acceptodds
Under review as a conference paper at ICLR 2027

Arrow Flow Matching for Reaction Mechanisms

13 views

Abstract

Chemists explain a reaction by drawing arrows. Each arrow moves one pair of valence electrons, and an elementary step made of such arrows creates no atom and no electron. We propose a generative model of these arrows. Its state is an integer bond–electron matrix, and its dynamics is a Markov chain trained by discrete flow matching. The central design choice is the elementary event, the change that one update of the chain applies to the state. The common choice is an edit of one entry of the matrix. We prove that a model which draws every entry independently conserves the electron count exactly only if it is deterministic in every entry, whatever its parameters. Arrow Flow Matching instead takes as its elementary event one move from a four-element alphabet, the matrix form of an arrow that chemists draw. Every move leaves the electron count unchanged, and the chain may stop only where a table of valid per-atom states allows it. Conservation therefore reduces to an algebraic identity and validity to a lookup on one softmax class, and both hold with probability one. As every update is an arrow, the trajectory of the model reads as a mechanism that a chemist can inspect. These guarantees cost almost no accuracy. Arrow Flow Matching is competitive with the most accurate model we train, and at its smallest width it already reaches a higher composite score than every other model at every width.

Primary area:
generative models
Keywords:
discrete flow matching, continuous-time Markov chains, reaction mechanism prediction, constrained generative models, conservation laws, arrow pushing, exact validity guarantees

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_w6LqOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper predicts reaction steps through discrete electron moves. Each move preserves the total electron count. A local validity table controls when the model can stop. The method achieves accuracy close to Discrete FlowER and produces interpretable reaction arrows. Experiments cover several model sizes and distribution shifts.

Critical strengths
  1. Each move preserves electrons by construction. The model does not need to learn this constraint from data. This directly addresses an important failure mode in reaction generation.

  2. The method retains competitive accuracy while producing explicit reaction arrows. Comparisons across model sizes and distribution shifts support the usefulness of this design.

  3. Fig. 2 is particularly helpful for understanding the proposed method.

Critical weaknesses
  1. The distinction from constrained autoregressive prediction is unclear. Eq. 13 trains next-move prediction on fixed action sequences. The paper should explain what the flow-matching formulation adds. This affects the assessment of its methodological novelty.

  2. The added value of learning arrow sequences is not fully established. Tab. 3 evaluates agreement with expert arrows. However, it lacks a comparison with arrows reconstructed from the same reactant–product pairs using the paper’s decomposition method. This control is needed to show what the learned trajectories add beyond the reaction endpoints.

Questions for authors
  1. How does Eq.13 differ from constrained autoregressive training on the same action sequences?

  2. What is the validity rate before candidate replacement? Which validity conditions are guaranteed by the local table?

  3. How often does deterministic factorization of the recorded endpoints recover the expert arrows in Tab. 3?

Reviewer_sUEWOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper models elementary reaction steps as sequences of four electron-conserving move types, with termination gated by per-atom validity rules. The resulting model produces interpretable electron-movement trajectories.

Critical strengths
  1. The method is coherent and clearly presented, with explicit distinctions between electron conservation, local chemical validity, and predictive accuracy.

  2. Comparisons across model sizes and shared-backbone variants provide useful evidence that structural guarantees can be imposed with a small accuracy penalty.

Critical weaknesses
  1. The experiments do not establish the downstream benefit of exact conservation. At the smallest width, AFM slightly underperforms Discrete FlowER on both step Top-1 accuracy and recursive pathway accuracy. The higher composite score partly reflects the inclusion of the constraints enforced by construction.

  2. The absence of a filtering-and-resampling baseline leaves the practical necessity of the proposed formulation unresolved. Discrete FlowER already achieves 97.67% validity and 95.53% electron conservation. These rates suggest that rejecting invalid or nonconserving candidates could be a competitive alternative, but its accuracy and computational cost are not evaluated.

Questions for authors
  1. Under equal computational budgets, how does AFM compare with Discrete FlowER augmented with validity/conservation filtering and resampling?

  2. Can the authors demonstrate that exact conservation improves a downstream outcome, such as correct pathways recovered per fixed search budget or fewer unsuccessful pathway expansions, beyond improving the constraint metrics themselves?

