acceptodds
Under review as a conference paper at ICLR 2027

LLM-to-Flow Distillation: Teaching Natural Language Goals to Scientific Generators

Abstract

Generative discovery in science requires coupling high-level scientific reasoning with the ability to generate samples respecting the structure of a design space. Yet these capabilities remain largely separated: foundation LLMs encode broad chemical and biological knowledge, while domain-specific flow- and diffusion-based generators capture the priors needed to produce valid sequences and geometries. We study how to align such scientific generators to high-level natural-language discovery goals, without requiring an explicit proxy reward, which is often unavailable or poorly captured by learned predictors in experimental sciences. We introduce LLM-to-Flow Distillation (LFD), in which a frozen LLM acts as an adaptive teacher for a domain-specific generator. Given a goal such as *'generate peptides with promising anticancer activity'*, the teacher constructs a curriculum of intermediate sub-goals aware of the flow’s current capabilities, and progressively directs them toward the original discovery goal. At each round, the LLM judges sampled candidates under the current objective. These judgments induce preferences that are iteratively distilled into the generator through DPO. This provides a dense learning signal even when the initial domain-specific model places negligible probability on designs satisfying the final goal, enabling adaptation toward poorly covered, out-of-distribution regions. Experiments on therapeutic peptides and small molecules structure design demonstrate that LFD enables a new form of generative control: aligning arbitrary domain-specific scientific generators to scientific discovery tasks expressed via high-level natural-language goals.

Primary area:
generative models
Keywords:
flow model fine-tuning, diffusion model fine-tuning, reward-based adaptation, generative discovery, scientific discovery

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_8MVfOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper couples the reasoning abilities of the frontier LLM models with the generative abilities of the flow and diffusion models for natural language guided scientific discovery. Given knowledge about the current abilities of the generator through samples and their observables, LLM proposes the next sub-goal to optimize the generator on and provides ordinal feedback for generator training through DPO. The frozen LLM, based on its broader scientific knowledge, proposes the sub-goals that eventually lead towards the user-specified natural language goal. The authors evaluate the proposed pipeline, LFD, on discovering therapeutic peptides and small molecules. Across 5 case studies, LFD outperforms the proxy reward guided baselines on the domain specific metrics while satisfying the goal descriptions in the input prompt.

Critical strengths
  • Well-motivated problem statement:
  1. Shifting from proxy reward functions to a more coarse natural language goal is important for scientific applications.
  2. The challenges of miscalibrated scores, negligible density over the goal-satisfying region are rightly identified and form a justified basis for pipeline design choices.
  • Goal 5 is a strong evidence in favor of the LFD pipeline
  1. The pre-trained model yields no hits in 38,500 samples; LFD finds one at round 6.
  2. Even the reward-guided baselines fail despite an LLM-generated binary reward for the full goal, which isolates the adaptive curriculum’s contribution.
Critical weaknesses
  • Unexplained experimental design choices:
  1. How are the number of rounds decided? The performance is sensitive to the number of rounds provided to the pipeline
  2. There is no information on the number of random seeds tested per experiment and how many candidates are finally proposed at the end of all the rounds
  3. Missing simple ablation on LFD but with only the user specified goal and no adaptive curriculum.
  • Several observables are trained predictors and implicitly steer the curriculum
  1. Peptide observables include predicted developability bands. Further, the non-hemolysis band implicitly informs about the toxicity metric.
  2. The curricula track the supplied descriptors closely, so the observables implicitly steer the search. No ablation over the observable set.
  3. Prompts for the LLM are highly verbose and customised per task. For example, Goal 5 explicitly forbids the carbon-only linkage requirement, a domain knowledge which wasn’t necessary in the prompt.
Questions for authors

Clarification Questions:

  1. For the sub-goal generation, how is it ensured that the batch generated by the student is representative of the generator’s current abilities? Given the finite sample regime, it could be a limited view of the generator's terminal density and hence the sampling noise could drive the curriculum. Would a larger batch or a different sampling seed recover the same sub-goal sequence?
  2. How does the DPO training deal with the skewness in the sizes of the positive and negative candidate sets, that influences the size of preference dataset?
  3. The experimental setup mentions that Gemini 3.1 Pro was used to construct the curriculum but Gemini 3.7 Flash was used for LLM judgements. What is the rationale behind using different LLM backbones for the two tasks?
  4. What is the standard deviation in table 1 and 2 over? It seems to be over multiple proposed candidates instead of different end-to-end runs of the discovery process, which raises a question on performance reproducibility.
  5. What is the distance metric in table 1?
  6. Couldn’t exclusion of abstained candidates lead to loss of potential pathway to the goal? Is it possible that a candidate is neither positive nor negative at the current round, but allows for future promising candidates? Further, even the candidates with unstable classification on 2 runs are discarded. What fraction of candidates are discarded on average in the experiments?
  7. What happens when the specified goal is unachievable, due to infeasible constraints? How does the pipeline handle this empirically?

Suggested experiments:

  1. How does the performance of LFD and the designed curriculum vary with the number of rounds T, since Appendix D mentions some subgoals are reused for multiple rounds.
  2. Currently the main case study results are on the Gemini models. Providing results on at least 1 more LLM would be useful to evaluate the robustness of the pipeline to the choice of LLM backbone.
  3. It might be interesting to see an ablation where the observables just contain a compact representation of the candidate, without any other measured properties.
  4. Continuing the previous point, how sensitive is the performance to the provided observables? From the provided curricula, it appears that the provided observables bias the intermediate goals of the LLM towards them, providing an implicit external guidance on the trajectory to follow.

Future Work

  1. It could be an interesting future study on what observables are most important in the prompt and is there a pattern that surfaces across the domains.

