{"id":"50f95cbe-5a29-443a-99d4-8444d9679826","arxiv_id":"2506.10186","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A rotationally aligned latent space lets non-equivariant diffusion models match equivariant model quality on 3D molecule generation.","lead":"This paper shows that a non-equivariant diffusion model can generate 3D molecules as well as specialized equivariant models, if molecules are first rotated into learned aligned orientations. The method, RADM, could make molecule generation faster and easier to scale with transformer architectures.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The ablation evidence for the learned alignment is confounded with the new autoencoder; without a same-autoencoder identity/fixed-rotation control, the central mechanism is not isolated.","rationale":"The reader's weakest assumption is that the learned rotation network, trained without a canonicalization constraint, produces alignments that reduce the complexity of the molecular distribution enough for non-equivariant diffusion to match equivariant models. I agree this assumption is load-bearing and empirically underdetermined, but my concern is more specific: the paper's main ablation does not actually isolate the contribution of the learned alignment. Table 2 compares RADM-GNN against GraphLDM/GraphLDM-aug, which differ in their autoencoder, decoder, and latent space, and against a PCA variant that uses no autoencoder. These are not controlled comparisons with respect to the rotation network. A direct control using the same RADM autoencoder but with identity or random fixed rotations would settle whether the alignment module is responsible for the reported improvements. Such a test is feasible and would not require new datasets or theory. Because the central mechanism is the paper's main contribution, and the current evidence leaves this control unrun, the appropriate verdict remains conditional: the empirical results are plausible and clearly reported, but the causal role of the learned alignment should be verified before accepting the paper's framing at face value.","tokens_in":14092,"tokens_out":12482,"duration_ms":173345,"concrete_test":"Run a control experiment on QM9 and GEOM-Drugs: keep the exact RADM autoencoder and diffusion pipeline, but freeze the rotation network to the identity, or to a random orthogonal matrix sampled per molecule (with the same marginal orientation distribution as the learned rotations), and retrain the same GNN and DiT diffusion backbones. Compare atom stability, molecule stability, validity, and uniqueness against Tables 1 and 2. If the control matches RADM within error bars, the learned alignment is not the causal ingredient and the central claim needs to be weakened.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that a learned, unsupervised SO(3) alignment is what lets a non-equivariant diffusion model match equivariant baselines. This rests on the assumption that Rθ, trained only to minimize reconstruction loss, genuinely produces a latent space that is easier for diffusion. The paper's main support, Table 2, does not isolate this mechanism. GraphLDM and GraphLDM-aug use GeoLDM's equivariant autoencoder (or no aligned pre-rotation), while RADM-GNN uses a newly trained autoencoder with a non-equivariant decoder and an additional rotation network; the PCA variant drops the autoencoder entirely. Thus the comparison conflates the learned rotation with changes in the autoencoder, decoder, and latent distribution. There is no theoretical guarantee that minimum-reconstruction alignments are good for diffusion: the objective has no canonicalization constraint and no diffusion-aware term. If a control with identity or fixed random rotations in the same autoencoder matches RADM's performance, the reported gains would come from the new autoencoder or training setup rather than from learned rotational alignment, and the title/central claim would overstate the role of alignment.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RADM, a non-equivariant latent diffusion model for 3D molecule generation. A graph neural network predicts a per-molecule SO(3) rotation, the molecule is rotated accordingly, and an autoencoder with an equivariant encoder and a non-equivariant decoder maps the rotated molecule into a latent space. A non-equivariant GNN or DiT is then trained as the denoising network in this aligned latent space. Experiments on QM9 and GEOM-Drugs report atom stability, molecule stability, validity, and uniqueness numbers that are competitive with the equivariant baselines EDM and GeoLDM, along with improved training and sampling efficiency. Conditional property prediction on QM9 also shows lower MAE than the reported EDM and GeoLDM baselines.","tokens_in":14317,"tokens_out":6012,"duration_ms":75405,"significance":"If the central mechanism is confirmed, the paper would provide a practical alternative to SE(3)-equivariant denoising networks for 3D molecule generation, with the scalability and speed advantages of transformer-based non-equivariant architectures. The paper is clearly written, includes a code repository, and reports a same-backbone ablation that shows large gains over the non-equivariant baselines GraphLDM and GraphLDM-aug. The reported results, if reproducible, would be a meaningful advance in relaxing architectural equivariance