acceptodds
Under review as a conference paper at ICLR 2027

Scalable Inference-time Steering in Molecular Design with Multimodal Meta Flow Maps

Abstract

Biological design problems naturally have continuous (e.g. molecular structures), discrete (e.g. DNA residue types) and often both modalities (e.g. protein structure-sequence co-design). Diffusion and CTMC-based methods have been developed for multimodal generation, and, for cross-modal tasks, the two are typically combined. However, it remains challenging to adapt these methods at inference-time to fulfill design constraints. Recent developments of one-shot samplers such as Meta Flow Maps (MFM) have demonstrated high controllability for image generation, showcasing the power of inference-time steering with accurately estimated reward gradients. Here, we adapt Meta Flow Maps to biological domains, and develop MFM framework both in discrete (dMFM) and multimodal settings (multiMFM). We demonstrate dMFM by designing DNA sequences with a target mechanical property. MultiMFM trained on molecular the datasets QM9 and GEOM-Drugs are applied to design molecules of a given physical property. In both cases, our steering algorithms achieve high accuracy with a single trained backbone, outperforming fine-tuning methods adapted for task-specific ends.

Primary area:
generative models
Keywords:
Flow matching, reward guidance, generative models, molecules, DNA

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.80 over 5 reviews, above 85% of the others.
Reviewer_AsioOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors incorporates the meta flow map (MFM) to biological application tasks. MFM is the inference steering method toward the posterior distribution p(x)e^r(x) without training. In this paper, the authors extend MFM to two biological task, discrete DNA design task and joint 3D small molecule design task. They propose discrete MFM (dMFM) and multi-modal MFM (multiMFM), which joints the continuous MFM and discrete MFM. Finally, they validate the proposed method on cyclizability-guided DNA design and conditional small-molecule generation tasks and compare it to various steering methods.

Critical strengths

The authors extend the meta flow map to discrete representation. This paper provides the empirical demonstration of dMFM and multiMFM on DNA and small-molecule design. Not only the biological sequence and 3D small molecule design, this framework can be introduced on multiple biological tasks.

Critical weaknesses
  1. The objective of MFM is shifting the sampling distribution to the posterior distribution p(x)e^r(x) via inference steering. However, the proposed dMFM is not validated that this can yield the same posterior distribution in discrete space. It would be beneficial to show the discrete toy example.
Questions for authors
  1. Recently, Riemannian mean flow (RMF; [1]) is proposed and demonstrated on DNA design task. Please compare against it on the same DNA task under comparable computational budgets, reporting the rewards or target errors. This comparison would help establish whether stochastic posterior look-ahead provides sufficient benefits to justify its additional sampling cost.
  2. Is the reward call count matched to other baselines?
  3. What are the limitations of the proposed method, and for what types of applications is it most suitable?
  4. Please report uncertainty estimates (std) for the proposed methods in Table 2.

Reference

  1. Woo, Dongyeop, et al. "Riemannian MeanFlow." ICML 2026
2 comments
AuthorsOpenReview ↗

benchmark RMF (FMAP), clarifying reward counts and limitations and uncertainties

Thank you for the questions. We added an exact discrete toy test of the tilted distribution, a comparison with RMF-style endpoint guidance, explicit reward-call counts, and standard deviations for Table 2.

W1. Toy Example

Yes. We tested it on all 256 sequences in {A,C,G,T}4. The base law p1 is known there, so the tilted target p(x)p1(x)er(x) can be normalized exactly. Over 500,000 guided samples, the empirical distribution q is within total variation distance 0.0058 of p, compared with 0.394 for the unguided p1. In other words, the guided samples get the probability of any set of sequences right to within 0.6 percentage points. The match improves with finer integration, and the best guidance coefficient in our sweep is the one the theory predicts. We will add the toy example into the revised manuscript.

Q1. Comparison with RMF

RMF guides with the reward gradient of a single deterministic endpoint, xtr(t1(xt)) (RMF, Sec. 3.4). We implemented this rule (FMAP) on our own base generator and reward, so the only difference from dMFM is the guidance signal: one deterministic continuation versus reward-weighted posterior samples. At a slightly lower billed budget, dMFM gives 10% lower error than FMAP, and it also beats DPS, FK, and best-of-N at similar budgets (Table Asio-1). We billed FMAP's entire teacher solve as one call, which favors FMAP.

Table Asio-1. Posterior guidance versus endpoint guidance and search on DNA.

MethodBilled budgetOracle error
dMFM, Eq. (4)8480.1281 ± 0.0015
RMF endpoint guidance (FMAP)8570.1425 ± 0.0034
DPS on the student8570.1407 ± 0.0040
FK, 9 particles8640.1351
Best-of-98640.1825

Cyclizability target y=0.30, held-out oracle; mean ± SEM for the guided methods.