Overall, I think the work is pretty interesting and I'll be open to improving my score based on the author responses to the questions.

6 comments
AuthorsOpenReview ↗

Rebuttal (1/3): Ablations and clarifications on LFD pipeline

We thank the Reviewer for recognizing the problem tackled of adaptation to natural-language goals as important, that we recognized the right challenges, and that Goal 6 (previously 5) is a strong evidence in favor of LFD. In the following, we aim to sharply address all limitations and questions mentioned within the review.


Number of rounds LFD is typically run continually: it does not require prescribing a fixed curriculum length, and TT can serve as an upper limit determined by available resources or need. In practice, we stop when the LLM judges that sufficiently many generated designs satisfy the final goal according to its belief, or when the computational budget is reached. For instance, for Goal 6, we stopped the run at round 6 when the LLM judged that a generated candidate matched the final goal requirements, for which we then confirmed this hit using the held-out checker (Figure 10).


Ablation of LFD without an adaptive curriculum We added a final-goal-only baseline using LLM categorical feedback, reported in the new subsection “Ablation on directly optimizing the final goal without a curriculum” (Appendix E.7, Table 11) of the updated manuscript. The LLM judges candidates directly against the user-specified goal, without intermediate sub-goals. We evaluate this baseline on Goal 2 (CPP) and Goal 5 (GSK3β):

MetricFinal-goal-only LFDLFD
CPP: pLM4CPPs (%) ↑14.21 ± 1.9469.06 ± 6.36
CPP: HemoPI2 (%) ↓41.27 ± 0.8837.02 ± 1.69
GSK3β: GNINA ↑6.17 ± 0.086.77 ± 0.11
GSK3β: Vina ↓−7.43 ± 0.10−8.84 ± 0.16
GSK3β: rDock ↓−8.81 ± 0.37−12.87 ± 0.57

This baseline clearly underperforms LFD on both tasks. Moreover, final-goal-only LFD cannot form preference pairs when no positive samples are available. Goal 6 clearly demonstrates this limitation: the initial FlowMol3 model generated no final-goal positives in more than 38,000 samples. Intermediate sub-goals provide the supervision needed to begin adaptation in this setting.


Trained predictors as observables and ablation over the observable set (compact representation) We regard these observables as task-agnostic because they describe general peptide properties relevant across most peptide-design tasks, rather than properties specific to anticancer activity. We agree, however, that this distinction is ambiguous here. To address this concern, we added an ablation removing the developability and non-hemolysis bands from the observables. We report this ablation under “Ablation on the observables” (Appendix E.1, Figure 11) in the updated manuscript. We evaluate six rounds of Goal 1, retaining the sequence and remaining physicochemical descriptors. At round six, Figure 11 shows mean AntiCP2-alt scores of approximately 43% without these bands and 40% with them, corresponding HemoPI2 scores are approximately 44% and 43%, respectively. The scores remain similar, with overlapping 95% confidence intervals, although the proposed sub-goals change. Thus, the supplied observables influence curriculum construction, but LFD still improves predicted anticancer activity without these bands while maintaining predicted hemolysis comparable to the run that includes them. We thank the Reviewer for motivating this ablation.


LLM prompts verbosity The prompts provide scientific context and prior domain knowledge, together with general instructions (Appendix B.2). Making this context explicit allows scientists to inspect and validate the assumptions guiding discovery. In the example raised by the Reviewer, now Goal 6, the carbon-only linkage instruction prevents the controller from adding a requirement absent from the user’s goal. It does not impose an additional chemical restriction. Such prompt templates can easily be drafted by an LLM from the discovery goal and can be reviewed by a scientist. Their necessary detail will depend on the teacher’s capabilities: we expect more capable models may require less biological and chemical context, although prompt simplification has not been evaluated here.


Stochasticity of finite-sample generated batch for subgoal generation Indeed a finite batch estimates the generator’s current capabilities (or coverage) without fully describing its density. As in minibatch stochastic optimization (e.g., stochastic gradient descent), noisy sample estimates can nevertheless support useful updates. Both candidate sampling and LLM proposals are stochastic: larger batches can reduce sampling noise, but need not recover an identical sub-goal sequence. We therefore examined curriculum variation across different seeds. Although the proposed sub-goals differ, the resulting curricula consistently improve evaluations against the original user-specified discovery goal, supporting the practical usefulness of finite-batch feedback despite variation in the subgoal sequence.

AuthorsOpenReview ↗

Rebuttal (2/3): Ablations, clarifications on replicability and metrics, and experiment on discarded judgments

Ablation on DPO training with unbalanced data We added “Ablation on sensitivity to unbalanced positive-negative class sizes” in Appendix E.6 of the updated manuscript. We compare 40 labelled candidates split into 20/20, 36/4, or 4/36 positives/negatives, and 80 candidates split into 40/40, 60/20, or 20/60. Each setting runs DPO from the same pre-trained model and is evaluated on 1,000 fresh peptides. Figure 13 shows similar results across these settings. Across all splits, CPP-predictor scores remain around 13% and predicted hemolysis around 42%, showing that DPO achieves comparable performance even with 9:1 or 1:9 positive-to-negative ratios in these experiments.


Why using Gemini 3.1 Pro and 3.7 Flash for different roles We used Gemini 3.1 Pro for curriculum construction, while assigning the more frequent candidate judgments to the cheaper Gemini 3.7 Flash to reduce API costs.


