acceptodds
Under review as a conference paper at ICLR 2027

Logical Guidance of Stochastic Interpolants

Abstract

Guidance and composition of generative models have applications in visual design, chemistry, and biology. Logical guidance composes conditional scores exactly for diffusion models. However, FM and SI learn velocity fields. We show that negation and exclusive disjunction admit exact conditional velocity composition on any interpolant, whereas conjunction, and therefore any formula containing one, is exact on Gaussian paths but in general incurs a drift-interaction error, so composed velocities need not sample the logical target. LoGSI solves this by using score-based re-equilibration as our principled mechanism and falling back to the Gaussian-path velocity rule as an approximate transport predictor. In non-Gaussian bridge problems, generation alternates transport steps with score re-equilibration steps. On image transfer (EMNIST to CMNIST, CelebA) and QM9 source-to-target molecular editing with logical attribute constraints, LoGSI improves rule conformity at the same network-evaluation budget over composed-velocity baselines, while also improving fidelity on the image tasks. More broadly, LoGSI applies wherever stochastic interpolants need guidance and logical combinations of constraints are natural.

Primary area:
generative models
Keywords:
Generative Models, Stochastic Interpolant, Guidance, Compositionality, Steering, Boolean Logic, Molecules

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_kycvOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors port logical guidance results that are typically well-known for diffusion to the stochastic interpolants/flow matching case. They show that the usual results tend to not hold in that (velocity field) case, and show how to correct for it instead. Empirical results are given for some simple logical tasks over QM9, EMNIST/CMNIST and CelebA.

Critical strengths
  • The presentation of the paper is rather clear, and rather easy to follow.
  • The experiments seem to demonstrate the advantage of the method for stochastic interpolants over more naive approaches, especially in Table 1.
Critical weaknesses
  • I think my overall qualm is with the initial motivation of this method at all, and perhaps something critical eludes my understanding. In the Gaussian case, diffusion, flow matching, and stochastic interpolants can really be seen as equivalent: flow matching, as the authors point out, can be sampled through a marginally matched SDE, diffusion can instead be sampled deterministically, and stochastic interpolants are meant to encompass both paradigms, and are typically identified with flow matching. So, it is unclear to me why the authors want to port this paradigm (which is interesting) to the stochastic interpolants case. And then, regardless of the equivalence that we can perhaps disagree on, why do the authors want stochastic interpolants instead of using diffusion models? more general paths? Perhaps I have missed out on this motivational bit, but I think this is really crucial to make this paper stronger.
  • Theorem 2’s hypotheses are extremely strong, if I understand correctly, especially the conditional independence of the atoms (in a neighbourhood around the considered point). I know that it is hard to formulate theorems in diffusion and ML in general, but I’m not certain that it is very informative. I understand this as some kind of counter-example to the equality between the fields, and am convinced of their inequality in general, but I am not certain that it is as general as hoped for. So, perhaps the “theorem” label is a bit too strong? Perhaps demoting it to a proposition would be better?
  • The plots for the experimental section are very hard to read. I cannot tell what is going on, really. One of the plots has also the legend overlapping with a quarter of the results. It would be very important to replot these to be able to read those numbers. However, I appreciate the willingness to produce so many baselines.
Questions for authors

I am mostly concerned about the motivation of the paper, as explained in the weaknesses. I think the paper is relatively sound, although largely incremental, even if a justification for SI's over diffusion is provided. I would be happy to revise my score should clarity in the experimental results, and, most importantly, in the motivation were addressed.

3 comments
AuthorsOpenReview ↗

Rebuttal by authors

We thank the reviewer for the clear feedback and for stating what would change their assessment. We address motivation first, since the reviewer identifies it as the main concern.

Motivation: why stochastic interpolants instead of diffusion? We agree with the reviewer that in the Gaussian case, diffusion, flow matching, and stochastic interpolants are essentially equivalent. Our Proposition 1 makes exactly this point: on Gaussian paths, velocity and score are affinely related, and LogDiff carries over to velocities unchanged. If both endpoints of our problem were Gaussian noise and data, there would be no reason to go beyond diffusion, and no need for our method, if not for applying the LoGDiff rules to the velocities and not the scores.

