out":1196,"duration_ms":11280,"temperature":1.0,"reasoning_tokens":1143,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:24:04.314892+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a trained END checkpoint and scan $t\\in(0,1)$ and molecules from the training set, computing the smallest singular value of each $3\\times3$ block of $U_\\varphi(x,t)$; a single exactly singular block, or one so small that the determinant and Woodbury-based inverse formulas blow up, would show that the construction fails on the very objects it samples.","supporting_citations":[{"cited_title":"Equivariant diffusion for molecule generation in 3d","cited_arxiv_id":null,"evidence_quote":"Defines the baseline equivariant diffusion model and the bond-inference lookup table; END generalizes its fixed forward process."},{"cited_title":"Equivariant energy-guided SDE for inverse molecular design","cited_arxiv_id":null,"evidence_quote":"Sets up the fingerprint-conditioned substructure task and supplies the guided EEGSDE baseline that CEND outperforms."},{"cited_title":"E (n) equivariant normalizing flows","cited_arxiv_id":null,"evidence_quote":"Provides the zero-center-of-mass subspace and the ambient-space Jacobian computation used to implement the invertible forward map on full coordinates."}],"review_version":1}