Standard deviation, seed tested, replicability of generative discovery The standard deviation in Table 1-2 is indeed over multiple proposed candidates. We want to clarify the replicability of our results. As is currently the standard in the community of discovery via generative models [1,2,3], we report in the manuscript the results of the best run obtained, akin to the so-called BoN (Best-of-N) procedure. This was done by evaluating LFD models after each subgoal gtg_t. In particular, the model is sampled, and the generated designs are judged by the LLM on the final goal gg. Then, we estimated with such judgments PLLM(g)P_{LLM}(g), we kept running the most promising runs, and killed the others. Notice that in this procedure there was no exposure of the held-out predictor used for evaluation. Such a procedure is also consistent with our limited API and compute budget, a constraint which is also very likely in practical applications where LFD can be applied. We stress that the less promising (and thus killed) runs could have resulted in a good (with respect to the downstream held-out predictor) final model and samples as well, yet due to the limited resources we have not continued such runs. We also note that the number of runs that manage to achieve good results in the end are always more than one, and the exact ratio (over all runs tested) depends on the difficulty of the goal tackled. In particular, Goal 6 was the most challenging one and although multiple runs could succeed at it (2/7 runs), several runs (5) have been killed at intermediate steps to conserve computational resources and API economic costs, as explained earlier. In simpler goals such as Goal 1, where optimality is not a binary signal, all the runs were progressing well according to the LLM assessment, and the killed ones were just the slowest ones at doing so.


Distance used in Table 1 We added an explicit definition in the paragraph “Distance from prior” in Appendix D.1 of the updated manuscript. The metric is the square root of the Jensen–Shannon divergence between the amino-acid frequency distributions of generated peptides and a reference set sampled from the pre-trained generator. Using base-2 logarithms gives a distance between 0 and 1, with larger values indicating a greater shift in amino-acid composition away from the prior. This metric is standard in peptide design as supported by the referenced works in Sec. 5.


Exclusion of abstained candidates Uncertain candidate may still lead to promising designs. Nonetheless, notice that based on the DPO loss, abstain candidates receive no direct penalty in the LFD update and can be re-sampled by the generator in future rounds. For Goals 1 (peptides), and 5 (molecules), we report the following respective % of (overall) candidates discarded per round of 8.0% and 11.2%. This filtering aims to improve the LLM reliability by avoiding judgements self-classified as uncertain, particularly where the LLM lacks sufficient evidence or expertise. An alternative would be to treat a subset of abstentions as positive, encouraging exploration in the spirit of “optimism in the face of uncertainty”. This could help explore additional pathways, but might also introduce incorrect preferences and increase computational costs. We empirically try to validate this idea by running LFD on Goal 1 by exactly turning into positives all the abstantions and the samples with at least one positive judgments. Such LFD implementation achieves comparable performance (although with a different subgoal progression) as the LFD run reported in the manuscript. In particular, AntiCP2-alt is 54.87 ± 1.22 and HemoPI2 is 40.42 ± 0.86 (respectivaly compared to the 53.54 ± 3.32 and 42.48 ± 2.04 of the manuscript). However, in this end-to-end run, it is difficult to isolate the role of the optimism due to the stochasticity of the LLM (and thus the generated subgoals) and of the peptides sampled at each step. Overall, we regard this as an interesting and promising extension of LFD.

AuthorsOpenReview ↗

Rebuttal (3/3): LFD with GPT-6.1 Sol, unachievable goals, and references

Unachievable goal It is straighforward to add as a preprocessing step of the LFD method a simple goal checker, which utilizes the frontier LLM scientific (e.g., chemical) and logical reasoning to assess whether the goal description admits at least a positive instance. For instance, the goal "generate small molecules with at least 2 cycles, but without cycles" clearly corresponds to an empty solutions set that can be recognized by an LLM. On the other hand, checking satisfiability of arbitrary goals is impossible in general. As a consequence, it is left to the scientist utilizing LFD the task of defining a goal believed to correspond to a non-empty solution set. Empirically, if the chosen goal indeed corresponds to an empty solutions set, LFD, run with an LLM that has sufficient understanding of the specific application, is not able to generate positive solutions. Instead, it attempts to relax the initial goal into a sequence of subgoals that allows to generate designs semantically closer to the goal description, yet not fully achieving it.


Ablation: results with another frontier LLM We added “Ablation on LLM model family” in Appendix E.3,of the updated manuscript, replacing both Gemini models with GPT-6.1 Sol, using high reasoning effort, on Goal 2 (peptides) and Goal 5 (small molecules):

MetricLFD with GeminiLFD with GPT-6.1 Sol
CPP: pLM4CPPs (%) ↑69.06 ± 6.3674.05 ± 2.37
CPP: HemoPI2 (%) ↓37.02 ± 1.6933.73 ± 0.66
GSK3β: GNINA ↑6.77 ± 0.116.69 ± 0.07
GSK3β: Vina ↓−8.84 ± 0.16−8.74 ± 0.11
GSK3β: rDock ↓−12.87 ± 0.57−12.93 ± 0.33

LFD with GPT-6.1 Sol achieves higher mean predicted CPP activity and lower mean predicted hemolysis, while retaining comparable docking scores. These results support robustness to the teacher’s model family across the two evaluated tasks. We also report the resulting curricula in the same subsection.


Let us know if these clarifications address your questions and, if so, would lead you to raise the score as you mentioned. We would be happy to clarify any additional doubts. Thanks again for your review!

References

[1] Mathematical discoveries from program search with large language models. Romera-Paredes et al., Nature 2024.

[2] Eureka: Human-Level Reward Design via Coding Large Language Models. Ma et al., ICLR 2024.

[3] Learning to Discover at Test Time, Yuksekgonul et al., ICML 2026.

Reviewer_UF1COpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces an online method for adapting a pretrained scientific flow model to a natural-language goal. The idea is: an LLM proposes intermediate goals based on the generator’s current samples, labels candidates as positive, negative, or uncertain, and the resulting preferences train the flow with DPO. Experiments cover peptide generation and small-molecule design.