Q2. clarifying whether reward calls are matched

Not exactly. We match the billed budget and report reward calls, backward work, and GPU time separately (Table Asio-2). On DNA, Eq. (4) makes the same forward and reward calls as Eq. (3) but needs no backward passes, which cuts GPU time by 49%; it is also 35% faster than direct GLASS. On QM9, multiMFM makes about as many reward calls as FK, but its backward passes make it slower per output than DPS or FK. To compare accuracy at matched sampling budget and PB-validity floor, Please refer to Table 4Jmz-3 replying to reviewer 4Jmz-3 where multiMFM outperforms baselines with PB-validity floor held at 75%.

Table Asio-2. Generator work, reward work, and measured sampling time.

TaskMethodGenerator forwardsReward forwardsBackward workGPU ms/output
DNAdMFM, Eq. (4)84875305.96
DNADirect GLASS, Eq. (4)84837709.17
DNAdMFM, Eq. (3)84875347 traversals11.74
QM9multiMFM, Eq. (3)358115230 generator + 115 reward39.5
QM9DPS1855757 generator + 57 reward15.0
QM9FK, 8 particles114412006.3
QM9Best-of-16204816071.7

DNA: one H200, length 50. QM9: one A100; the timed settings is the same as the accuracy runs.

Q3. Best use and limitations

Our method is most useful for hard design problems at limited inference budgets: rare target values, or several requirements that must hold at once. On GEOM-Drugs, best-of-4 wins on the two easiest radius-of-gyration targets, but multiMFM wins on the six harder ones, with 19% lower error at the hardest. On DNA, dMFM keeps its advantage up to length 400 with four motifs required at once. Because the sampler does not depend on the reward, it can be reused when objectives change. The main limitations are that guidance strength must be tuned against validity (too much guidance worsens both error and validity), that search does better under very tight limits on sequence-statistic drift, and that results depend on the quality of the distilled sampler and the reward model.

Q4. Standard deviations in Table 2

We now report mean ± SD over three seeds, with 1,000 outputs per property each, for multiMFM and steer-search (SS). These are new evaluation runs and replace the numbers in the submission.

Table Asio-3. Table 2 results with standard deviations.

PropertymultiMFM, all outputsmultiMFM, PB-validmultiMFM-SS
Cv0.444 ± 0.0060.415 ± 0.0120.414 ± 0.015
μ0.818 ± 0.0570.740 ± 0.0530.528 ± 0.027
α1.166 ± 0.0301.113 ± 0.0181.009 ± 0.041
Gap592 ± 12562 ± 6413 ± 6
HOMO270 ± 7259 ± 13243 ± 9
LUMO395 ± 16372 ± 11265 ± 12

Thank you for response and additional experiments and analysis. I'll maintain my score (3).

Reviewer_okH9OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper adapts Meta-Flow-Map to joint discrete-and-continuous modeling and proposes Multimodal-Meta-Flow-Map. It validates the approach on two biological design tasks: cyclizability-guided DNA design and property-guided molecule design. In addition, the paper demonstrates the scalability and robustness of the proposed inference-time scaling method. However, the paper suffers from poor readability, and its main experimental results lack comparison against a key baseline (GLIC, ICML 2026).

Critical strengths

The paper adapt MFM to DNA sequence and property-guided molecule design and show the effectiveness of MFM-based inference-time-scaling.

Critical weaknesses
  1. Missing comparison with a key baseline, GLIC. GLIC (https://arxiv.org/pdf/2606.06303, ICML 2026) is a closely related work on inference-time steering for molecular/biological sequence design. The main experiments, including Table 2, do not compare with GLIC, so it is difficult to judge the actual advantage of the proposed method over the state of the art in this direction.
  2. Comparability of experimental results is questionable. Because Table 2 directly quotes numbers from TFG-Flow without clarifying whether the settings are identical, the comparison may be unfair. If the data splits, reward functions, or evaluation protocols differ, the reported gains lose credibility.
  3. Writing and figure clarity need improvement. Beyond the Figure 2 / Line 185 inconsistency and the grammatical error in the definition of the Discrete Meta Flow Map, the overall readability is poor, making the technical details harder to follow.
Questions for authors
  1. The authors state that the results in Table 2 are taken from TFG-Flow. Can you clarify whether the experimental setting in this paper (data split, reward function, evaluation metric, model size, etc.) exactly matches that of TFG-Flow? If not, please provide a fair comparison under a unified setting.
  2. Why is there no comparison with GLIC (https://arxiv.org/pdf/2606.06303, ICML 2026)? GLIC also addresses inference-time steering for molecular design and is highly relevant. Please explain how GLIC differs from your method in terms of problem formulation, assumptions, and empirical performance.
2 comments
AuthorsOpenReview ↗

compare with TFG and benchmark GILC

Reply to Reviewer okH9

Thank you for the review. Our QM9 setting closely matches TFG-Flow's, multiMFM's molecules are more often PB-valid, and its low errors in Table 2 hold before any validity filtering. We also compare with GILC, both in method and against its published QM9 results.

Q1. Comparability with TFG-Flow

Our settings are mostly identical. multiMFM and TFG-Flow use the same QM9 dataset, and the guide and oracle are trained on the same QM9 splits. We use the guide and evaluation oracle released by TFG-Flow. Both multiMFM and TFG-Flow sample from the same exponential tilt, using the exponential of a squared-error reward. The only difference is that multiMFM is a 3k-parameter transformer, 15× smaller than the 5M-parameter model that TFG-Flow trains. We report PoseBusters (PB) validity, which is stricter than the RDKit validity that TFG-Flow reports. We re-evaluated the PB validity of TFG-Flow's molecules, and multiMFM's molecules have higher validity on every property (Table okH9-1).