Diffusion models map Gaussian noise to the data distribution. Our setting is different: we transport between two data distributions, ρ0→ρ1\rho_0\to\rho_1, starting from a concrete source sample. This is known as a bridge problem. Examples are editing a given letter into a digit with a prescribed colour (CMNIST experiments), a given face into one with glasses and a hat (CelebA experiments), or a given molecule into one with a ring and an oxygen atom (QM9 experiments). Bridge models, such as stochastic interpolants with data-dependent couplings [1], I2SB [2] or Schrödinger bridges, are the standard tool for this kind of data-to-data transport, and we use stochastic interpolants because they cover this case in a single framework.

Once the source is non-Gaussian, the equivalence the reviewer mentions breaks down. The velocity is no longer an affine function of the score (Eq. 9 has an additional drift term), and this is exactly why LogDiff-style composition fails for velocities (Thm. 2). In other words, the regime in which diffusion and SIs stop being interchangeable is precisely the regime our paper addresses. We made this motivation more explicit in the introduction, with the Gaussian-path equivalence stated as the starting point and the data-to-data setting as the reason to go beyond it.

Strength of the assumptions in Theorem 2. Conditional independence of the atoms is not an additional assumption of Theorem 2. It is the same assumption under which LogDiff's AND rule gives the exact composed score. We impose it deliberately: under conditional independence the scores compose exactly, so Theorem 2 shows that the velocities still fail to compose, and isolates the failure from any violation of conditional independence. Without the assumption, the scores would not compose exactly either, and the comparison would not be meaningful. The neighbourhood is needed only so that the identity can be differentiated in x. Within this setting, the theorem is more than a counterexample: it gives the bias in closed form for every interpolant, coupling and pair of endpoints, and an exact criterion (Δt=0\Delta_t=0) for when composition is exact. The explicit counterexample is in App. A.3. That said, we agree that the label may suggest more generality than intended, and we have renamed it to Proposition 2 in the revision.

Readability of the plots. We agree, and thank the reviewer for pointing out the overlapping legend. We have replotted Figs. 2–4: the legends moved outside the panels. Further, we coloured the curves by method instead of by logical statement, and showed only one logical statement in the FID (CMNIST) and novelty (QM9) plots. These changes shall provide an improved readability of the plots.

Incremental contribution. Please see our answer to Reviewer TebU, where we explain what is new relative to LogDiff and to standard predictor-corrector sampling. In short: (i) we show that LogDiff-style composition does not carry over to velocities on general stochastic interpolants, i.e. when bridging two arbitrary (non-Gaussian) data distributions, and give the resulting bias in closed form (Thms. 1–2); (ii) unlike standard predictor-corrector sampling, our corrector is needed even in continuous time, as it removes a bias of the predictor rather than discretization error; (iii) we show that the Gaussian-path rule is a good transport proxy (Table 1)

We thank the reviewer again for the constructive feedback. We hope that the clarified motivation and the planned revisions address the reviewer's main concerns and, as the reviewer kindly offered, that they would consider revising their score. We are happy to discuss any remaining questions.

[1] Albergo et al. (2024). Stochastic Interpolants with Data-Dependent Couplings. Proceedings of the 41st International Conference on Machine Learning (ICML).

[2] Liu et al. (2023). I²SB: Image-to-Image Schrödinger Bridge. Proceedings of the 40th International Conference on Machine Learning (ICML).

I thank the authors for their clarification on the motivation, which corresponded to my "In the Gaussian case, [...]", and "more general paths?" comments. I am aware of the Schrödinger Bridge literature too. I appreciate their useful edits in the introduction, which highlight the central contribution of the work, and which is now more obvious. Although I am not entirely convinced by the magnitude of the contribution, I raise my score to a 3, reflecting its soundness.