Critical strengths
  1. The motivation for teaching natural language goals to scientific generators is clear and reasonable.
  2. The method is simple and direct.
Critical weaknesses
  1. Some claims (promising activity, binding, low toxicity, and therapeutic) may be overstated. The discovery results are computational results, not evidence of scientific or therapeutic success.
  2. In my understanding, the method does not provide a guarantee that its curriculum remains a valid relaxation of the final goal or reaches it.
  3. "The ground-truth value is hidden from the LLMs and used only for evaluation." How could the authors assure this statement? For example, QM9 (2014) has a high probability of being trained by SOTA LLMs.
  4. The authors did not consider the retrosynthesis, which is an important evaluation aspect for molecular design.
  5. In my opinion, this paper is more like LLM-guided preference supervision informed by pretrained scientific knowledge; the claim of using first-principles scientific reasoning is neither specified nor rigorously verified.
  6. The method is presented as aligning “arbitrary” scientific generators and handling “OOD discovery goals,” but the experiments use two generator families and a small set of tasks, with one structured-constraint example reporting a hit rather than a systematic success rate. The terminology and scope seem to be broader than the demonstrated evidence.
Questions for authors
  1. The peptide results use activity and toxicity predictors, and the molecular results use docking scores and computational checks. What evidence supports describing the designs as therapeutically active, binding, or low in toxicity?

  2. The paper states that ground-truth values were hidden from the LLMs. How did you assess or mitigate possible pretraining exposure to public datasets such as QM9?

  3. Why did the authors not evaluate synthesizability or retrosynthetic accessibility for the generated molecules?

  4. What does “first-principles scientific reasoning” mean operationally in LFD, and how did you verify that the LLM’s reasoning is scientifically reliable for the discovery goals?

  5. Given the limited number of generator families and tasks, what evidence supports the claims of alignment for “arbitrary” generators and OOD goals? For Goal 5, what is the success rate across independent runs and batches?

6 comments
AuthorsOpenReview ↗

Rebuttal (1/3): clarification on the computational results and memorization analysis

We thank the Reviewer for recognizing that the motivation for teaching natural language goals to scientific generators is clear and reasonable, and that the proposed method is simple and direct. In the following, we aim to sharply address all limitations and questions mentioned within the review.


Computational vs experimental results Within the submitted version of the work, it was already clearly stated both within the Experiments Sec. 5 (under the "Baselines and evaluation" paragraph), as well as in the Conclusion, that we perform and present only "in silico" evaluations. This is a standard term to denote computational, rather than experimental (i.e., wet-lab, or in vitro and in vivo) results. Computational evaluation is standard practice for methodological contributions at machine learning conferences, including work on molecular and biological domains.


Guarantee of valid relaxation This is primarily an experimental paper, and does not aim to formally define valid relaxations or provide theoretical guarantees for the generated curriculum. Establishing meaningful guarantees for a curriculum generated by prompting a frontier LLM would require strong assumptions and abstractions about the teacher’s behavior. We instead present a method that is to our knowledge the first to solve instances of a new problem class: fine-tuning of scientific generators (e.g., flow and diffusion models) via natural-language goals (see Fig. 1). Indeed, this highly practically-relevant problem setting might lead to new theoretical questions and modeling frameworks, to which the community can contribute, to develop understanding and/or better methods for this new problem class.


Hidden ground-truth value We thank the Reviewer for raising this point. Indeed, the previous phrasing was somewhat ambiguous. Therefore, within the updated version of the manuscript (Sec. 4) we have replaced the sentence "The ground-truth value is hidden from the LLMs and used only for evaluation." with the following correct sentence: "The LLMs is given candidates’ observables and task descriptions, but neither the computed target-property values nor access to the corresponding computational evaluators. The LLMs’ predicted scores are compared against these held-out values to assess the accuracy of their scientific judgments.". This correctly expresses that held-out predictors are not given as observables to the LLM.

While in general it is impossible to inspect the training data of frontier LLMs, we have added an analysis of memorization in the new main-text paragraph “LLM training data and memorized designs” (Sec. 5) of the updated manuscript. For the anticancer task, we assess whether LFD regurgitates sequences that the LLM teacher may have seen during pretraining. Specifically, we compare generated peptides against 31,466 distinct reference sequences from six databases, including AntiCP 2.0, used to train the held-out anticancer predictor. The new Table 2 reports:

Reference overlapLLMLFDDiffusionNFT
Exact match34%0%0%
≥90% normalized edit similarity56%0%0%
≥80% normalized edit similarity77%0%0%

Direct LLM generation frequently reproduces known sequences or close variants, consistent with memorization. In contrast, LFD, like DiffusionNFT, shows no detectable exact or near-copying in the evaluated samples at the reported thresholds. These results provide encouraging evidence against sequence regurgitation by LFD. This analysis does not rule out pretraining exposure to held-out property information. The revised statement in Sec. 4 concerns the information withheld during the experiments, it does not claim that the evaluated designs or their properties were unseen during pretraining.

While the lack of memorization is encouraging, prospective wet-lab (e.g., in vitro) validation is needed to assess LFD’s utility independently of potential overlap between the teacher’s training data and the held-out evaluator. Our strong computational results motivate this next step.

AuthorsOpenReview ↗

Rebuttal (2/3): Synthetic and Retrosynthetic accesibility and LLM reasoning assessment

Synthetic or retrosynthetic accessibility Several influential recent papers on molecular flow and diffusion models, published at ICML, ICLR, NeurIPS, and AISTATS [1,2,3,4,5], evaluate molecular validity and generation quality without reporting synthetic or retrosynthetic accessibility. Their omission is therefore not unusual in this methodological literature, and these evaluations should not be regarded as a uniformly established standard. Nonetheless, we consider them important for practical molecular design and subsequent wet-lab testing. Following the Reviewer’s suggestion, we conducted new experiments assessing both synthetic and retrosynthetic accessibility, reported in Appendix D.2.2, Tables 5–6, of the updated manuscript.

