{"id":"39793f7e-1fa2-4f66-b19a-7d742849823b","arxiv_id":"2505.13791","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Quetzal is an autoregressive 3D molecule generator that matches diffusion-model sample quality on QM9 and GEOM while sampling much faster and enabling exact likelihood computation.","lead":"Quetzal generates 3D molecules one atom at a time, using a transformer to choose each atom type and a small diffusion network to place its position. It produces molecules comparable to diffusion models on standard benchmarks, but with faster sampling and exact per-atom likelihoods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Atom-order dependence undermines the flexible-task claims more than the core generation result; ordering robustness checks are needed.","rationale":"The paper's core result—that Quetzal matches diffusion sample quality at lower sampling cost with a transformer plus per-token DiffMLP—is credible and well supported by standard benchmarks, several ablations, and honest discussion of limitations. The NLL estimator and GEOM split issues raised by the reader are real but secondary to the architecture's main claim. The most load-bearing vulnerability is the ordering premise. It is not an internal inconsistency, because the model is explicitly defined over ordered sequences; but it is a generality limitation that directly contradicts the abstract's 'natively handles variable-size tasks' and 'accepts any 3D structure' framing. Since the authors already show the failure mode in Table 4, the concern is confirmed rather than speculative. The proposed reordering test would quantify how severe the constraint is in a realistic downstream setting and would force either a rewording of the flexible-task claims or a demonstration of robustness. This does not change the reader's conditional verdict; it sharpens the central condition.","tokens_in":16560,"tokens_out":12277,"duration_ms":128105,"concrete_test":"Run Quetzal's hydrogen-decoration checkpoint on 1,000 QM9 test molecules whose heavy atoms are reordered three ways: original .xyz order, RDKit canonical order, and a seeded random order. Strip hydrogens, decorate, and report correct-hydrogen count and median RMSD after Hungarian matching for each order. If canonical or random order drops correct-H rate below 50% (or to OpenBabel's level), the paper should reword the hydrogen-decoration/scaffold-completion claims as in-distribution demonstrations contingent on .xyz order, not as general native capabilities.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The weakest load-bearing point is the assumption that .xyz atom order is a consistent, localized generation order that downstream tasks inherit. The paper is unusually honest about this: Section 4.4 says hydrogen decoration works because in QM9 hydrogens appear last in the .xyz files, Section 5 says the model performs poorly under random atom orderings, and Table 4 (B.2) shows random permutation collapses molecule stability from 87.6% to 25.9% and xyz2mol validity from 99.1% to 80.7%. These are not cosmetic weaknesses: the abstract advertises that without any architectural changes Quetzal natively handles variable-size tasks like hydrogen decoration and scaffold completion, and Section 3 claims it accepts any 3D structure as input. If real inputs have arbitrary atom order (PDB files, crystallography output, or canonicalized SMILES order), the model's validity and the native flexible-task contribution do not transfer. The unconditional-generation quality claim is much less threatened because sampling defines its own learned order; the concern is specifically about the generality and flexibility claims.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Quetzal, an autoregressive generative model for 3D molecules. The model treats a molecule as an ordered sequence of atoms and factorizes p(M) = product over i of ptype(a_{i+1} | prefix) pcoord(x_{i+1} | prefix, a_{i+1}). A causal transformer predicts the next atom type, and a per-token Diffusion MLP trained with the DiffLoss objective of Li et al. (2024a) predicts the continuous next position. The authors report state-of-the-art results among autoregressive baselines, near-diffusion-level quality on QM9 and GEOM, faster sampling, exact per-atom divergence-based likelihoods, and zero-training adaptation to hydrogen decoration and scaffold completion. The paper includes ablation studies and openly released code.","tokens_in":16734,"tokens_out":5497,"duration_ms":49235,"significance":"If the central claims hold, this is a meaningful step for autoregressive 3D molecule generation: it shows that a simple hybrid of causal transformer and per-token diffusion can close much of the quality gap with equivariant diffusion models while offering exact likelihoods and flexible output sizes. Strengths include the clean decomposition, the use of standard benchmarks with several metrics and multiple evaluation runs, the frank disclosure of the model's order sensitivity and of the