- Rating
- 2
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper introduces KRONOS, a framework for autoregressive 3D molecule generation in the continuous latent space of a Unified Autoencoder. A diffusion head models the next-token distribution, while a stop classifier determines generation length. Fragment-local encoding and a FIM-inspired mixed training strategy enable unconditional and fragment-conditioned generation within a single model. Experiments on QM9 and GEOM-Drugs evaluate generation quality, fragment completion, conformational stability, and sampling efficiency.
Critical strengths
1. Unified task support: A single model accommodates both unconditional generation and fragment-conditioned completion, with molecular size determined during generation.
2. Promising unconditional performance: KRONOS improves several generation-quality metrics over the evaluated autoregressive baselines on QM9 and GEOM-Drugs.
3. Multiple evaluation perspectives: The experiments include xTB refinement, sampling-efficiency measurements, and ablations on atom ordering, model size, and the training mixture.
Critical weaknesses
1. Limited methodological novelty: The core generative formulation is a relatively direct combination of existing approaches: UAE-3D already provides a unified latent representation of molecular topology and geometry; MAR established continuous-token autoregression with Diffusion Loss; and models such as QUETZAL and InertialAR already perform autoregressive 3D molecule generation without a separate molecular autoencoder. Variable-length generation and the use of Diffusion Loss therefore provide limited grounds for establishing a distinctive methodological contribution. The main additional designs are fragment-local encoding and mixed fragment-conditioned training. While these are relevant adaptations, the overall methodological advance appears incremental relative to the established representation and generative framework.
2. Insufficient evidence for the added value of latent autoregression: Although Tables 1–2 and Appendix A.8 demonstrate improvements over the evaluated baselines, the experiments do not sufficiently isolate the benefit of the latent AR formulation. In particular, there is no controlled comparison between autoregressive and diffusion priors operating on the same UAE representation. Comparisons with direct atom-level AR also leave the effects of the representation confounded with model capacity, training budget, and sampling implementation, making it difficult to assess whether the gains justify the additional autoencoder and training stage. Moreover, the ablation reports only unconditional QM9 performance, without measuring the corresponding conditional-generation gains or establishing the claimed trade-off on GEOM.
3. Incomplete and insufficiently documented autoregressive comparisons: The claim of leading unconditional performance among autoregressive methods is difficult to assess without comparisons to relevant approaches such as Geo2Seq, InertialAR and Uni-3DAR. InertialAR and Uni-3DAR provide released checkpoints, while Geo2Seq provides an official reproduction pipeline. These comparisons should use a common evaluation implementation, with differences in data preprocessing, conformer selection, dataset splits, and sampling protocols stated explicitly. Tables 2 and 3 also use a retrained QUETZAL despite the availability of an official GEOM checkpoint. Retraining may be appropriate to match the data setup, but the paper does not explain this choice or document the retraining budget and checkpoint-selection criterion. Evaluating the released checkpoint as an additional reference, with its training-data differences clearly identified, would help assess the robustness of the comparison.
4. Insufficient validation of fragment-conditioned generation: Table 3 reports validity, stability, and uniqueness, but does not directly evaluate whether completions preserve the supplied fragment or successfully attach new structure to it. Algorithm 1 jointly decodes the seed and generated latents, while Appendix A.3 describes a decoder that integrates information across fragments. Fixing seed latents therefore does not by itself guarantee preservation of the seed’s atom identities, internal bonds, or coordinates. Appendix A.13 further acknowledges sensitivity to the fragment’s placement within a reference conformer, with out-of-distribution placements producing shorter and less reliable completions. The evaluation consequently leaves unclear both whether the conditioning constraints are satisfied and how well the method works with independently specified fragments. Seed retention, successful attachment, and sensitivity to centering and rigid transformations should be quantified, including the fraction of all completions that simultaneously preserve the seed, form a connected structure, and satisfy chemical-validity checks.
Questions for authors
Please see the Weaknesses section above.
I thank the authors for the detailed rebuttal and the additional experiments.
The fragment-preservation measurements fully resolve my second concern: the conditioning fragment is recovered exactly, with coordinate deviations at the reconstruction accuracy of the autoencoder.
On my first concern, the UDM-3D comparison is the right experiment, but its conclusion rests on a baseline the authors were unable to reproduce at its published level (76.3% against the reported 93.5%, while KRONOS reaches 92.6%). I appreciate the transparency here, and the independent reproduction attempt the authors cite does support their reading. Still, the attribution of the gain remains open, and the comparison between latent and direct atom-space autoregression is not settled by this experiment.
I therefore maintain my rating of 2 (weak rejection).