We first assessed the synthesis-agnostic checkpoints from the main paper using SA score [9] and Retro* score [10], as reported under “Evaluation of original (synthesis-agnostic) LFD” (Appendix D.2.2, Table 5) in the updated manuscript. Original LFD achieves the lowest mean SA (lower is better) and best mean scores on all three docking metrics among the original adapted methods. Its RS (higher is better) exceeds FlowGRPO and DiffusionNFT but remains below DPO.

LFD can be made synthesis-aware simply by including synthetic or retrosynthetic accessibility in the discovery goal. We added two variants under “Evaluation of synthesis-aware LFD” (Table 5): (i) prompt-only, without providing SA/RS observables, and (ii) prompt plus SA/RS observables. Their mean results are:

LFD variantSA ↓RS (%) ↑GNINA ↑Vina ↓rDock ↓
Original, synthesis-agnostic2.44 ± 0.0471.2 ± 3.26.77 ± 0.11−8.84 ± 0.16−12.87 ± 0.57
Synthesis-aware: prompt only2.23 ± 0.0779.3 ± 4.56.76 ± 0.04−8.26 ± 0.20−11.94 ± 0.38
Synthesis-aware: prompt + SA/RS2.11 ± 0.0684.7 ± 5.66.79 ± 0.06−8.23 ± 0.10−12.15 ± 0.19

Both variants improve SA and RS over original LFD, while retaining better mean docking scores than pre-trained FlowMol on all three metrics. The variant with SA/RS observables achieves lower mean SA (2.11) and higher mean RS (84.7%) than pre-trained FlowMol and the DPO, FlowGRPO, and DiffusionNFT baselines trained without a synthesis objective (Table 5).

We additionally ran “Baselines via multi-objective scalarization” (Table 6): DPO, FlowGRPO, and DiffusionNFT with binding-plus-SA or binding-plus-RS rewards at binding:synthesis weights of 50:50, 80:20, and 95:5. Both synthesis-aware LFD variants match or improve upon the best mean docking scores across these 18 configurations, without selecting explicit scalarization weights. LFD with SA/RS observables also achieves the highest mean RS: 84.7%, versus at most 81.0% among scalarized baselines. Some baselines obtain lower SA, but with lower RS and significantly lower docking scores than this LFD variant.


First-principles reasoning and LLM reasoning reliability

  • (First-principles reasoning) We have removed the term "first-principles" from the entire updated manuscript. As one can expect, the contributions presented within the original paper hold unchanged. We thank the Reviewer for catching this linguistically inappropriate wording.

  • (LLM reasoning reliability) The authors team includes chemistry and biology experts that have assessed key parts of the LLM reasoning in multiple goals. This is possible due to LFD allowing interpretable sub-goals and traces. We have also consulted external (computational) chemistry experts in molecular and/or protein design. They have assessed the LLM-generated sub-goals, traces, as well as final generated structures as chemically and biologically reasonable from a scientific standpoint.

AuthorsOpenReview ↗

Rebuttal (3/3): Experiment on a new scientific generator, OOD goals, and references

Arbitrary generator The proposed method (LFD) can be utilized, by construction, with any generative model that can be adapted via DPO. This is arguably the case for most, if not all, current generative models employed in scientific discovery. This includes discrete flow and diffusion models, their continuous counterparts, GFlowNets, and (e.g., local) LLMs. LFD in principle can also be used to distill frontier LLM knowledge even into explicit probability distributions over known design spaces, e.g., a categorical distribution, as often employed for library generation for high-throughput experimental technology [6]. The experimental results within the original submitted version offer experimental assessment on both discrete (discrete flow models) and continuous generative models (flow models), widely used in sequence and structure design, and in scientific discovery tasks in general. Within the updated version, we have added a new goal on peptides (new Goal 3, sec. 5) tackled via a third model class: a discrete diffusion model widely used for peptide design [7]. Together, these experiments demonstrate LFD across discrete flow, continuous flow, and discrete diffusion models, spanning both sequence and structure generation. This provides empirical support across three model classes, while the broader methodological applicability is conditional on a generator supporting the required DPO update.


OOD goals and systematic success rate

  • (OOD discovery) Prior formalizations, including the notion used in ActFlow [8], allow for varying degrees of out-of-distribution rather than a strict binary classification. While the degree of OODness is difficult to estimate for Goals 1–5, Goal 6 (previously Goal 5) provides clearer empirical evidence of a highly OOD target. Based on this observation, we have updated the writing making it more conservative, revising the introduction and conclusion to refer to this single demonstrated OOD discovery goal and updated Insight 6 accordingly.

  • (Goal 5 success rate) For Goal 5 (now Goal 6), two of seven initiated runs produced verified hits. Once these hits were confirmed, we stopped the remaining runs at intermediate stages to limit compute and API costs. Thus, 2/7 (28.6%) is the observed fraction of successful runs under this stopping procedure. Our emphasis on the strongest verified discoveries follows established precedents in FunSearch [11], Eureka [12], and TTT-Discover [13], which search over multiple attempts and evaluate the best solutions found.


We hope these additional experiments are satisfactory evidence to address your questions. If so, we would kindly ask the reviewer to consider increasing their score, or to let us know which concerns remain.

