- Rating
- 3
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This work introduces HIVE, which repeatedly applies a single equivariant model to refine Hamiltonian predictions and rank per-structure prediction errors within the same rollout. It uses a flow-based ODE formulation similar to QHFlow, but constructs a deterministic path starting from a zero source so that every step estimates the same target Hamiltonian. The stepwise predictions are averaged to obtain the final Hamiltonian, while their dispersion provides an error-ranking score without requiring labels at inference or a separate uncertainty model. The authors report lower Hamiltonian MAE than previous results on three of four periodic benchmarks, while matching the prior result at the reported precision on the fourth. They also evaluate accuracy across different step counts using the same checkpoint, assess error ranking under structural shifts using newly computed DFT labels for three PBE systems, and separate the gains from averaging from improvements in the individual step estimates.
Critical strengths
The joint construction of prediction and error ranking within a single rollout is interesting. Since every step estimates the same Hamiltonian target, the roles of the mean and dispersion are clear. The ranking score requires neither an additional ensemble nor extra field evaluations beyond the rollout already used for prediction. I also appreciate that the authors clearly explain the inability to detect errors shared across steps and interpret the score as a ranking signal rather than a calibrated error estimate. In Hamiltonian learning, where additional data can be obtained through DFT calculations, deciding which predictions should be recomputed is an important problem. The paper explains this need intuitively and helps draw attention to it.
The experiments and analyses provide thorough support for the central claims. By varying only the step count at a fixed checkpoint, the authors show that accuracy improves over the first few steps on all four benchmarks and distinguish the benefits of averaging from improvements in the individual step estimates. For structural shifts, they compute new DFT labels and include comparisons against displacement RMSD, ranking within individual displacement levels, and risk–coverage evaluation on the original test sets. These controls help assess whether the dispersion captures information about structure-specific prediction errors.
Critical weaknesses
Evidence for the practical utility and robustness of error ranking remains limited. The high pooled correlations in Figure 3 span a wide error range across different displacement levels. Although within-level and original test-set evaluations are also provided, the in-distribution correlation for bilayer graphene is 0.442, lower than the pooled result, and the main analyses focus on one checkpoint per system. In addition, the active-learning pilot in Appendix C.4 shows no consistent advantage over random selection when acquisition is restricted to new candidates. The demonstrated acquisition benefit is therefore mainly avoiding redundant selection of structures already seen during training. I acknowledge that the authors explicitly state these limitations, but it remains difficult to judge how effectively error ranking translates into more efficient data acquisition.
The practical advantage over existing methods is difficult to assess after accounting for model size and computational cost. Increasing K at a fixed checkpoint provides evidence that additional inference computation improves accuracy, but the baseline results in Table 1 are taken from the literature and do not match model size or training and inference budgets. HIVE uses larger models than some existing methods, so its resource requirements should be considered alongside the accuracy gains. It would be helpful to include parameter counts and inference times measured under consistent conditions in Table 1, and to report training and inference costs separately. This would clarify whether HIVE offers a better combination of accuracy and error ranking than existing predictors or ensembles under the same compute budget.
Questions for authors
Performance relative to computational cost. Could you report parameter counts and inference times measured under consistent conditions for the models in Table 1? With training and inference costs reported separately, and accounting for model size and geometry feature caching, how does HIVE compare with existing predictors or ensembles in Hamiltonian MAE and risk–coverage performance under the same inference budget?
Contributions of the time and Hamiltonian inputs. Could you compare the prediction accuracy of geometry-only, geometry+time, and geometry+time+H models using the same backbone and training conditions?
Reproducibility of error ranking. Does the in-distribution and within-level ranking performance of remain similar across the already trained ensemble members?
Follow-up Response
I appreciate authors' efforts in rebuttal. My concerns have been addressed so I increase my rate. Please include all new results in the revised paper .