GEOM split issue, and the public code. The main reservations concern the strength of the generality claims in light of the demonstrated order dependence and the interpretability of the GEOM NLL result under a leaky split.","major_comments":[{"comment":"The paper correctly discloses that Quetzal is not permutation-invariant and that random atom permutations lower xyz2mol validity from 99.1% to 80.7% and molecule stability from 87.6% to 25.9%. However, the abstract and Section 3 claim that Quetzal 'accepts any 3D structure as input' and 'natively handles variable-size tasks.' These claims are stronger than the evidence supports, since hydrogen decoration and scaffold completion are demonstrated only for the specific atom orderings seen in the training data (e.g., hydrogens last in QM9). Please either add an experiment showing robustness under realistic input orderings (e.g., RDKit canonical order, PDB order) or qualify the claims to say the model requires the training distribution's atom order. This issue is load-bearing for the paper's third contribution.","section":"Section 5, Section 4.4, Table 4 (Appendix B.2)"},{"comment":"The GEOM evaluation uses a random split in which up to 30 conformers per molecule are distributed across train and test; the paper itself notes that 'most molecules in the test set have conformers that are seen in the training set.' Consequently the reported GEOM NLL (-313.63) and generation metrics may largely reflect memorization rather than generalization. I ask for an additional evaluation on a molecule-disjoint split or, at minimum, a clear statement that the reported numbers are not a test of generalization to unseen molecules. Without this, the 'state-of-the-art NLL on GEOM' claim is not fully supportable.","section":"Section 4.3 and Appendix B"},{"comment":"The NLL column in Table 1 mixes different estimators across models: Quetzal numbers come from the exact per-atom ODE likelihood computed with Ndiff=60, while baseline numbers are copied from prior papers and may be ELBOs or use different solvers. Please report the estimator and discretization for every entry, or restrict the comparison to models evaluated with the same protocol. Also, the phrase 'exact likelihood' in Section 3.1 should be qualified as exact only in the limit of the ODE integrator, since Ndiff=60 is a finite discretization.","section":"Table 1 and Section 3.1"}],"minor_comments":[{"comment":"There are recurring formatting errors from missing spaces (e.g., 'we presentQuetzal', 'we proposeQuetzal', 'Quetzaluses') and some table headers are cramped; please fix these presentation issues.","section":"Throughout"},{"comment":"The caption states 'We assume uniqueness is 100%.' This assumption should be justified, since it directly affects the reported validity x uniqueness values for the baselines.","section":"Table 2 caption"},{"comment":"The paper explains Quetzal's reduced validity x uniqueness on QM9 as 'signs of overfitting' on the fixed atom orderings. Consider adding an explicit analysis, such as nearest-neighbor distances to the training set, to support this interpretation rather than leaving it as a hypothesis.","section":"Section 4.2"},{"comment":"The scaffold completion results are purely qualitative and the paper defers quantitative evaluation to future work. Please clearly label this section as a demonstration rather than an evaluated capability, or provide a quantitative metric.","section":"Section 4.5"},{"comment":"Given that 'we need to know the best order in which to generate atoms' is described as a central limitation, a short discussion of possible remedies (e.g., order learning, canonicalization, or order-robust training) would strengthen the paper.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's the quick read on Quetzal. The core idea is simple and honest: take MAR's Diffusion Loss, apply it per-atom to continuous 3D coordinates in a causal transformer, and you get an autoregressive molecule generator that samples much faster than all-atom diffusion and gives exact per-atom likelihoods. On QM9 and GEOM it matches or beats prior autoregressive models and is competitive with diffusion baselines on validity and uniqueness. That part I believe. The speed numbers are also credible—the architecture genuinely reduces the number of expensive transformer passes.\n\nWhat's new: the application of DiffLoss to 3D molecules, the demonstration that it scales to GEOM, exact likelihood via a 3x3 Jacobian, and the variable-size tasks like hydrogen decoration. None of these are fundamentally new frameworks, but it's the first time an autoregressive 3D generator reaches diffusion-level quality with these extra capabilities. That's a useful contribution.