References

[1] SemlaFlow -- Efficient 3D Molecular Generation with Latent Attention and Equivariant Flow Matching, Irwin et., AISTATS 2025

[2] All-atom Diffusion Transformers: Unified generative modelling of molecules and materials, Joshi et al., ICML 2025

[3] Variational Flow Matching for Graph Generation, Eijkelboom et al., NeurIPS 2024

[4] Accelerating 3D Molecule Generation via Jointly Geometric Optimal Transport, Hong et al., ICLR 2025

[5] Controlled Generation with Equivariant Variational Flow Matching, Eijkelboom et al., ICML 2025

[6] Breaking the Synthesis Barrier for AI-Designed DNA Libraries, Sussex et al., NeurIPS 2026

[7] PepTune: De Novo Generation of Therapeutic Peptides with Multi-Objective-Guided Discrete Diffusion, Tang et al., ICML 2025

[8] Active Flow Expansion for Out-of-Distribution Discovery: from Theory to Molecules, De Santi et al., NeurIPS 2026

[9] Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions, Ertl et al., Journal of cheminformatics 2009

[10] Retro*: learning retrosynthetic planning with neural guided A* search, Chen et al., ICML 2020

[11] Mathematical discoveries from program search with large language models. Romera-Paredes et al., Nature 2024.

[12] Eureka: Human-Level Reward Design via Coding Large Language Models. Ma et al., ICLR 2024.

[13] Learning to Discover at Test Time, Yuksekgonul et al., ICML 2026.

Reviewer_cEQpOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper proposes a new method to align domain-specific generative models to discovery goals that are stated in natural language and for which a reward function may be impractical to get. This is a real issue as competing frameworks rely on reward-guided adaptation (eg, FlowGRPO). Rewards may be either unavailable, poorly fit to the task or they may lead to reward hacking.

The authors propose LLM-to-flow Distillation (LFD) where a frozen, off-the-shelf LLM is used as an adaptive teacher that guides the training of the generative model. The authors tackle two issues: first, LLMs appear in all cases to be bad at guessing the absolute scoring of a given candidate in all the cases considered. The second issue is that the first version of the generative model may place little or no mass on the region of interest (ie, not produce any valid candidate). The first issue is solved by using paired comparisons and DPO to train the generative model on these pairs. The second issue is addressed through sub-goals design which are performed iteratively based on the generative model current capabilities. Sub-goals are redefined dynamically throughout training to reach the ultimate goal.

The model is validated in two parts. In section 4, LLM feedback reliability is assessed across three domains (QM9 MolLogP, MatterGen crystal hull energy, PepDFM peptide hydropathy) for absolute scored and paired comparisons. The conclusion of that assessment is that pairwise assessment has an accuracy of about 75%, which is much more encouraging than the absolute score guessing from the models. In section 5, LFD is tested in five natural-language goal conditioned training loops (PepDFM cell-penetrating and anticancer, FlowMol3 in D2 docking, GSK3β docking and a structurally constrained kinase campaign). Compared to DPO, classifier guidance, FlowGRPO, and DiffusionNFT, LFD reports the best held out scores all tasks.

Critical strengths
  1. The problem formulation is new, well formulated and the solution fits its purpose. The Goal 5 result demonstrates it cleanly. Equation 2 is a formulation I had not seen stated this directly, and it is the right one for a large class of real discovery settings. That the adaptive curriculum succeeds where a correct-but-sparse reward fails is the clearest demonstration of the soundness of the solution.

  2. The curriculum and judgments are legible and auditable by experts. This is a meaningful advantage over reward-guided adaptation, and it can even lead to further scientific insights in how to design research roadmaps. It also makes the method's limitations inspectable (see comments below), which is something to be praised.

Critical weaknesses
  1. My first concern is that the held-out evaluators are not independent of the teacher. I think this undercuts the reward-hacking claim. To put it simply: LDF mitigates reward hacking by not overfitting to one classifier, therefore observing reward hacking would be very surprising. Additionally, the LLM has been trained on the model that provided the reward (published characterizations of those datasets are part of the literature the LLM teacher has read) so agreement between the classifier and the LLM is definitional. The main should carry the caveat. Showing that reward hacking exists for the other methods is nevertheless useful, and showing how LDF compares is nice but I wouldn't draw too much on that.

  2. I believe the experiments cannot separate the three things the method does, so the contribution is poorly localized. My understanding is that LFD bundles (a) LLM goal decomposition into an adaptive curriculum, (b) LLM categorical judgment of candidates, (c) DPO as the optimizer. Every experiment varies all three at once against baselines that have none of (a) or (b) and a different (c). The implications are that, first, curriculum and optimizer are confounded: LFD is only ever run with DPO and the baselines are only ever run without a curriculum. "LFD beats FlowGRPO" is consistent with many things at once (the curriculum helping, DPO suiting this regime, or both). Second, component (b) appears to be doing less work than Section 4 implies: in Appendix D nearly every sub-goal is a deterministically checkable predicate ("exclude cysteine", "at least two aromatic rings and an aliphatic nitrogen", the whole of Goal 5). Say one were to code an exact checker, like a python script (which LLMs can perfectly do), it would strictly dominate a two-pass LLM vote.

Questions for authors
  • Checker ablation: For the sub-goals that admit an exact predicate I'd like to see what happens when the LLM proposes the rung as now, but a deterministic checker labels the batch? I believe a win would be genuinely interesting (that would be a slightly different paper, but stronger in my opinion).

  • Curriculum vs optimizer: What happens if we run FlowGRPO or DiffusionNFT on the LFD curriculum: same sub-goal sequence, same per-rung labels, different optimizer?

  • Each round forms the full Cartesian product of stable positives and negatives, which is roughly 400 heavily correlated pairs from 40 samples. Are results stable under subsampled or matched pairings?

  • Appendix D.2.3 lists the Goal 5 curriculum as g0, g1, g3 with no g2, while Section 5 refers to "the third sub-goal g2". Which is correct?

  • I believe having the LLM emit each sub-goal as an executable predicate (SMARTS, sequence rule) rather than prose would give an exact p(Ω_{g_t}) per round. Currently, curriculum progress and judge drift are not separable.