Table okH9-1. PB validity (%) on the six QM9 properties.

MethodCvC_vμαGapHOMOLUMO
TFG-Flow83.0 ± 0.383.8 ± 0.580.4 ± 0.680.9 ± 0.784.2 ± 0.987.3 ± 0.4
multiMFM90.390.690.590.891.590.9

We also reran TFG-Flow on its multiobjective benchmark, where multiMFM has lower error on all three tasks (Table ifYC-5 in our reply to Reviewer ifYC).

To check that filtering does not drive the Table 2 numbers, we recomputed multiMFM's errors on all outputs and on the PB-valid subset of the same runs (Table okH9-2). Filtering changes the error by only 4–10%, and even without filtering, multiMFM's error is below the published TFG-Flow value on five of six properties.

Table okH9-2. Filtering audit on the same multiMFM runs.

PropertyTFG-Flow published MAEmultiMFM all-output MAE ± SDmultiMFM PB-valid MAE ± SDmultiMFM PB validity
CvC_v1.750.444 ± 0.0060.415 ± 0.01290.3%
μ0.8170.818 ± 0.0570.740 ± 0.05390.6%
α2.321.166 ± 0.0301.113 ± 0.01890.5%
Gap804592 ± 12562 ± 690.8%
HOMO364270 ± 7259 ± 1391.5%
LUMO941395 ± 16372 ± 1190.9%

Gap/HOMO/LUMO in meV; other units as in manuscript Table 2.

Q2. GILC

Thank you for the pointer; we will cite and discuss GILC (arXiv 2606.06303). GILC is training-free and approximates future outcomes with independent per-token distributions from the pretrained denoiser. We instead distill a sampler that draws joint future samples, capturing dependencies between positions and, for molecules, between atom types and coordinates, at a one-time training cost. Importantly, GILC only handles discrete variables (atom types), so the property-guidance comparison is not like-for-like.

Even so, multiMFM has lower error than GILC's best reported variant on five of six properties over all outputs, and on all six after PB filtering (Table okH9-3).

Table okH9-3. QM9 property MAE against GILC's published results.

MethodCvC_vμαGapHOMOLUMO
GILC-PG (published)1.530.8082.15783374875
multiMFM, all outputs0.4440.8181.166592270395
multiMFM, PB-valid0.4150.7401.113562259372

GILC-PG is the best variant in GILC, Table 4. Its header lists α before μ, but the bound rows it copies from Bao et al. (2023) show the second column is μ and the third is α. Gap/HOMO/LUMO in meV.

Q3. Writing and figures

Thank you for catching the Figure 2 / Line 185 inconsistency and the grammatical error in the dMFM definition; we will fix both. We will also reorganize the method section around three questions: what the posterior sampler learns, how it is used during guidance, and which approximations the practical estimators make. Notation will be defined at first use, training will be separated from inference-time steering, and each table will state its evaluation population, uncertainty measure, and protocol.

Thank you for your reply. I have updated my rating from weak reject to weak accept.

Reviewer_4JmzOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper extends Meta Flow Maps to discrete sequence generation (dMFM) and joint discrete–continuous molecular generation (multiMFM). The central idea is to distill a reward-independent posterior sampler, use it to generate multiple endpoint look-aheads from an intermediate state, and estimate a reward-based guidance direction. The framework is evaluated on DNA cyclizability control, QM9 property targeting, and GEOM-Drugs fingerprint matching.

The approach is technically interesting, and reusing one posterior sampler across different rewards is attractive. However, the current evaluation does not sufficiently isolate the contribution of MFM-based guidance or establish its practical computational advantage.

Critical strengths
  • Useful extension to discrete and multimodal generation. The paper develops a concrete discrete posterior-sampling construction and combines it with continuous MFM components. Jointly modeling atom identities and coordinates is a meaningful application of this framework.

  • Reusable guidance infrastructure. The distilled MFM is independent of the reward and can be reused across property objectives. This is useful when design requirements change frequently, although suitable reward evaluators must still be available.

  • Explicit discussion of approximation limits. The paper acknowledges finite-sample bias, the relaxation involved in differentiating discrete predictions, and the distinction between network evaluation counts and actual runtime.

Critical weaknesses

The incremental contribution over existing inference-time guidance methods is not sufficiently isolated. Reward-guided generation at inference time is already well established in diffusion and flow models, including continuous guidance, particle-based steering, multimodal guidance such as TFG-Flow [A], FK steering [C] and TFG [D], and guidance for biological sequences such as Dirichlet Flow Matching [B]. The paper acknowledges several of these directions and reports selected comparisons. However, the central question remains insufficiently answered: what specific advantage does posterior distillation provide over existing guidance mechanisms, beyond extending the MFM construction to additional representations?

