acceptodds
Under review as a conference paper at ICLR 2027

MOLRGEN: A Benchmark and Verifier for Reasoning LLMs on De Novo Molecular Generation

Abstract

Recent large language models have shown strong performance on tasks with verifiable outcomes, but their use in de novo molecular generation remains limited by the lack of training environments where rewards can be computed without reference molecules. We introduce MolRGen, a benchmark and molecular verifier for post-training and evaluating LLMs on de novo molecular generation. MolRGen contains approximately 4,500 protein-pocket targets, resulting in 50k multi-objective optimization prompts combining docking scores with molecular properties such as QED, synthetic accessibility, logP, and physicochemical descriptors. Unlike caption-based generation or molecule-editing benchmarks, MolRGen evaluates molecules proposed from scratch by computing rewards at generation time. We benchmark general-purpose and chemistry-specialized open-source LLMs and introduce a diversity-aware top-k metric to measure whether models can generate a diverse set of high-scoring molecules. Finally, we use the verifier to fine-tune a 128B LLM with GRPO, showing improved performance, at the cost of a diversity-exploitation trade-off.

Primary area:
datasets and benchmarks
Keywords:
Large Language Models, Reasonning, Verifiers, Benchmark, Molecular Modeling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_hsxSOpenReview ↗
Rating
3(+1)
Confidence
2(−2)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MolRGen, a benchmark aiming at text-based DE NOVO molecule generation. The benchmark consists of 4500 structures and 12 classical molecule properties. Based on the proposed dataset, the author trains a specific LLM, RL-Molstral, which gains better performance than baselines. The results support that GRPO training can benefit reasoning capabilities of the LLM on these DE NOVO tasks.

Critical strengths
  1. The writing structure of the paper is clear. The authors follow a "new DE NOVO problem --> new dataset proposed --> train a specific model on it --> evaluation from many aspects" pipeline. This makes the paper easier to follow.

  2. The scale of the proposed dataset is large. This may reduce the problem of randomness.

  3. The main evaluation consists of both general LLMs and chemistry-specific LLMs. This is necessary in evaluations of text-based molecule generation.

Critical weaknesses
  1. In fact, the current story makes the novelty of this paper limited. The authors claim that they want to explore performing "open-ended molecular design", and this has been explored in [1][2]. The "molecule verifier", like RDKit calculations or similarity calculations, is also not a new technique. [3][4] also utilizes a GRPO-style training process to teach general LLMs how to learn knowledge of molecule generation. [5] has explored the topic of RL helping open-ended generation. Indeed, verifiers help LLMs learn, but this is not new, compared with previous work. Regarding the works mentioned above, I think that "novel, new" observations from the proposed benchmark seems to be weak. While I appreciate the artifacts the authors propose, the authors may point out the key differences between this work and previous benchmarks & RL models.

  2. Additionally, some important ablations and baselines are missing. Chem-R and MolReasoner I mentioned, and some other baselines should be compared. Additionally, you may add supervised fine-tuning as an ablation. The evaluation does not seem strong enough without these comparisons. You may have some important results in the supplementary material, however, following ICLR 2027 instructions, you may either put them in your main text, or use an "Appendix" just after the references to show these results.

Here is another small suggestion: you may redesign Table 1, as the current version is not clear enough.

[1] Li, Jiatong, et al. "Speak-to-structure: Evaluating llms in open-domain natural language-driven molecule generation." Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 2026.
[2] Hao, Li, et al. "Beyond chemical qa: Evaluating llm's chemical reasoning with modular chemical operations." Advances in Neural Information Processing Systems 38 (2026).
[3] Wang, Weida, et al. "Chem-r: Learning to reason as a chemist." Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 2026.
[4] Zhao, Guojiang, et al. "Molreasoner: Toward effective and interpretable reasoning for molecular llms." arXiv preprint arXiv:2508.02066 (2025).
[5] Narayanan, Siddharth, et al. "Training a scientific reasoning model for chemistry." Advances in Neural Information Processing Systems 38 (2026): 157671-157710.

Questions for authors

See the weaknesses part. I think the core problem is the limited novelty compared to previous works (open-ended design, GRPO training for LLMs on molecule generation). So I recommend rejection. I have some other questions:

Do you use the same prompt template in your training and evaluation? If so, will it influence the results? If not, can you show the different prompt templates?