5 comments
AuthorsOpenReview ↗

Rebuttal (1/2): Memorization analysis

We thank the Reviewer for recognizing that our problem formulation is new, well-formulated, and the right framing for a large class of real discovery settings, as well as the experimental results in Goal 6 (previously 5) showcasing the soundness of the proposed solution, and for praising the interpretability of our method. In the following, we aim to sharply address all limitations and questions mentioned within the review.


Evaluator dependence and memorization. We thank the reviewer for this important point. Held-out during LFD does not imply independent of LLM pretraining. We have therefore removed “independent” when describing evaluators, softened the reward-hacking claim, and added an explicit caveat in the new main-text paragraph “LLM training data and memorized designs.” (Sec. 5). Although the LLM utilized by LFD never observes the held-out evalutator and therefore cannot adapt to it the generated designs, frontier-LLM training data are undisclosed, exposure to relevant datasets or predictor descriptions cannot be excluded. We have clarified this caveat in the updated text. Our comparison supports improved held-out performance, rather than guaranteeing immunity to reward hacking. Wet-lab validation of novel designs, which clearly goes beyond the scope of this work, is needed to assess whether the LLM’s chemical judgments generalize to new experimental outcomes beyond agreement with potentially memorized datasets or predictors. As a computational step toward assessing this aspect, we introduced an analysis of memorization for the anticancer task to determine whether LFD regurgitates sequences that the LLM may have seen during pretraining. Specifically, we compare generated peptides against 31,466 distinct reference sequences from six databases, including AntiCP 2.0, which is the dataset with which the held-out predictor we utilized for evaluation has been trained. The new Table 2 (Sec. 5) reports:

Reference overlapLLMLFDDiffusionNFT
Exact match34%0%0%
≥90% normalized edit similarity56%0%0%
≥80% normalized edit similarity77%0%0%

Here, LLM denotes peptide (sequences) generated by directly prompting Gemini 3.1 Pro with the anticancer goal, LFD denotes peptides sampled from the discrete flow (PepDFM) adapted via LFD, and DiffusionNFT denotes the reward-guided baseline without LLM supervision. Direct LLM generation frequently reproduces known sequences or close variants, thus showing significant memorization. In contrast, LFD shows no detectable exact or near-copying in the evaluated samples, as DiffusionNFT. These results suggest that LFD does not regurgitate sequences memorized by the teacher due to task-related pre-training information.

AuthorsOpenReview ↗

Rebuttal (2/2): Additional requested ablations

Localizing the roles of curriculum, judgment, and optimization. We thank the reviewer for these constructive suggestions. We agree that end-to-end comparisons alone cannot identify each component’s contribution, and we have spent significant effort to add targeted ablations, discussed in the following.

  • (Checker ablation with LLM-generated executable predicates) We thank the Reviewer for mentioning this interesting idea. We modified LFD so that the LLM can generate deterministic program checkers for well-defined sub-goals whose satisfaction can be assessed exactly from the available observables, as reported in Appendix E.4, Tables 9 and 10 of the updated manuscript, under the new subsection “Ablation on replacing LLM judgments with executable checkers”. We tested the first two sub-goals of Goal 5 (previous Goal 4). The LLM generated Python/RDKit checkers and for each sub-goal, the checker matched the saved LLM judgments on all 128 molecules, showing 100% agreement across 512 label comparisons. Thus, replacing LLM judgments leaves the DPO preference sets exactly unchanged for these sub-goals and samples. Nonetheless, in general verified checkers would also prevent possible errors from LLM judgments. Moreover, each checker can be reused across batches under the same sub-goal, replacing repeated LLM judgment queries with local execution, thus reducing API costs.

  • (Ablation on LFD optimizer) We replaced the DPO update with iterative DiffusionNFT using LLM-generated scores from 1 to 100 for each sub-goal, as reported in Appendix E.2, Table 7 of the updated manuscript, under the new subsection “Ablation on LFD optimizer update”. LFD with DiffusionNFT obtains an AntiCP2-alt score of 52.94 ± 2.05%, compared with 53.54 ± 3.32% for DPO; the corresponding HemoPI2 scores are 44.08 ± 1.21% and 42.49 ± 2.04%. These are mean predictor scores with 95% confidence intervals. LFD therefore retains similar predicted anticancer activity with DiffusionNFT, while DPO achieves lower mean predicted hemolysis. This demonstrates compatibility with an alternative optimizer, although the feedback format also changes. Nonetheless, as presented in Sec. 4, such positive results via score-based feedback might not hold in general, depending on reliability of the LLM employed for feedback.

  • (Ablation on directly optimizing the final goal without curriculum) We added a baseline using LLM categorical feedback directly with respect to the final goal, without intermediate sub-goals, as reported in Appendix E.7, Table 11 of the updated manuscript, under the new subsection “Ablation on directly optimizing the final goal without a curriculum”. This baseline clearly performs significantly worse than LFD on both tasks: its mean CPP-predictor score is 14.21% versus 69.06% for LFD, and its mean GNINA score is 6.17 versus 6.77; LFD also achieves better mean Vina and rDock scores. Moreover, final-goal-only LFD cannot form preference pairs when no positive samples are available. Goal 6 (previous Goal 5) clearly demonstrates this limitation: the initial FlowMol3 generated no final-goal positives in more than 38,000 samples. Intermediate sub-goals provide the supervision needed to begin adaptation in this setting.