constraints without losing sample quality. However, the attribution of these gains specifically to the learned rotational alignment is not yet cleanly established by the presented experiments, and the cross-paper baseline comparisons leave some uncertainty in the headline claims.","major_comments":[{"comment":"Section 4.2 and 4.3, Table 1: Baseline numbers are copied from previous papers with different training setups, and no error bars are given for those baselines. The claim of comparability to state-of-the-art equivariant models would be substantially strengthened by re-evaluating EDM and GeoLDM with the same evaluation code and, if feasible, the same training budget. At minimum, the paper should state the training steps, batch sizes, and sampling protocols used by the baselines so that the comparison is transparent.","section":"Section 4.4, Table 2"}],"minor_comments":[{"comment":"Section 4.2: The sentence 'RADM DiT-S improves the performance drastically compared with previous non-equivariant models' is informal; specify which metrics and which baselines are meant.","section":"Section 4.5, Table 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is well written and the empirical results are promising, but the main causal claim about learned rotational alignment needs a cleaner ablation. The requested identity-rotation and random-rotation controls are straightforward and should be decisive. The cross-paper baseline comparison is a secondary concern. I see no ethical or novelty issues."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Set aside the benchmark table for a second: the new thing here is the learned rotation. The authors train a small GNN to output an SO(3) rotation per molecule, rotate the coordinates, and then run a completely non-equivariant autoencoder and diffusion model in that aligned latent space. The SVD-based rotation parameterization is sensible, and the visualizations suggest the alignment actually arranges rings in similar orientations. On QM9 and GEOM-Drugs the method lands close to GeoLDM, and the DiT backbone gives a real efficiency win. Code is out. That is a clean, useful contribution.\n\nThe soft spot is the ablation. Table 2 is supposed to validate the alignment itself, but RADM-GNN uses the new non-equivariant autoencoder plus learned rotations, while GraphLDM and GraphLDM-aug use GeoLDM's equivariant autoencoder, and the PCA baseline drops the autoencoder entirely. So the comparison confounds the learned alignment with the change of autoencoder and latent representation. The missing control is the same autoencoder with identity rotation or random fixed rotations. Without that, you cannot attribute the gains to the rotation network. The paper's footnote explains why an equivariant decoder would kill the gradient, but that doesn't address the control. This is fixable, but it is load-bearing for the title claim.\n\nOther notes: baseline numbers are copied from other papers with different training setups; GEOM-Drugs numbers lack error bars (the text says std is negligible after rounding, but the QM9 table shows stds, so why not report them); and the assertion that equivariance is not necessary is stronger than the evidence—what's shown is that one non-equivariant model with a learned alignment can match one equivariant baseline on two datasets. That's a useful empirical finding, not a general theorem.\n\nI'd send this to review. The idea is original, the experiments are mostly reproducible, and the weakness is specific and addressable. If the authors add the identity-rotation control and re-run the baselines in a unified setup, the paper would be much stronger. Worth a serious referee's time; I'd assign it to someone who cares about symmetry in generative models.","headline":"Genuinely new idea and strong results, but the ablation doesn't isolate the learned alignment; the missing identity-rotation control is the main weakness.","tokens_in":14805,"tokens_out":3635,"would_cite":true,"duration_ms":42180,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims equivariant denoising is not required for high-quality 3D molecule generation: a learned per-molecule rotation in latent space lets a plain diffusion model match equivariant baselines while training and sampling faster.","keywords":["3D molecule generation","diffusion models","rotational alignment","equivariance relaxation","latent diffusion","non-equivariant networks","diffusion transformer","molecule generation benchmarks"],"falsifier":"Train the identical pipeline but freeze the rotation network's output to a per-molecule random $SO(3)$ rotation sampled once at encoding time, leaving every other component unchanged, and compare QM9 molecule stability: if the random-rotation variant stays near RADM-GNN's 85.8% instead of sliding toward GraphLDM's 70.5%, the learned alignment is not what drives the gain. A complementary check is to apply the trained rotation network to held-out molecules and measure whether the orientation of shared substructures, such as ring-plane normals, is significantly more concentrated than under random rotation — a uniform orientation distribution would show the alignment itself is vacuous.","tokens_in":13902,"feed_emoji":"🧪","tokens_out":19854,"duration_ms":183267,"temperature":0.7,"pith_summary":"The paper challenges the dominant assumption of 3D molecule generation: that the denoising network must be $SE(3)$-equivariant, meaning its predictions rotate and translate in lockstep with the input molecule. It argues this constraint is a design choice, not a necessity, and proposes RADM, which first learns a per-molecule $SO(3)$ rotation with a small graph network and then trains a completely non-equivariant latent diffusion model on the aligned molecules. The rotation is learned without supervision, purely from reconstruction loss through an autoencoder whose decoder is deliberately non-equivariant, so that orientation determines the loss and the network is pushed to find simplifying alignments. On QM9 and GEOM-Drugs, RADM matches the best equivariant latent diffusion baseline on validity and atom stability and cuts sampling time by up to a factor of seven, with scaling from small to base transformer giving consistent gains. If the claim holds, molecule generation can adopt the scalable, well-optimized architectures of vision and language instead of waiting on specialized equivariant layers.","feed_headline":"Equivariance is optional once molecules are rotationally aligned","feed_subtitle":"A learned per-molecule rotation lets a plain transformer rival equivariant specialists and sample up to 7x faster.","key_machinery":"The load-bearing component is the learned rotational alignment. A vanilla GNN maps a molecule to an unconstrained matrix $M \\in \\mathbb{R}^{3\\times 3}$, which is projected onto the rotation group by the SVD-based map $R = \\mathrm{SVD}_+(M) = U\\,\\mathrm{diag}(1,1,\\det(UV^\\top))V^\\top$; this sample-dependent rotation is applied to the atomic coordinates before encoding. The rotation network is trained jointly with the autoencoder using only the reconstruction loss, and the non-equivariance of the decoder is what makes the gradient informative — if both encoder and decoder were equivariant, the $L^2$ reconstruction error would be constant under rotation and the rotation network would have nothing to learn. The same fixed autoencoder feeds a latent diffusion model whose denoiser is either a vanilla GNN or a DiT-style transformer with attention masking for variable atom counts, leaving permutation equivariance as the only inductive bias.","core_discovery":"On the paper's own terms, the discovery is that the equivariance constraints built into state-of-the-art 3D molecule diffusion models can be relaxed without losing sample quality, provided the data is pre-aligned. A rotation network maps each molecule to an arbitrary $3\\times 3$ matrix that is projected onto $SO(3)$ — the rotation group — by a singular value decomposition, and the resulting rotation is applied to the atomic coordinates before they enter the autoencoder. Because the decoder is deliberately non-equivariant, the reconstruction loss is sensitive to orientation, so the rotation network receives a gradient that pushes molecules toward orientations the encoder can represent compactly. The aligned latent space is then simple enough for a plain permutation-equivariant denoiser: on GEOM-Drugs, RADM-DiT-B reaches 85.0% atom stability against GeoLDM's 84.4% with equal 99.3% validity, and on QM9 it approaches GeoLDM's molecule stability (87.3% vs 89.4%) while exceeding its validity (94.1% vs 93.8%). The paper frames these results as evidence that a well-aligned representation, not equivariant layers, is what carries the performance.","pith_inferences":["The same recipe — a learned transform that removes nuisance symmetries, followed by a symmetry-free generative model — should transfer to other structured generation problems such as point clouds, protein backbones, or crystal structures, though the paper does not test those settings.","Because the alignment objective is purely reconstructive, nothing prevents task-aware alignment: coupling the rotation network to downstream objectives such as property prediction could produce frames even more favorable than the reconstruction-only ones reported here.","A sharper ablation than the ones in the paper would train the autoencoder with fixed random rotations per molecule, isolating whether the gain comes from the learned consistency of the frames or merely from the autoencoder's denoising.","The approach never requires the aligned frame to be canonical or consistent across re-samplings, treating alignment only as nuisance reduction; this weaker goal is what lets it sidestep known impossibility results for continuous canonicalization."],"forward_implications":["RADM-DiT-B matches the state-of-the-art equivariant GeoLDM on GEOM-Drugs atom stability (85.0% vs 84.4%) and validity (99.3% each), and exceeds it on QM9 validity (94.1% vs 93.8%), so non-equivariant generators can be practically competitive.","Sampling 100 molecules takes 7 seconds with RADM-DiT-S and 19 seconds with RADM-DiT-B, versus 49 seconds for GeoLDM and 55 seconds for EDM, transferring the speed of parallel transformer architectures to molecule diffusion.","Scaling the denoiser from DiT-S to DiT-B improves every quality metric on both datasets, indicating that non-equivariant molecule diffusion inherits the scaling behavior of standard deep learning models.","In conditional generation, RADM-DiT-B achieves lower mean absolute error than both EDM and GeoLDM on all six QM9 properties tested.","The learned rotations are visually meaningful: after alignment, shared