Reviewer_s6TdOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper models an elementary reaction step as a Markov chain on integer bond–electron matrices. It argues that such a model should be chemically valid, conserve electrons exactly, and be interpretable as a chemist's arrows. It first applies discrete flow matching directly to the matrices (Discrete FlowER). It then proves that an entrywise head can conserve electrons only if it is deterministic (Theorem 1). Finally, it proposes Arrow Flow Matching (AFM), whose updates are moves from a four-type alphabet that conserve electrons by construction, with a valence table gating when the chain may stop. Against nine reimplemented baselines, AFM is the only model with perfect validity and conservation. Its accuracy is close to Discrete FlowER's, and it has the highest composite score.

Critical strengths

A clean structural principle. Enforcing a hard linear constraint through the elementary event, rather than through penalties or repair, is a clear idea that may extend beyond chemistry. The guarantees cost little accuracy, and the engineering is reported candidly. This includes enumerating the valence table against the evaluator and fixing a reconstruction bug for all models.

A distribution over mechanisms. AFM is the only model that defines a normalised distribution over valid arrow sequences, so mechanisms can be sampled, scored and edited. There is early evidence that this helps. AFM adapts faster from ten examples of an unseen move combination (0.607 vs 0.554 top-1), and its uncertainty concentrates on the first decision at the reaction centre.

Critical weaknesses

The necessity of AFM is not established. All three stated requirements appear achievable without it. Validity and conservation can be obtained by filtering Discrete FlowER's candidates, of which only 2–4% fail; filtering is argued against in Appendix E.3 but never measured. Arrows can be obtained by decomposing Discrete FlowER's predicted change with the paper's own Appendix H procedure.

The interpretability result in Table 3 does not settle this. The paper states that AFM's arrows match the chemist's whenever its product is correct, so the agreement seems to come from the deterministic decomposition rather than from learning. Meanwhile, the FlowER baselines are denied arrows unless every intermediate update conserves electrons.

Theorem 1 does not close the gap either. It concerns spread distributions, while the emitted candidate is an arg-max, and its proof is close to trivial. Accuracy offers no independent reason to switch. Discrete FlowER is ahead on top-1 and path@1, and AFM trails it on unseen move combinations and with 1% of the data. What only AFM provides, a distribution over mechanisms, is the real contribution. Yet it appears only in appendix analyses, none of which compare against filtering plus post-hoc decomposition.

The headline claims rest on evaluation and framing choices. The composite score effectively counts validity twice. For matrix models, atom and proton conservation fall below 1 only when a candidate is invalid. Validity of 1 means agreement with the same RDKit checker used to build the valence table. It is also partly enforced by fallback: failing candidates are replaced by an earlier stoppable state, possibly the reactant itself, and how often this happens is not reported.

The flow-matching framing is loose. Training is next-move cross-entropy on one greedily ordered sequence, which is teacher-forced autoregression with action masking. Some intermediate states are unchemical artefacts of splitting arrows into moves, which weakens the claim that the trajectory reads as a mechanism.

Questions for authors
  1. What are top-k and path@1 for Discrete FlowER and Continuous FlowER after a validity and conservation filter?
  2. How does Table 3 change if the FlowER arrows are obtained by decomposing their final predicted change? Can arrow order be evaluated?
  3. Does AFM beat filtered Discrete FlowER with post-hoc decomposition on editing mechanisms, calibrating alternative mechanisms, and few-shot adaptation?
  4. How often do the fallbacks fire, and how often is the emitted candidate the reactant itself?
1 comment

The new experiments resolve my concerns about fallbacks and the checker, and Table R16 is a useful addition. However, they also confirm that filtered Discrete FlowER matches AFM on accuracy and validity, and that arrow agreement follows from any correct product. The case for AFM now rests on efficiency and mechanism-level capabilities. That is a different contribution from the one the paper makes. The supporting evidence is preliminary: several results are partly guaranteed by construction or lack a chance baseline, and the evaluation-cost comparison depends on how evaluations are counted. I have raised my score slightly to reflect the resolved concerns, but I think the reframed contribution needs to be developed in the paper itself rather than in the rebuttal.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.