This requires more than comparisons with published numbers or search-based baselines. In particular, a controlled comparison between distilled MFM guidance, direct GLASS posterior guidance, and a compatible non-distilled guidance baseline, using the same generator, rewards, and wall-clock budget, would help separate improvements due to posterior estimation, distillation, and additional inference computation. Without this evidence, the practical advantage of the proposed extension remains unclear.

A. TFG-Flow: Training-free Guidance in Multimodal Generative Flow B. Dirichlet Flow Matching with Applications to DNA Sequence Design C. A General Framework for Inference-time Scaling and Steering of Diffusion Models D. TFG: Unified Training-Free Guidance for Diffusion Models

Questions for authors

Under the same base generator, reward evaluator, and validity protocol, how does multiMFM compare with best-of-N, direct GLASS guidance, and compatible diffusion/flow guidance baselines at matched wall-clock cost? Please include the additional training cost of MFM distillation.

How does the total compute required to maintain a fixed level of target attainment and diversity grow with sequence length and reward selectivity? Does the favorable scaling observed in gradient diagnostics translate into end-to-end design performance?

3 comments
AuthorsOpenReview ↗

Comparison at matched cost and scaling of compute to maintain fixed target attainment

Thank you for pushing us to isolate what posterior distillation adds.

Q1. What does posterior distillation contribute?

To isolate distillation, we fixed the base generator, reward, guidance rule (Eq. (4)), window, strength, and billed budget, and changed only how the posterior look-aheads are produced: by the distilled student, or by direct GLASS integration with four splits between posterior samples (N) and integration steps (S). dMFM has 14% lower error than the best GLASS split at y=0.30 (Table 4Jmz-1). More GLASS steps give better samples but fewer of them, while the student gives a good sample in one evaluation.

Table 4Jmz-1. Controlled distillation ablation at billed budget 848.

Posterior samplerNSOracle error, y=0.30
dMFM16 × 10.1705 ± 0.0037
Direct GLASS16 × 10.5681 ± 0.0074
Direct GLASS8 × 20.1973 ± 0.0028
Direct GLASS4 × 40.2722 ± 0.0067
Direct GLASS2 × 80.4122 ± 0.0079

Held-out oracle, mean ± SEM. All rows use the same guidance strength.

dMFM is also faster. With Eq. (4), it takes 35% less GPU time than the most accurate GLASS allocation (8 × 2), and Eq. (4) runs 49% faster than Eq. (3) because it needs no backward passes (Table 4Jmz-2).

Table 4Jmz-2. Batched DNA sampling time.

Posterior sampler and guidance ruleNSGPU ms/output
dMFM, Eq. (4)16 × 15.96
dMFM, Eq. (3)16 × 111.74
Direct GLASS (best split), Eq. (4)8 × 29.17

Single H200, length 50.

On molecules, Table 4Jmz-3 compares methods with the same generator, guide, oracle, and PoseBusters protocol. Billed evaluations count the FLOPs of all forward and backward passes, including the reward, in units of one generator forward pass. At the same PB-valid floor 75% and compute, multiMFM outperforms benchmarked methods at higher accuracy.

Table 4Jmz-3. Shared-generator polarizability comparison.

MethodBilled budgetAll-output MAE ± SDPB validity
multiMFM4701.667 ± 0.05778.5%
DPS4732.477 ± 0.07279.5%
Best-of-45123.592 ± 0.09298.2%
FK-811446.972 ± 0.44198.4%

MAE in Bohr³, over all outputs with no filtering or selection.

Distillation cost. Training the QM9 student took about 73 minutes on one H200. The student has 363k parameters and is shared across all six QM9 property objectives, so this cost is paid once per base generator, not once per reward.

Q2. Length, selectivity, and diversity

Yes, the favorable scaling in the gradient diagnostics does translate into end-to-end design performance. To show this, we measured the sampling time needed to hold accuracy and diversity fixed as length and target difficulty grow. We use a four-motif DNA task (TTTTTC, AAAATT, GAATTC, CACGTG) whose reward is high only when all four motifs are present. The target is the unguided mean plus k standard deviations (k=3, or the harder k=5). We fixed the requirement of normalized error ≤ 0.5 and at least 95% unique outputs.

Increasing sequence length 8X from 50 to 400 bps increases dMFM's sampling time by 4.0X at k=3 and 5.5× at k=5, while the posterior sample count it needs stays fixed (N=8 and N=16). The sampling compute at fixed accuracy increases linearly as sequence length increases. Where FK also qualifies, dMFM is faster; at k=5, FK does not reach the accuracy requirement at any tested budget.

Table 4Jmz-4. Sampling cost at fixed accuracy and diversity requirements.

LengthkdMFM s/100 outputsdMFM errorFK s/100 outputsFK error
5031.660.261 ± 0.0162.050.453 ± 0.038
20033.220.211 ± 0.0118.360.328 ± 0.076
40036.680.294 ± 0.02019.450.241 ± 0.023
5052.040.359 ± 0.026Not reached1.436 ± 0.060
20055.170.260 ± 0.020Not reached1.116 ± 0.111
400511.230.412 ± 0.024Not reached0.954 ± 0.056

Error is the mean absolute target error divided by the unguided standard deviation. Both methods share the base model, reward, and targets, and use no backward passes. Times are on one H200 and include reward evaluation; at k=5, FK error is its best over tested particle counts.