\n\nThe soft spots are mostly in the evaluation claims. The NLL column in Tables 1 and 2 mixes estimators: Quetzal's exact ODE NLL is not directly comparable to EDM's or SymDiff's ELBO-based numbers. The 'state-of-the-art NLL on GEOM' claim is weakened by this and by the known data leakage in the random conformer split—the authors themselves point this out. Similarly, the scaffold completion section is explicitly qualitative, and hydrogen decoration works only because QM9 lists hydrogens last; the authors are upfront about the ordering sensitivity. That doesn't kill the core generation result, but it means the 'natively handles flexible tasks' claim is oversold. Real-world PDB or crystallography inputs won't have the training ordering.\n\nI agree with the stress-test's framing: the ordering dependence threatens the generality claims more than the unconditional generation quality. But the paper already discloses this in Sections 4.4 and 5, so it's not a hidden flaw. Still, a revision should add order-robustness experiments or, at minimum, reword the claims to say the tasks work under the training distribution's ordering.\n\nCitation pattern is clean, code link is provided, ablations are reasonable. The paper deserves a serious referee; I'd recommend accept after minor-to-moderate revision.","headline":"Solid autoregressive alternative to diffusion for 3D molecules; core generation and speed claims hold, but NLL comparisons and flexible-task generality need sharpening.","tokens_in":17280,"tokens_out":2088,"would_cite":true,"duration_ms":19124,"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":"Quetzal builds 3D molecules atom-by-atom and matches diffusion quality.","keywords":["autoregressive generation","3D molecule generation","diffusion loss","causal transformer","DiffMLP","hydrogen decoration","scaffold completion","exact likelihood"],"falsifier":"Permute the atom order of molecules in the QM9 test set, or reorder the bare heavy-atom scaffolds before hydrogen decoration, and measure generation quality or decoration accuracy. The paper itself reports xyz2mol validity falling from 99.1% to 80.7% under random atomic permutations; if a different ordering scheme restores high validity, then the ordering is a modeling choice rather than a hard limitation.","tokens_in":16335,"feed_emoji":"⚛️","tokens_out":5619,"duration_ms":50425,"temperature":0.7,"pith_summary":"Quetzal treats each 3D molecule as an ordered sequence of atoms and generates it one atom at a time. A causal transformer predicts the next atom's type, while a small per-atom diffusion model, the DiffMLP, predicts the continuous 3D position conditioned on the already-built prefix. The paper claims this simple hybrid closes most of the quality gap between autoregressive and diffusion 3D molecule generation: on QM9 and GEOM it beats all autoregressive baselines and reaches diffusion-level validity, while sampling substantially faster. Because the coordinate diffusion model acts on a single 3D position at a time, exact likelihood scores are tractable, and variable-size tasks like hydrogen decoration and scaffold completion work without any architectural change.","feed_headline":"Quetzal builds 3D molecules atom-by-atom and matches diffusion quality","feed_subtitle":"A causal transformer plus a tiny per-atom diffusion model cuts sampling cost while keeping quality competitive.","key_machinery":"The load-bearing object is the Diffusion Loss, applied as a per-token Diffusion MLP (DiffMLP) that models the distribution of the next atom's 3D position. The DiffMLP is a conditional denoising diffusion model over a single coordinate vector, trained with denoising score matching and sampled with the Heun integrator on geometrically spaced time steps. A causal transformer processes all prefixes in one forward pass to produce prefix embeddings; a second transformer stack incorporates the next atom type and produces a conditioning vector; the DiffMLP then maps noisy coordinates, timestep, and conditioning vector to the clean position. This separation means only O(n) transformer forward passes are needed per molecule, while all diffusion steps go through a cheap MLP, which is what makes fast sampling and exact 3x3 Jacobian likelihood computation tractable.","core_discovery":"The central claim is that a standard causal transformer combined with Diffusion Loss—a per-token diffusion model for continuous coordinates—is enough to make autoregressive 3D molecule generation competitive with state-of-the-art diffusion models. Quetzal factorizes the molecule distribution into a categorical next-atom-type prediction from a GPT-style transformer and a continuous next-position prediction from a lightweight diffusion MLP over 3D coordinates. On QM9, Quetzal reaches xyz2mol validity of 98.6% and validity-by-uniqueness of 94.0%, ahead of previous autoregressive models and comparable to or better than strong diffusion baselines. On GEOM, it is the first autoregressive model demonstrated at that scale, with lookup validity-by-uniqueness of 95.3% and the best reported negative log-likelihood of -313.63. It also samples 22.5x faster than SymDiff on QM9 at 30 diffusion steps, and 128x faster than EDM on GEOM. The same trained model performs hydrogen decoration with 99.8% correct hydrogen counts