Robustness to subsampled and matched preference pairs. We added the subsection “Ablation on subsampling preference pairs” in Appendix E.5, with results in Figure 12 of the updated manuscript. We evaluate two Goal 2 sub-goals, fixing the LLM labels on 36 positive and 36 negative samples, the model before the update, and the training budget. We compare (i)(i) the full Cartesian product of 1,296 pairs; (ii)(ii) random subsets selected uniformly without replacement; and (iii)(iii) balanced subsets formed by combining one-to-one matchings without repeating a pair. For (ii)(ii) and (iii)(iii), we use subsets of 36, 72, 144, or 324 pairs. As shown in Figure 12, random and balanced subsets achieve CPP-predictor scores comparable to those obtained with the full Cartesian product. Both subsampling strategies achieve similar results with fewer distinct preference pairs in these two single-update experiments. Pair subsampling therefore offers a practical refinement of LFD, enabling adaptation with potentially smaller preference sets while retaining comparable performance. We thank the Reviewer for mentioning this suggestion.


Appendix D.2.3 Goal 5 Curriculum We thank the Reviewer for finding this inconsistency. We updated the appendix in the submitted manuscript.


Thanks again for your thoughtful review, and let us know if we solved all your doubts! If so, we would kindly ask the reviewer to consider raising their score to show support for the paper.

Dear Reviewer,

Thank you for your constructive feedback which has strengthened our paper. Since you did not note any remaining question on concern, we wanted to check whether anything is still left unclear on your side. If there is any specific clarification that would move the paper toward a clear accept, please let us know, and we would be happy to provide it.

Thank you again for taking the time to review our paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We thank the Reviewers for recognizing our problem formulation as well-motivated (8MVf), new, well-formulated (cEQp), and the right one for a large class of real discovery settings (cEQp), our method as simple and direct (UF1C), the legibility and auditability of its curriculum and judgments as a meaningful advantage over reward-guided adaptation (cEQp), and the results on Goal 5 (now Goal 6) being the clearest demonstration of the soundness of the solution (cEQp) and strong evidence in favor of LFD (8MVf).

To tackle all points raised by the Reviewers, we used the additional main-text page and expanded the appendix by 12 pages, adding a sixth discovery goal (new Goal 3), synthetic and retrosynthetic accessibility evaluations, experiments with another frontier LLM and with another scientific generator, and six targeted ablations of the method’s components. The revisions are in blue text within the updated manuscript. We summarize the main contributions and additions below.

Molecules synthetic and retrosynthetic accessibility

We added synthetic and retrosynthetic accessibility evaluations (Appendix D.2.2, Tables 5–6): (1) original LFD achieves the lowest mean SA and best mean docking scores among synthesis-agnostic adapted methods, (2) adding synthesis requirements to the prompt improves both SA and Retro*, and (3) providing SA/Retro* as observables further improves both. Against baselines using weighted binding and synthesis rewards, both synthesis-aware variants match or exceed the best mean docking scores, and the variant with observables also achieves the highest mean Retro*.

Arbitrary generator claim and new discovery goal with another scientific generator By construction, LFD can be applied to any generative model that supports the required DPO update. The original experiments assess LFD using both discrete and continuous flow models. In the updated manuscript, we added a new peptide discovery task using PepTune, a masked discrete diffusion model over peptide SMILES (Goal 3, Sec. 5.1, Fig. 6; Appendix D.1.3, Table 4). Our experiments thus now cover three generator classes, spanning sequence and molecular-structure generation, providing broader empirical support for LFD’s methodological applicability.

Extra discovery results with a different frontier model

We added “Ablation on LLM model family” (Appendix E.3, Table 8), replacing both Gemini models with GPT-6.1 Sol on peptide and small-molecule tasks. Mean predicted CPP activity increases from 69.06% to 74.05%, while predicted hemolysis decreases from 37.02% to 33.73%, while docking scores remain comparable. These results show that effective LFD adaptation is not specific to the original LLM family.

LLM training data and memorization

We added an explicit pretraining caveat and a memorization analysis in the new paragraph “LLM training data and memorized designs.” (Sec. 5, Table 2). Against 31,466 reference peptides from six databases, direct LLM generation produces 34% exact matches and 56%/77% overlap at the 90%/80% similarity thresholds, LFD produces 0% at all three thresholds. These results provide evidence against exact or near-copying in the evaluated LFD samples.

Method Ablations

We added targeted experiments to ablate several aspects of LFD, including the roles of the curriculum, feedback, optimizer, and training pairs, as asked by the Reviewers:

  • Executable checkers (E.4, Tables 9–10): generated programs show 100% agreement with both LLM passes for two molecular sub-goal predicates and four exact peptide clauses, allowing repeated judgments of these predicates to be replaced by local execution.
  • Optimizer (E.2, Table 7): replacing DPO with DiffusionNFT retains similar predicted anticancer activity (52.94% versus 53.54%), although the feedback format also changes.
  • Curriculum (E.7, Table 11): final-goal-only LFD underperforms full LFD on CPP activity (14.21% versus 69.06%) and GSK3β docking (GNINA 6.17 versus 6.77).
  • Observables (E.1, Fig. 11): removing the developability and non-hemolysis bands retains improvements in predicted anticancer activity with comparable predicted hemolysis.
  • Preference pairs (E.5, Fig. 12): random and balanced subsets of 36–324 pairs achieve comparable results to all 1,296 pairs in two controlled single-update experiments.
  • Class imbalance (E.6, Fig. 13): single-update results show limited sensitivity to positive–negative class sizes, including 9:1 and 1:9 ratios.

Ultimately, we wish to point out that we may have underemphasized LFD’s central novelty: to our knowledge, it is the first method to fine-tune scientific generators such as flow and diffusion model toward natural-language discovery goals, without requiring a task-specific proxy reward, and by distilling goal-related information from frontier (closed-weights) models.

We hope these new results and the detailed responses below address the Reviewers’ concerns and make the contribution, evidence, and scope of LFD clearer.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.