Have you ever tried applying your trained model on some other open-source text-based molecule generation benchmarks? If so, you can prove your concept that your verifier "really helps".

4 comments

1. Novelty and positioning ([1]–[5]). We agree that none of the individual components of MolRGen is new in isolation — RDKit descriptors, GRPO, and verifier-based rewards all have precedents. Our claim is not any single component but the composition: a reference-free RL environment where rewards for de novo molecules are computed at generation time, and the findings it enables. Concretely, we see MolRGen as complementary to the cited works:

WorkTask typeGround truth / rewardDe novo?
Speak-to-StructureNL-driven generationOne-to-one molecule–text mappings; pre-defined answer moleculeNo (retrieval of a defined target)
ChemCoTBenchChemical reasoning / modular operationsAnnotated QA answersNo
Chem-RMulti-task chemistry QA / reasoning via GRPOVerifier on answers to known questionsNo
MolReasonerMolecular reasoning incl. text-to-moleculeMatches a described target molecule; test-time RLNo (caption-based)
Science0-cChemistry scientific reasoningQA-style verifiersNo
MolRGenPocket-conditioned de novo generationOnline-computed docking + descriptor rewards; no reference moleculeYes

The distinction we consider essential: in [1]–[5], the verifier checks agreement with a known answer (a described molecule, a QA label, a reaction outcome). MolRGen's verifier instead scores novel molecules against a protein pocket and a property profile at generation time, which is what makes it usable as an RL environment rather than an evaluation set. We also note that [1]–[5] contain no structure-based objective: no cited work conditions on a protein pocket or runs docking inside the training loop. To the best of our knowledge, the combination of (i) reference-free pocket-conditioned prompts, (ii) docking-in-the-loop rewards for GRPO, and (iii) a diversity-aware evaluation exposing the diversity–exploitation trade-off of RL fine-tuning is new as a whole.

2. Missing baselines and ablations.

  • Chem-R and MolReasoner: We are running the evaluations of MolReasoner on MolRGen prompts, and we are coming back to you whenever the results are available. For Mol-R, no model weights have been released on HuggingFace, so we will be unable to run a comparison within the reviewing period. We note that both methods target different tasks than MolRGen — multi-task chemistry QA and reasoning [3], and molecular understanding with description-to-molecule generation [4] — neither conditions on a protein pocket, and their training signals are verifier scores against known answers rather than rewards computed for novel molecules at generation time.
  • SFT ablation: We agree this comparison would be interesting, and we considered it carefully (tools to build SFT datasets are provided in our code). First and mainly, the topic of our work is on the creation of an environment to evaluate the skills of language models in de novo drug discovery, and show that this environment can be used for RL post-training. Additionally, training a large 128B model forced a choice between RL post-training experiments and an SFT experiment within our compute budget.

(All supplementary materials will be concatenated to the main paper for the final submission)

  1. Prompt templates in training and evaluation. We support the use of any prompts, and provide tools to generate custom textual prompts given a task’s metadata, but we used the same prompt templates across training and evaluations up to simple adaptations for multi-property optimization.

  2. Transfer to other benchmarks. We have not evaluated the trained models to other benchmarks, however regarding the generalization abilities of the model, we did evaluate it, since all proteins in the test set are unknown to the model (none leak from the training to the validation and test set).

We hope these answers clarify the contributions of MolRGen relative to prior work, and we remain at the reviewer's disposal for any further discussion.

Reviewer_hsxSOpenReview ↗

Response to Rebuttal

Thank you for your response!

I will wait for the additional results as you mentioned.

Best,

Reviewer hsxS

Dear reviewer hsxS,

We just obtained the results of the evaluation of the MolReasoner checkpoint (post-trained on reference-based molecular generation) on our tasks. The performance is lower than of ChemDFM-R, mostly due to the model struggling with instruction following (the model mistakes SMILES and SELFIES notations for instance, or does not generate the compounds in parseable form).