We thank the reviewer for engaging with our response and for raising their score; we are glad the revised introduction makes the central contribution clearer.

Reviewer_bKxSOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper works on the problem of guiding diffusion/flow models under constraints specified as arbitrary Boolean formulas, or to my understanding as someone who has experience with diffusion/flow models but not specifically in this subarea, this studies conditional generation under multiple conditions and Boolean means the generation target is defined by logic statements AND, OR, NOT of those conditions. There is a prior work on logical composition that focuses only on Gaussian priors (diffusion models). This paper identifies a gap between that we cannot recursively compose velocity fields anymore (the asymmetry mentioned in the paper). The authors extend the prior Gaussian-only work to any prior by a predictor-corrector framework. It first performs the same linear combination rule directly as in the prior work LogDiff, then it performs the langevin correction step using the conditional score function. So basically, to my understanding, they use LogDiff style linear combination to get an approximate velocity field to go forward in time; however, since the prior is not Gaussian, such a composed velocity will be drifted (e.g. in Appendix A3, there is an example that a distant cluster v_4 causes error in the composed velocity); to resolve this, the authors claim that the conditional score field at time t is always exact, so after an approximate velocity step forward in time, the authors performs corrections using the conditional score filed at fixed time t to get the sample back to high density regions as the correction step. This approach overall makes sense to me; however, this framework requires training a separate score network in addition to the velocity network, and generation can be computationally expensive as well since every discrete step forward in time requires additional correction steps.

Critical strengths
  1. This work theoretically identifies the gap in applying LogDiff style composition in logical guidance of flow models under non-gaussian prior.
  2. I greatly appreciate the example provided in Appendix A3. Frankly speaking, this paper is a bit difficult to understand, unrelated to the author's presentation. The example in Appendix A3 makes it clear why there is an error with LogDiff style composition.
Critical weaknesses
  1. This framework requires training a separate score network in addition to the velocity network, and generation can be computationally expensive as well since every discrete step forward in time requires additional correction steps.

  2. See question 3

Questions for authors
  1. Are the score and velocity networks identical in architecture and size? Have you considered using a smaller score network for reduce computational cost? This comes with the cost of accuracy, it would be interesting to see how good the final generation is versus different correction fidelity.

  2. If i am reading the figures correctly, in figure 9, validity of molecules seems to drop as K increases. Do you think this is due to conditional dependence? This framework assumes conditional independence, however, unlike color/digit, chemical properties are often correlation/dependent.

  3. To my understanding, OR-ME has exact velocity decomposition as in LogDiff, why in figure 2, increasing corrector steps still shows a noticeable increase in conformity? The authors attribute this to neural network approximation and discretization error. However, if this is the case, is it possible that the gain for other tasks are also from this? Some clarification is needed to understand the empirically contribution of this work.

3 comments
AuthorsOpenReview ↗

Rebuttal by authors

We thank the reviewer for the careful reading and for the very accurate summary of our method.

Computational cost and the separate score network (W1, Q1). The score network is not specific to LoGSI. Any SDE sampler of a general stochastic interpolant needs it, including the CompSI baseline. The cost of the correction steps is already accounted for in our experiments: all comparisons are at matched NFE, and the score evaluations in the corrector count toward the budget. The K=0 baselines spend the same budget on additional transport steps instead.

To answer Q1 directly: yes, in the paper the velocity and score networks are separate and identical in architecture and size (App. B.3). This is not necessary, however. Velocity, score and classifier can share one backbone with separate heads. We tested this on CMNIST:

  • Size: the shared model has about the same parameter count as a single network, which halves the total (16.3M → 8.2M parameters).
  • Training: it only adds a loss term, so training cost does not increase.
  • Sampling speed (120 NFE, batch size 128): the shared model is 3.03× faster for AND, 1.97× for AND¬, and 1.66× for OR-ME and OR-CI. The exact numbers depend on the implementation.
  • Quality: conformity changes by less than 8% relative to the separate models. FID decreases for OR-CI (−0.6, −1%) and OR-ME (−1.6, −6%), and increases for AND (+5, +21%).