On molecules, the same trend holds for selectivity: in a radius-of-gyration task on GEOM-Drugs with eight targets of increasing difficulty, best-of-4 wins on the two easiest, while multiMFM wins on all six harder ones, with 19% lower error at the hardest.

AuthorsOpenReview ↗

Follow up on our response

Dear Reviewer 4Jmz,

Thank you again for reading our paper and for pushing us to isolate what posterior distillation actually contributes. We have run the controlled ablation you asked for: holding the base generator, reward, guidance rule, window, strength, and budget fixed and changing only how the posterior look-aheads are produced, dMFM reaches 14% lower oracle error than the best direct-GLASS split and takes 35% less GPU time than it. We have also compared multiMFM against DPS, best-of-4, and FK-8 under a shared generator, oracle, and PoseBusters protocol at matched billed budget, and we report the one-time distillation cost (about 73 minutes on one H200 for a 363k-parameter student that is shared across all six QM9 objectives).

On your scaling question, we measured the sampling time needed to hold accuracy and diversity fixed as length and reward selectivity grow. An 8× increase in sequence length, from 50 to 400 bp, costs ~5x sampling time with the posterior sample count unchanged, while FK never reaches the accuracy requirement at k=5 at any tested budget, and on GEOM-Drugs multiMFM wins on the six harder of eight radius-of-gyration targets. So the favorable gradient diagnostics do carry over to end-to-end design performance.

Would be great if you can take a minute to see if these address your concerns. We are happy to clarify further or run additional comparisons if needed.

Best, Authors

Reviewer_4JmzOpenReview ↗

thanks for the rebuttal

The authors did a great job at addressing my concerns. I was already biasing toward a weak acceptance. In light of other reviews, I am happy to keep my original score.

Reviewer_QF8KOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Recent one-shot samplers, such as Meta Flow Maps (MFMs), have demonstrated scalable inference-time steering through efficient estimation of reward gradients. This paper extends the MFM framework to biological design tasks by introducing discrete MFM (dMFM) and multimodal MFM (multiMFM). The authors evaluate dMFM on DNA sequence design with target mechanical properties and multiMFM on molecular generation with target physical properties. In both settings, the proposed methods improve target attainment while reusing a single reward-independent MFM across different objectives without task-specific retraining.

Critical strengths
  1. Scalable inference-time steering for multimodal biological data.

The paper extends Meta Flow Maps to discrete and joint discrete–continuous settings, enabling reward-guided generation of DNA sequences and molecules with both atom types and 3D coordinates. A key advantage is that the same pretrained MFM can be reused across different reward functions without task-specific fine-tuning. The experiments on DNA, QM9, and GEOM-Drugs demonstrate the practical applicability of the proposed framework across different modalities and design objectives.

Critical weaknesses
  1. Inexact reward gradients for discrete variables.

The proposed method relies on continuous relaxation of categorical variables to compute reward gradients. However, differentiating through simplex-valued predictions does not generally yield the exact gradient associated with categorical posterior samples. Although the authors acknowledge this limitation, the paper does not sufficiently investigate the resulting approximation error or its impact on steering performance. This raises concerns about the theoretical justification of the discrete guidance formulation and its reliability in practice.

Questions for authors
  1. Could you quantify the impact of the continuous relaxation on gradient estimation and steering performance? In particular, how does the current estimator compare with the gradient-free alternative in Eq. (4)?

  2. Could you report wall-clock inference time and computational costs, including reward evaluations and backward passes, in addition to NFEs? This would help substantiate the scalability claims.

  3. Table 2 combines published baseline results with multiMFM results evaluated only on PB-valid molecules. Could you clarify the validity filtering used for the baselines and provide comparisons under a consistent evaluation protocol, reporting both MAE and validity?

3 comments
AuthorsOpenReview ↗

Answers to Question 1-2

Thank you for raising these points. The gradient-free rule in Eq. (4) evaluates the reward on hard one-hot sequences, so it sidesteps the relaxation. Gradient-free rule matches the accuracy of a tuned Eq. (3), and because it needs no backward passes, it cuts sampling time by 49%. Below we also give a full cost breakdown and show that multiMFM's low errors hold before any validity filtering.

Q1. Continuous relaxation and Eq. (4)

We agree that a reward evaluated on a soft categorical prediction does not in general give the exact categorical gradient. To measure how much this matters, we compared Eq. (3) with the gradient-free Eq. (4), giving both the same tuning sweep over guidance windows and sample allocations. Eq. (4) feeds hard one-hot sequences to the reward and does not differentiate through either the reward or the posterior sampler. In the following tables Billed Evaluations measure total compute: we count the FLOPs of every forward and backward pass, including those of the reward, and express them in units of one generator forward pass.

Table QF8K-1. DNA guidance accuracy with both rules tuned the same way.