and low RMSD, and completes molecular scaffolds, because generation is conditioned on an arbitrary prefix.","pith_inferences":["If the ordering assumption transfers to other ordered 3D data, the same architecture could apply to proteins or crystal structures where a canonical atom order exists, but the paper only tests molecules with inherited .xyz ordering.","Exact per-token likelihood opens a concrete path to importance-sampled Boltzmann generators and reward-based finetuning, which the paper names as future work.","A testable extension would be learning or inferring the atom generation order (for example via masked diffusion), removing the reliance on fixed .xyz order; the paper explicitly flags this as the main open problem."],"forward_implications":["Autoregressive models can be made competitive with diffusion models for 3D molecular generation without discretizing coordinates or using equivariant layers.","Generation speed improves by one to two orders of magnitude because the expensive transformer runs once per atom rather than once per diffusion step.","Exact per-atom log-likelihood becomes computable, enabling direct likelihood-based model comparison and potential use in importance sampling.","Because the model accepts arbitrary-size inputs and outputs, tasks like hydrogen decoration and scaffold completion can be solved with the same checkpoint and no retraining.","The approach scales with standard transformer infrastructure, so larger models and datasets can be leveraged directly."],"supporting_citations":[{"why":"Supplies the Diffusion Loss method that lets an autoregressive transformer generate continuous-valued tokens, the core mechanism Quetzal adapts to 3D coordinates.","marker":"Li et al., 2024a"},{"why":"Provides the diffusion training and sampling framework, including preconditioning, the Heun integrator, and geometrically spaced timesteps, used by the DiffMLP.","marker":"Karras et al., 2022"},{"why":"Defines the EDM baseline and the QM9/GEOM train-validation-test splits and evaluation protocol that Quetzal is measured against.","marker":"Hoogeboom et al., 2022"},{"why":"Contributes the Symphony autoregressive baseline and the xyz2mol validity and uniqueness evaluation used throughout the experiments.","marker":"Daigavane et al., 2023"},{"why":"Provides the QM9 dataset of quantum chemistry structures used for unconditional generation and hydrogen decoration experiments.","marker":"Ramakrishnan et al., 2014"},{"why":"Provides the GEOM dataset of diverse drug-like conformers used to demonstrate scaling to a large, diverse benchmark.","marker":"Axelrod and Gomez-Bombarelli, 2022"},{"why":"Contributes the Mol-StrucTok autoregressive baseline and the tabulated xyz2mol results that Quetzal compares against.","marker":"Gao et al., 2024"}],"fun_headline_variants":["Quetzal: autoregressive 3D molecules, diffusion-level quality","Atom-by-atom autoregressive 3D generation now rivals diffusion","Quetzal: transformer + diffusion loss for fast, scalable molecules","Quetzal: 3D molecule builder, diffusion quality without diffusion cost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model depends on a consistent, chemically localized atom order inherited from .xyz files, and if real inputs arrive in a different order, generation quality collapses.","fun_headline_variants_meta":{"raw":{"variants":["Quetzal: autoregressive 3D molecules, diffusion-level quality","Atom-by-atom autoregressive 3D generation now rivals diffusion","Quetzal: transformer + diffusion loss for fast, scalable molecules","Quetzal: 3D molecule builder, diffusion quality without diffusion cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000703,"raw_usage":{"total_tokens":3188,"prompt_tokens":976,"completion_tokens":2212,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":2143}},"tokens_in":592,"tokens_out":2212,"duration_ms":15829,"temperature":1.0,"reasoning_tokens":2143,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:11:11.134812+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Permute the atom order of molecules in the QM9 test set, or reorder the bare heavy-atom scaffolds before hydrogen decoration, and measure generation quality or decoration accuracy. The paper itself reports xyz2mol validity falling from 99.1% to 80.7% under random atomic permutations; if a different ordering scheme restores high validity, then the ordering is a modeling choice rather than a hard limitation.","supporting_citations":[{"cited_title":"GEOM , energy-annotated molecular conformations for property prediction and molecular generation","cited_arxiv_id":null,"evidence_quote":"Provides the GEOM dataset of diverse drug-like conformers used to demonstrate scaling to a large, diverse benchmark."},{"cited_title":"Tokenizing 3D Molecule Structure with Quantized Spherical Coordinates","cited_arxiv_id":"2412.01564","evidence_quote":"Contributes the Mol-StrucTok autoregressive baseline and the tabulated xyz2mol results that Quetzal compares against."}],"review_version":1}