With a shared backbone, the overhead of LoGSI is therefore small, at the price of a slight drop in conformity and, for AND, a higher FID. We expect tuning the shared architecture reduces this gap. When quality is the priority, separate networks remain preferable. We will add a cost table to the appendix.

Validity drop on QM9 (Q2). We believe the main reason is that, with more corrector steps, the samples lose the validity they inherit from their source molecule. Each sample starts from a real, valid QM9 molecule. At small K, the edit stays close to this source (low Kabsch RMSD, Fig. 4), so the result is usually a slightly modified valid molecule. As K grows, the corrector moves the samples further toward pt(x∣φ)p_t(x\mid\varphi) as learned by the model and away from the specific source, which is the trade-off we discuss in the limitations (samples approach pt(⋅∣φ)p_t(\cdot\mid\varphi) rather than pt(⋅∣φ,x0)p_t(\cdot\mid\varphi,x_0)). The Kabsch RMSD rising with K reflects this. Validity then shifts from "inherited from the source" toward the validity of the generative model itself, which, as for any learned 3D molecular generator, is below 100%. In addition, many samples only satisfy the rule after a structural change, such as adding a ring or an oxygen atom, which is a harder edit than leaving the molecule almost unchanged. This is consistent with conformity among valid molecules rising with K while validity decreases: the corrector enforces the logical constraint as intended, at the price of moving further from the valid source. In contrast, CompSI and CompVel (K=0K=0) retain validity similar to LoGSI at K=1K=1, but achieve substantially lower conformity to the target logical constraints compared to higher KK.

Conditional dependence between has_rings and has_oxygen, as the reviewer suggests, may contribute as well. If conditional independence is violated, the composed score no longer targets pt(x∣φ)p_t(x\mid\varphi) exactly. The discretization error of the Langevin steps with a learned score is another possible secondary factor. We cannot rule these out and mention them now in the limitations. We also made the explanation in Section 5 more precise, by adding the discussion from above.

Why do correctors help on OR-ME (Q3)? The reviewer is right that by Thm. 1 the composed velocity is exact for OR-ME, so on OR-ME there is no drift-interaction error to correct. The gain there comes from the generic benefit of score correction: the corrector reduces errors that the learned transport accumulates along the trajectory. This is not only a discretization effect, because under compute matching the baseline already uses the finer time grid. The same generic effect is present for the other formulas too, so the reviewer's concern is justified. The theory is, however, reflected in the baseline: the composed velocity reaches 0.97 conformity on OR-ME, where it is exact, but only 0.80 on AND, where Thm. 2 predicts a bias. LoGSI brings all formulas to 0.92–0.99. Moreover, the transport ablation (Table 1) shows that the gains are not explained by generic correction alone: at identical corrector budget, only changing the transport field moves CMNIST AND conformity from 0.15 to 0.65 to 0.90. Since AND is also harder than OR, the two effects cannot be fully separated. We have clarified this in Section 5 (see also our answer to Reviewer TebU).

We thank the reviewer again for the questions, which helped us to sharpen the paper. We are happy to discuss any further points.

Thanks for the response, my concerns are addressed. However, I will maintain my score and low confidence as I have limited background in this specific subarea.

We thank the reviewer for their response and are glad that our answers addressed their concerns.

Reviewer_TebUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies logical guidance of generative models, i.e. steering models to satisfy a boolean expression over desired properties. Specifically, the authors extend LoGDiff (logical guidance of diffusion models) to velocity-based generative models. The main finding is that scores and velocities do not obey the same composition rules in general. Under the required assumptions, the authors show that logical score composition induces exact velocity composition on Gaussian paths. For general stochastic interpolants, negation and mutually exclusive disjunction remain exact. Conjunction can introduce a drift-interaction error on general interpolants. The proposed method, LOGSI, alternates between a transport step using an approximate composed velocity and a Langevin correction step using the composed conditional score to move samples toward the desired distribution. Experiments on EMNIST-to-CMNIST transfer, CelebA transfer, and QM9 molecular editing show improved constraint satisfaction over composition-only baselines.