Guidance ruleBilled budgetOracle error ± SEM
dMFM, Eq. (4)8480.1281 ± 0.0015
dMFM, Eq. (3), tuned9000.1321 ± 0.0029
dMFM, Eq. (3), published allocation8600.1443 ± 0.0026

Cyclizability target y=0.30, held-out oracle.

After tuning, the two rules reach similar accuracy, and Eq. (4) does so with a 6% smaller budget. Discrete steering therefore does not depend on the soft-input reward gradient. The gradients themselves also agree: in the cases where both rules give stable estimates (at t=0.7 or 0.8), the mean cosine similarity between their steering directions is 0.87.

Q2. Wall-clock time and complete cost

Table QF8K-2 breaks down the cost of each DNA sampler, including reward evaluations and backward work. With identical forward counts, dropping the backward passes (Eq. (4) vs. Eq. (3)) cuts GPU time by 49%. The distilled sampler also uses 35% less GPU time than the most accurate direct-GLASS allocation we tested. All timings include reward evaluation during guidance.

Table QF8K-2. DNA forwards, backward work, and GPU sampling time.

MethodGenerator forwardsReward forwardsBackward workGPU ms/output
dMFM, Eq. (4), N=1696 base + 752 student75305.96
Direct GLASS, Eq. (4), N=8, two steps848 base37709.17
dMFM, Eq. (3), N=1696 base + 752 student75347 joint traversals†11.74

Single H200, length 50, batched sampling. †Each backward traversal covers all 16 posterior draws and their reward evaluations. Timing and accuracy are separate runs.

At the same budget of 848, with the generator, reward, Eq. (4), guidance window, and strength all fixed, dMFM also gives 14% lower error than direct GLASS (0.1705 vs. 0.1973). This isolates the gain from distillation.

At the same PB-valid floor 75% and compute, multiMFM outperforms benchmarked methods at higher accuracy.

Table QF8K-3. Molecular accuracy at comparable inference budgets.

ExperimentMethodBilled evaluations/outputAll-output MAEPB validity
QM9 polarizabilitymultiMFM4701.667 ± 0.05778.5%
QM9 polarizabilityDPS4732.477 ± 0.07279.5%
GEOM radius of gyration, k=12multiMFM3760.659089.8%
GEOM radius of gyration, k=12Best-of-44000.809784.4%

The posterior sampler is trained once per base generator and reused across rewards. Training the QM9 student took about 73 minutes on one H200.

AuthorsOpenReview ↗

Answers to Question 3

Q3. Validity filtering in Table 2

First, multiMFM produces more PB-valid molecules than TFG-Flow on every property (Table QF8K-4). The MAEs reported by TFG-Flow are computed over all of its outputs, which include up to 20% PB-invalid molecules, some of them disconnected.

Table QF8K-4. PB validity (%) on the six QM9 properties.

MethodCvμαGapHOMOLUMO
TFG-Flow83.0 ± 0.383.8 ± 0.580.4 ± 0.680.9 ± 0.784.2 ± 0.987.3 ± 0.4
multiMFM90.390.690.590.891.590.9

So that our side of the comparison does not depend on filtering either, we recomputed multiMFM's errors on all outputs as well as on the PB-valid subset of the same runs (Table QF8K-5). Filtering changes the error by only 4–10%, and even without filtering, multiMFM's error is below the published TFG-Flow value on five of six properties.

Table QF8K-5. Filtering audit on the same molecular runs.

PropertyTFG-Flow published MAEmultiMFM all-output MAE ± SDmultiMFM PB-valid MAE ± SDmultiMFM PB validity
Cv1.750.444 ± 0.0060.415 ± 0.01290.3%
μ0.8170.818 ± 0.0570.740 ± 0.05390.6%
α2.321.166 ± 0.0301.113 ± 0.01890.5%
Gap804592 ± 12562 ± 690.8%
HOMO364270 ± 7259 ± 1391.5%
LUMO941395 ± 16372 ± 1190.9%

For a comparison under a single protocol, the DPS control in Table QF8K-3 uses the same generator and reports errors over every output, with no filtering or selection. multiMFM still has 33% lower error there.

For steer-search, errors are computed on the selected candidates, while PB rates refer to the underlying trajectories. We will update the manuscript with same evaluation protocols as TFG-Flow.

Thank you for your detailed response and clarifications. I appreciate the authors' efforts to address my concerns. I will maintain my rating of weak accept, leaning toward a positive assessment.

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

This paper extends Meta Flow Maps to discrete and multimodal settings for biological sequence and molecular design. The authors propose dMFM for DNA sequence design and multiMFM for molecular generation with both atom types and 3D coordinates. The main advantage is that a single reward-independent MFM can be reused for different design objectives at inference time, without retraining the base generative model. Experiments on DNA sequence design, QM9, and GEOM-Drugs show improvements in target property optimization compared with unguided generation and several baselines.

Critical strengths
  • The idea of using one pretrained MFM for different reward functions is interesting and practical. It avoids training a separate generative model for every design objective.

  • The authors test the method on both DNA sequences and molecular generation, which shows its potential beyond standard continuous generation tasks.

