Important question: Have you tried basing your method on a well-established force field instead of assuming harmonic oscillator independent of dihedral angles? I think it is possible your method could be faster or more reliable than conventional optimizations using the same force field, but I am not sure, it would have to be demonstrated.
Here is a long set of comments. These mix comments about clarity and suggestions for improvements that you might want to do in future work, assuming that the authors will want to improve this for publication, possibly in a different venue. I hope these are helpful:
Review: "Prediction Is Not Action: Reliability-Guided Analytic Refinement of Molecular Conformers" (ICLR 2027 submission #38017)
General comment
The paper proposes a lightweight model (0.77M parameters) that repairs residual bond-length and bond-angle defects in generated conformers. A graph-only network predicts ideal local geometry; a learned "Reliability" gate, the analytic bond/angle Jacobian and a learned step size τ turn the residuals into a single small Cartesian update. On ETFlow outputs, Validity3D rises from 38.6% to 59.3% with only 0.007 Å displacement, and the gain transfers to unseen scaffolds and to DiTMC. The problem is relevant and the experimental bookkeeping is careful.
I have three main concerns. First, the novelty is overstated: the method is one steepest-descent step on a harmonic bond/angle potential, the standard Wilson B-matrix machinery of geometry optimizers, with learned parameters. Second, the mechanistic claims are weakly supported: all action-control ablations move the structure less than the full model, and the authors' own routing study (Table 14) shows that conditioning the learned components on the Source, the paper's central idea, has no measurable effect. Third, the baselines are insufficient: no learned refiners, no displacement-matched classical baselines, no non-learned controls, no downstream task, and the V3D ceiling of the GEOM reference data is never reported.
The paper could become a useful contribution if reframed more modestly and supported by these experiments. Since the requested changes go beyond presentation, I recommend major revision.
Detailed questions
- Overstated novelty.
(a) Appendix A.2 shows that the update is one steepest-descent step on a weighted harmonic bond/angle potential, using the transpose of the Wilson B-matrix. Each learned quantity has a classical counterpart: µ ↔ force-field equilibrium values, r/σ² ↔ force constants (e.g., the cited Lindh model), τ ↔ step length/trust radius. This is the machinery of redundant internal-coordinate optimizers (Gaussian's Berny optimizer, geomeTRIC). Please state this in the main text and position the work against these optimizers.
(b) Separating a target from the step taken toward it is standard (line search, trust-region methods, Levenberg–Marquardt, inverse-variance weighting). The novelty is that these quantities are learned; the "first" claim (l. 076) should be softened.
(c) The slogan "Prediction Is Not Action" claims superiority over two alternatives that are never tested: (i) action without prediction, an equivariant network (e.g., EGNN) that directly regresses per-atom displacements, at equal parameters and displacement; (ii) prediction as action, an exact back-transformation that sets bonds and angles to µ (dihedrals fixed). Ablations inside the framework cannot support the slogan.
- Ablations are confounded with displacement. All ablated variants in Table 3 move the Source less than the full model (0.004–0.007 vs. 0.010 Å), which alone can explain their lower V3D. Please compare variants at matched displacement (e.g., V3D vs. RMSD curves over τ), and explain how the fixed τ = 0.00432 Å was chosen.
- Source-conditioning is not supported by Table 14. With no learned component reading the Source (mask 0000), V3D is 56.9%, the same as the final configuration (0101: 56.8%); the authors' own estimated effects (+0.1–0.3 pp) are not significant. This contradicts the paper's central claim. What seems to help is a learned per-primitive weight, not Source conditioning. Please reconcile with Tables 3 and 13. With one seed per mask, Table 14 also cannot justify the final configuration.
- Inference-time ablations overstate component importance. σ is also an input to the Reliability and τ heads, so replacing it after training feeds them out-of-distribution inputs. Indeed, the inference-time swap costs 12.9 pp (Table 3), while retraining without learned σ (J0 vs. J1, Table 13) costs only 5.5 pp (albeit on a different cohort, with a single seed and a fixed step size). Please report retrained ablations (three seeds).
- The gain is essentially bond-length correction. Bond-only reaches 49.2% vs. 50.3% for the full model; Angle-only barely changes the angle error. Please state this and tone down the Bond–Angle allocation contribution.
- Single step. Why only one step with one τ per molecule? Since residuals are computable, iterating (and/or per-atom step sizes) may give much higher V3D at still small displacement. Please test it.
- Targets ignore the Source conformation.
(a) µ is graph-only, so all conformers of a molecule are pulled toward the same values, although equilibrium angles and bonds depend on conformation (gauche vs. anti angles, twisted conjugated bonds, ring puckering).
(b) The training Reference is a random conformer of the molecule, not one matched to the Source, so the model learns ensemble-average geometry.
Please report how much each bond/angle varies across a molecule's GEOM ensemble relative to the V3D tolerance, and compare with training on the closest Reference conformer.
- Unfair classical baselines. MMFF94s and GFN2-xTB appear to have been run without restraints (1.0 and 0.41 Å RMSD), i.e. they also change the conformation. Please compare in the same regime: positional restraints at several strengths (V3D vs. RMSD curve), frozen torsions, and few optimizer steps.
- Missing non-learned controls. Ideal bond lengths are largely tabulated, the typed-context statistics already used for s_stat could also provide reference values, and the gain comes mostly from bonds (point 5). Please report: (i) the same update with µ, σ from the table, r = 1 and constant τ; (ii) exact correction to table values with fixed dihedrals; (iii) a few MMFF/UFF steps with only bond and angle terms. If these approach 59.3%, the network adds little.
- No learned baselines. Flow-Matching Refiner, GeoOpt-Net and GOLF are cited but not compared.
- V3D of the Reference data. Please report the V3D pass rate of the GEOM Reference conformers. Since xTB-optimized structures pass only 59%, the GEOM (xTB) references likely fail often too. Then the model imitates targets that V3D considers invalid, and part of the gain may come from pushing borderline cases just past the threshold.
- Practical relevance. Which application needs V3D-valid structures within 0.01 Å of the input? Please show a downstream benefit.