Critical strengths

The paper provides a theoretical characterization of when logical composition extends from scores to velocities.

The proposed method improves constraint satisfaction across both image and molecular transfer tasks compared with composed-velocity ODE and SDE baselines at matched network-evaluation budgets.

Critical weaknesses

Writing. I found the paper challenging to follow.

The novelty seems limited. The methods builds directly on LogDiff's score-composition results and combines them with established transport and Langevin correction steps. While the characterization of velocity composition error is useful, the overall methodological advance appears limited.

Limited empirical evidence. The experiments demonstrate improvements on three benchmarks, but do not clearly establish whether these gains specifically address the identified drift-interaction error or reflect more general benefits of score correction. The practical value for editing also remains unclear because stronger correction improves constraint satisfaction but also reduces similarity to the original input. It would be useful to compare methods at the same level of input preservation.

Questions for authors

No additional questions beyond the concerns raised above.

4 comments
AuthorsOpenReview ↗

Rebuttal by authors

We thank the reviewer for the careful summary and constructive criticism. We address the three weaknesses in turn.

Writing. To make the paper easier to follow, we have (i) moved the algorithm (Alg. 1) from the appendix into the main text, so the sampler is described in one place, and (ii) added a table to Section 4 summarising when composition is exact: for each logical operator (NOT, AND, OR-ME, OR-CI), whether score and velocity composition are exact on Gaussian paths and on general interpolants. If the reviewer has further passages that were hard to follow, we would be grateful to hear about them and will revise those parts accordingly.

Novelty. We agree that the individual building blocks (LogDiff score composition, transport steps, Langevin correction) are known, and we do not claim them. Our contribution is the following:

  1. Velocity composition is not automatic, and we characterise exactly where it fails. The natural way to port LogDiff to flow matching / SIs is to apply its weights to velocities. We show this is exact for NOT and OR-ME on any interpolant (Thm. 1), but every conjunction node adds a closed-form bias. That bias factorises into a probabilistic prefactor and a drift interaction Δt\Delta_t, which gives an if-and-only-if criterion (Thm. 2, now Prop. 2). Without this result, the naive approach silently samples the wrong distribution.
  2. The role of the corrector is different from standard predictor–corrector sampling. In Song et al. (2021), predictor and corrector target the same marginal, and the corrector was designed to only remove discretization (i.e. integration) error. However, in LoGSI the predictor is biased also in the continuous time limit (i.e. without integration error) (Thm. 2, now Prop. 2). The corrector is the mechanism that makes sampling exact. Therefore the corrector step is needed even as Δt→0\Delta t \to 0.
  3. The choice of transport field matters. The Gaussian-path rule is exact at every NOT/OR-ME node, which makes it a good proxy (this is a new result of the LoGSI paper). Table 1 confirms this. (See also the answer of the next point)

Do the gains address the drift-interaction error, or [...] from generic score correction? Reviewer bKxS raised this too.

The theory is clearly reflected in the baseline: the composed velocity reaches 0.97 conformity on OR-ME, where it is exact (Thm. 1), but only 0.80 on AND, where Thm. 2 predicts a bias. LoGSI closes this gap and brings all formulas to 0.92–0.99. The corrector also improves OR-ME, because it corrects any deviation of the transport from the composed-score target, whether caused by the drift-interaction bias or by errors of the learned transport. Since AND is also inherently harder than OR, we cannot fully separate the two contributions from this comparison alone.

The transport-velocity ablation (Table 1) shows, however, that the gains are not explained by generic score correction. There, the corrector budget is identical (K=5) and only the transport field changes. On CMNIST AND, conformity rises from 0.15 with the unconditional velocity, to 0.65 with the averaged velocity, to 0.90 with our proxy velocity, while score-only correction reaches just 0.03. The same ordering holds on CelebA (0.07 / 0.29 / 0.63) and QM9 (0.27 / 0.31 / 0.34). What matters is the combination: a transport field that is exact at every NOT/OR-ME node and biased only at conjunctions, followed by a corrector that removes the remaining bias. We have added this discussion to Section 5.