Critical weaknesses
  • The quality of gradient estimation is still a concern. For DNA design, the cosine similarity with reference gradients is only 0.18–0.32 at the selected sampling budget (N=8). It is therefore unclear how reliable the estimated gradients are, especially for more complex or larger design problems.

  • The molecular design results mainly rely on learned property predictors. It is unclear whether the generated molecules truly have the desired properties or simply exploit errors in the predictors. Also, reward guidance decreases molecular validity, and the proposed search strategy only partially addresses this issue.

Questions for authors
  • The gradient estimation accuracy for DNA design seems relatively low at N=8. How much does increasing the number of posterior samples improve the final design performance? Is N=8 sufficient for longer sequences?

  • Have the authors checked the properties of generated molecules using independent quantum-chemical calculations, rather than only the reward predictors used during optimization?

  • Could the authors provide a more direct comparison of computational costs with existing methods, including wall-clock time and the number of reward evaluations?

  • How well would the method work when optimizing multiple objectives simultaneously, especially when molecular validity and different property objectives conflict?

4 comments
AuthorsOpenReview ↗

Answers to Q1-3

We thank the reviewer for these questions, which led to several new experiments. We wish to highlight Table ifYC-5 which shows multiMFM outperforms TFG-Flow using exactly TFG-Flow’s multiobjective benchmark and reward setups. Independent quantum chemistry confirms that the generated molecules reach their requested polarizability, not just a low predictor error. On DNA, final design quality is already close to saturated at N=8 posterior samples. At matched compute, multiMFM gives 33% lower polarizability error than DPS.

Q1. Gradient accuracy, sample count, and longer sequences

The low cosine similarity at N=8 matters less for final designs than the number suggests. Table ifYC-1 shows the final cyclizability error as N grows. Going from 2 to 4 samples cuts the error by 36%, but going from 8 to 16 cuts it by only 4%. For reference, direct GLASS reaches a cosine similarity of 0.243 at N=8, below dMFM's 0.286.

Table ifYC-1. Posterior sample count and final cyclizability error.

Posterior samples NMean oracle error
20.2192
40.1408
80.1335
160.1281

For longer sequences, dMFM with N=16 has lower error than both search baselines at every length from 50 to 400 (Table ifYC-6 under Q4).

Q2. Independent quantum chemistry

Yes. We checked the polarizability of generated molecules against independent quantum-chemical values (Table ifYC-2). For the 117 molecules whose structures match a QM9 entry, we used the existing DFT labels, and for the other 180 we ran xTB. As a control, we randomly reassigned the requested targets among the molecules within each group, answering whether each molecule is closer to its own target than to someone else's.

Table ifYC-2. Independent checks of polarizability targeting.

Generated moleculesIndependent checkError to original targetsError after shuffling targets
With a QM9 match (117)Existing DFT labels1.4688.972
Without a QM9 match (180)Calibrated xTB calculations4.03610.77

MAE in Bohr³, lower is better. xTB values are calibrated on 400 QM9 molecules.

Against the original targets, the error is 84% lower than after shuffling in the DFT check and 63% lower in the xTB check. So the molecules do reach their requested polarizability, including those with no QM9 match. The independent errors are higher than what the learned oracle reports, so the predictor is somewhat optimistic.

Q3. Computational cost

Billed evaluations measure total compute: we count the FLOPs of every forward and backward pass, including those of the reward, and express them in units of one generator forward pass. At the same PB-valid floor 75% and compute, multiMFM outperforms benchmarked methods at higher accuracy.

Table ifYC-3. Target accuracy at comparable inference budgets.

ExperimentMethodBilled evaluations/outputTarget errorPB validity
QM9 polarizabilitymultiMFM4701.667 ± 0.05778.5%
QM9 polarizabilityDPS4732.477 ± 0.07279.5%
GEOM radius of gyration, k=12multiMFM3760.659089.8%
GEOM radius of gyration, k=12Best-of-44000.809784.4%
DNAdMFM, 16 samples × 1 step8480.1705 ± 0.0037—
DNADirect GLASS, 8 samples × 2 steps8480.1973 ± 0.0028—

At deployment, higher PB validity is required and the number of NFEs for all methods needs to increase.

For wall-clock time, the gradient-free rule in Eq. (4) cuts batched sampling time by 49% relative to Eq. (3). It uses the same forward passes but no backward passes, and its accuracy is similar. The distilled sampler also takes 35% less GPU time than direct GLASS (Table ifYC-4).

Table ifYC-4. Measured DNA sampling time.

MethodPosterior samples × stepsGPU ms/output
dMFM, Eq. (4)16 × 15.96
Direct GLASS, Eq. (4)8 × 29.17
dMFM, Eq. (3)16 × 111.74

Single H200, length 50, batched sampling. The posterior sampler is trained once and reused across rewards.

AuthorsOpenReview ↗

Answer to Q4

Q4. Multiple objectives and validity

Thank you for suggesting this benchmark. We followed the multiobjective setup in Table 2 of TFG-Flow and steered two properties at once in three tasks (Cv + , + , and + ). multiMFM outperforms TFG-Flow on all three tasks, with higher PB validity (Table ifYC-5).