substructures such as rings appear in similar orientations across molecules, evidence that the autoencoder discovers chemically sensible frames without supervision."],"supporting_citations":[{"why":"EDM is the equivariant diffusion baseline whose dataset splits, noise schedule, and evaluation protocol RADM inherits, and the main quality target it must match.","marker":"Hoogeboom et al. (2022)"},{"why":"GeoLDM is the state-of-the-art equivariant latent diffusion model RADM compares against; its encoder architecture and early-stopping rule are reused so the only changes are alignment and the non-equivariant denoiser.","marker":"Xu et al. (2023)"},{"why":"EGNN supplies the equivariant layer used in the autoencoder encoder and the template whose coordinate update is dropped to obtain the vanilla GNN denoiser.","marker":"Satorras et al. (2021)"},{"why":"DiT is the transformer backbone whose parallelizability gives RADM its training and sampling efficiency gains.","marker":"Peebles & Xie (2023)"},{"why":"Supplies the efficient SVD-based SO(3) projection used to turn the rotation network's matrix output into a valid rotation within automatic differentiation.","marker":"Brégier (2021)"},{"why":"Provides the analysis of rotation parameterizations on which the choice of the SVD+ projection for gradient-based learning is based.","marker":"Geist et al. (2024)"},{"why":"Latent diffusion motivates shaping a latent space with an autoencoder, the mechanism the alignment builds on.","marker":"Rombach et al. (2022)"},{"why":"QM9 is the primary benchmark dataset whose split and property set define the reported quality and conditional-generation numbers.","marker":"Ramakrishnan et al. (2014)"},{"why":"GEOM-Drugs is the larger benchmark (up to 181 atoms) where RADM matches or exceeds GeoLDM's atom stability and validity.","marker":"Axelrod & Gomez-Bombarelli (2022)"}],"fun_headline_variants":["Rotational alignment removes the need for equivariant 3D diffusion","Per-molecule SO(3) alignment lets non-equivariant models compete","Align molecules first, then plain diffusion matches equivariant models","Skip equivariant layers: rotational alignment does the heavy lifting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise, stated by the paper itself as an intuition (Section 3.1), is that reconstruction-driven alignment genuinely helps: rotating each molecule into a learned orientation is what lets a plain diffusion network match equivariant models, and the paper offers no guarantee that this alignment transfers beyond the datasets and molecule sizes tested.","fun_headline_variants_meta":{"raw":{"variants":["Rotational alignment removes the need for equivariant 3D diffusion","Per-molecule SO(3) alignment lets non-equivariant models compete","Align molecules first, then plain diffusion matches equivariant models","Skip equivariant layers: rotational alignment does the heavy lifting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000189,"raw_usage":{"total_tokens":1331,"prompt_tokens":935,"completion_tokens":396,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":324}},"tokens_in":551,"tokens_out":396,"duration_ms":5320,"temperature":1.0,"reasoning_tokens":324,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:32:16.057850+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the identical pipeline but freeze the rotation network's output to a per-molecule random $SO(3)$ rotation sampled once at encoding time, leaving every other component unchanged, and compare QM9 molecule stability: if the random-rotation variant stays near RADM-GNN's 85.8% instead of sliding toward GraphLDM's 70.5%, the learned alignment is not what drives the gain. A complementary check is to apply the trained rotation network to held-out molecules and measure whether the orientation of shared substructures, such as ring-plane normals, is significantly more concentrated than under random rotation — a uniform orientation distribution would show the alignment itself is vacuous.","supporting_citations":[{"cited_title":"G., Vignac, C., and Welling, M","cited_arxiv_id":null,"evidence_quote":"EDM is the equivariant diffusion baseline whose dataset splits, noise schedule, and evaluation protocol RADM inherits, and the main quality target it must match."},{"cited_title":"and Xie, S","cited_arxiv_id":null,"evidence_quote":"DiT is the transformer backbone whose parallelizability gives RADM its training and sampling efficiency gains."},{"cited_title":"R., Frey, J., Zhobro, M., Levina, A., and Martius, G","cited_arxiv_id":null,"evidence_quote":"Provides the analysis of rotation parameterizations on which the choice of the SVD+ projection for gradient-based learning is based."},{"cited_title":"High-resolution image synthesis with latent diffusion models","cited_arxiv_id":null,"evidence_quote":"Latent diffusion motivates shaping a latent space with an autoencoder, the mechanism the alignment builds on."},{"cited_title":"O., Rupp, M., and Von Lilienfeld, O","cited_arxiv_id":null,"evidence_quote":"QM9 is the primary benchmark dataset whose split and property set define the reported quality and conditional-generation numbers."},{"cited_title":"and Gomez-Bombarelli, R","cited_arxiv_id":null,"evidence_quote":"GEOM-Drugs is the larger benchmark (up to 181 atoms) where RADM matches or exceeds GeoLDM's atom stability and validity."}],"review_version":1}