Trade-off between conformity and input preservation. We agree with the reviewer that this is an important point for editing. There is an inherent trade-off: to satisfy the logical constraint, a sample often has to change, and the more corrector steps we use, the closer the samples get to pt(⋅∣φ)p_t(\cdot\mid\varphi) and the further they move from the specific source sample. We discuss this in the limitations (samples approach p(⋅∣φ)p(\cdot\mid\varphi) rather than p(⋅∣φ,x0)p(\cdot\mid\varphi,x_0)), and it is visible in our results: LPIPS and Kabsch RMSD to the source grow with K (Figs. 4, 7, 8). The number of corrector steps K therefore acts as a control that lets the user choose how much conformity to trade for preservation. Even at the largest K, the edits stay well below the distance between random source–target pairs, so the samples remain related to their source.

We agree that LoGSI does not yet explicitly control how far a sample may move from its source, and we see this as an important direction for future work, for example by adding a term to the sampler that keeps the sample close to the starting point, or by targeting p(⋅∣φ,x0)p(\cdot\mid\varphi,x_0) directly. We have stated this more clearly in the limitations.

We thank the reviewer again for the constructive feedback. We hope that our answers address the reviewer's concerns, and if so, we would appreciate it if the reviewer would consider raising their score. We are happy to discuss any remaining questions.

AuthorsOpenReview ↗

Rebuttal period ending soon

Dear Reviewer TebU,

we have done our best to provide timely answers to the reviewer’s concerns, and we have modified the document accordingly.

However, we have not received any further feedback on the provided material.

Since the rebuttal period is ending, we would greatly appreciate if the reviewer would consider responding to our rebuttal.

With Kind Regards,

The Authors

Thank you for the response. I appreciate the theoretical distinction between score and velocity composition. However, I remain uncertain whether this characterization and the resulting sampling procedure constitute a sufficiently substantial advance over prior work. I therefore maintain my score.

We thank the reviewer for their response. When one wants to edit samples or bridge between two data distributions, applying LogDiff's rules to velocities is the obvious way to obtain logical control, but our results show that this samples the wrong distribution at every conjunction. We identify exactly when and why this happens, and provide a general framework for logical composition with stochastic interpolants, whose corrector targets the exactly composed distribution under the same assumptions as LogDiff. The method is simple and training-free, which we see as a strength rather than a limitation. We therefore respectfully maintain that the contribution is substantial.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of discussion and revisions

We thank all reviewers and the AC for their time and feedback.

Discussion Reviewer bKxS confirmed that their concerns have been addressed. Reviewer kycv found that the revised introduction makes the main contribution clearer and raised their score. Reviewer TebU agreed with the theoretical distinction between score and velocity composition. Their remaining question is about the significance of the contribution, not its correctness. In our view, it matters that the natural way of composing velocities fails at conjunctions as soon as one edits samples or bridges two data distributions. We characterise exactly when this happens and give a simple method that requires no retraining of the underlying stochastic interpolants.

Revisions (marked in blue in the updated PDF):

  • Introduction: clearer motivation. On Gaussian paths, diffusion, flow matching and stochastic interpolants are equivalent; the data-to-data (bridge) setting is where they differ and where our results apply (kycv).
  • Algorithm 1 moved to the main text, and a new table showing when composition is exact (TebU).
  • Section 5: discussion of drift-interaction bias vs. generic score correction, and a revised explanation of the validity drop on QM9 (TebU, bKxS).
  • Limitations: the conditional-independence assumption, and the trade-off between conformity and staying close to the source sample, with explicit distance control as future work (TebU, bKxS).
  • Figures 2–4 replotted for readability; Theorem 2 renamed to Proposition 2 (kycv).

We will also add a cost comparison of separate vs. shared-backbone networks to the appendix (bKxS).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.