Table ifYC-5. Multiobjective steering on QM9 (TFG-Flow benchmark).

MethodTask 1: C_v + μTask 2: Δε + μTask 3: α + μ
C_vμΔε (meV)μαμ
Cond-Flow (published)4.96 ± 0.071.57 ± 0.011456* ± 191.57 ± 0.009.33 ± 0.011.54 ± 0.01
TFG-Flow (our rerun)1.956 ± 0.0151.030 ± 0.034980.7 ± 29.61.046 ± 0.0552.609 ± 0.2381.169 ± 0.034
PB validity83.1%85.2%85.8%
multiMFM0.668 ± 0.0120.682 ± 0.016525.3 ± 11.20.706 ± 0.0211.380 ± 0.0020.723 ± 0.022
PB validity91.0%89.2%88.7%
multiMFM-SS†0.625 ± 0.0100.655 ± 0.010484.5 ± 8.20.670 ± 0.0191.352 ± 0.0340.689 ± 0.016
PB validity†95.5%95.1%95.4%

*TFG-Flow’s published Δε values are in milli-Hartree, not meV; converted, they are 1456 meV (Cond-Flow) and 1263 meV (TFG-Flow).

On DNA, we extended the manuscript's two-motif reward to four motifs (TTTTTC, AAAATT, GAATTC, CACGTG), combined with a smooth minimum so the reward is only high when all four are present. dMFM lowers error by 84% relative to FK-16 at length 50 and stays ahead of FK-16 and best-of-32 up to length 400, with a smaller budget than either (Table ifYC-6). Because the posterior look-ahead scores each continuation against the full specification, guidance moves toward sequences that meet all the requirements together.

Table ifYC-6. Four-motif conjunction error across sequence lengths.

LengthUnguideddMFM, Eq. (4)FK-16Best-of-32
500.17430.01150.07210.1046
1000.12850.00690.04790.0744
2000.09700.00540.03480.0542
4000.06960.00600.02370.0380

Billed budgets: dMFM 848, FK-16 1536, best-of-32 3072.

For validity, guidance strength sets the trade-off with property accuracy. Pushing strength past the range makes both polarizability error and PoseBusters validity worse (Table ifYC-7), so ablating guidance strength gives a practical way to set the strength using both accuracy and validity.

Table ifYC-7. Molecular property accuracy and validity in the strength ablation.

Guidance strengthPolarizability MAEPB validity
2.51.64478.5%
2.751.74673.6%

512 molecules; MAE in Bohr³.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer ifYC,

Thank you so much for reading through our paper and for the questions. We have ablated the number of samples NN, simulated the properties of designed molecules with direct quantum calculations with xTB, and compared costs with existing methods. Thank you for also proposing the multiobjective experiment which we have done exactly following TFG's protocol and showed improvement, and our PB-validity is also higher. Would be great if you can take a minute to see if they address your concerns. We are happy to clarify further if needed.

Best Authors

Dear Reviewer ifYC, could you check if the responses from the authors have addressed your concerns?

Comments on the paper

2 comments
Area_Chair_on6NOpenReview ↗

Reviewer-author discussion kickoff

Dear Reviewers,

Thank you for your review. The authors have now provided their rebuttal. We would appreciate it if you could take some time to read their responses and engage in the discussion, particularly on whether your concerns have been adequately addressed.

Thank you for your time and contributions to the review process!

AuthorsOpenReview ↗

Summary of revised manuscript

Dear Area Chair and Reviewers,

Thank you for the detailed feedback. We have incorporated the rebuttal experiments into the revised manuscript, with changes marked in blue. The main revisions address four questions raised in the reviews.

What does posterior distillation add? On DNA, we held the generator, reward, guidance rule, and sampling budget fixed. Distillation reduced target MAE from 0.1973 for the best tested direct-GLASS configuration to 0.1705. In separate timing measurements, sampling time fell from 9.17 to 5.96 ms per output (§3.1).

How comparable are the molecular results? Table 2 now reports MAE over all generated outputs, with standard deviations across three seeds for every QM9 property. We distinguish published baselines from comparisons using our own generator. On the shared generator, multiMFM achieved polarizability MAE of 1.667 versus 2.477 for DPS at nearly identical generator compute (§3.2).

Does steering work for multiple objectives, and do the gains survive an independent check? On all three paired-property tasks, multiMFM improved both property errors and molecular validity relative to our TFG-Flow rerun (Table 4). For polarizability, independent DFT labels and calibrated xTB estimates both placed generated molecules closer to their requested targets than to shuffled targets (Figure 5).

How does runtime scale at fixed quality? Figure 6 reports measured sampling time at common accuracy and diversity thresholds. At 400 bp and the three-standard-deviation target, the fastest tested passing configurations took 66.8 ms per output for dMFM and 194.5 ms for Feynman–Kac steering.

We also clarified the assumptions behind the ideal steering equations and documented the guidance modifications used in practice. The appendix gives the experimental settings, uncertainty estimates, and compute accounting behind these comparisons.

Thank you for your time and consideration.

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.