This could be due to its size (only 7B), making it less robust when evaluated on slightly out-of-distribution tasks. You can find below the performances of MolReasoner:

ModelSizenr=16n_r = 16nr=16n_r = 16nr=24n_r = 24nr=24n_r = 24nr=32n_r = 32nr=32n_r = 32nr=48n_r = 48nr=48n_r = 48
top-1top-16top-1top-16top-1top-16top-1top-16
Gemma-431B0.650.370.670.480.680.520.70.56
Mistral-Small-4128B0.60.320.620.440.640.480.660.52
MiniMax-M2229B0.610.310.640.430.660.480.680.42
RL-MolStral-g16128B0.640.450.660.530.670.550.690.57
ChemDFM-R14B0.620.310.650.410.660.460.690.5
Ether024B0.550.220.580.310.590.360.620.42
MolReasoner7B0.390.080.420.120.450.150.480.22

We will include this new baseline with a discussion of its behavior (similarly to the Ether0 section in the appendix) in the revised version of the paper.

Reviewer_hsxSOpenReview ↗

Response to Authors' Rebuttal

Thank you for your extra experiments. I would like to raise my score to 3; however, the novel contributions of this paper are still not so convincing, and I will lower my confidence. I could not strongly support acceptance of this paper.

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

The work introduces  a benchmark and verifier for evaluating  for  a de novo molecular generation task using Large Language Models (LLMs). The proposed benchmark and verifier dataset, MOLRGEN, consists of  4,500 protein-pocket targets and 50,000 prompts.  The work introduces  molecule verifiers  based on docking scores and molecular properties, and  demonstrates the use of reinforcement learning (RL) fine-tuning.  MOLRGEN furthermore allows  RL training without pre-computed reference molecules and claims to better capture the complexity of chemical space.

Critical strengths
  1. The benchmark includes the molecule verifiers for computing reward function and it is an important element for training and inference.
  2. Diversity-aware top-k scoring is a useful evaluation metric.
Critical weaknesses
  1. Practical Molecular Optimization (PMO), PDBbind, BindingDB, PINDER, etc. with pockets and inhibitors are available. It is not clear regarding the choice of SAIR for creating datasets.
  2. The benchmark does not include any discussion on diversity of protein targets, and the potential biases in those targets. This aspect is important for drug discovery targets.
  3. Although the molecular verifiers are a good addition in the benchmark, most of them are weak verifiers in the drug discovery domain without strong experimental or molecular dynamic simulations. For example, better docking scores do not indicate a better chemical space. No discussion on Stereoselectivity or other cell-specific properties in the verifiers indicate limited applicability of the benchmark and verifiers for drug discovery tasks.
  4. The evaluations on the state-of-the-art commercial models are not performed and this limits the evaluation and benchmark's scope.
  5. The use of Tanimoto similarity is not a good measure for diversity-aware top-k scoring.
  6. The dataset includes upto three properties plus docking score which limits again the scope for real-life applications to drug discovery.
Questions for authors
  1. What are possible biases that can be introduced by the current dataset during training?
  2. How is the diversity-aware scoring based on cosine similarity?
  3. Minor changes in molecules (having high Tanimoto score) often have different properties, and hence, diversity-aware based on the Tanimoto score may not be an ideal way to score differences between molecules. How will your evaluations perform on such molecules?
  4. How do your benchmarks and rewards ensure better chemical space exploration?
  5. What are the reasons for domain-specific LLMs poor performance on the benchmarks? Furthermore, a better performance of the Gemma-4 can be explained by the multimodal?
2 comments

We thank reviewer DtKD for the constructive comments. We address each pointed weakness and question below.

1. Dataset choice (SAIR vs. PMO / PDBbind / BindingDB / PINDER). Our central design constraint is a reference-free setting: we require pocket targets for which no co-crystallized ligand is used as input or conditioning signal, so that rewards are computed online rather than against known binders. We agree these alternative resources are important and will add a discussion comparing them although we want to point out that PINDER covers protein–protein interaction interfaces rather than small-molecule pockets, and PMO defines ligand/scaffold optimization tasks inherited from prior benchmarks rather than pocket-conditioned generation.

2. Proteins target diversity and biases. We include in the first section of the supplementary material a discussion on the target proteins (function, Uniprot annotation score, …). We will enrich this section with a broader discussion on the provenance of the data and the related biases (some structures are generated with cofolding models, which might introduce a bias at training time).

3. Verifier strength (docking and 2D descriptors as proxies). We agree that docking scores and physicochemical descriptors are weak proxies for drug-discovery viability as discussed in our limitations section, however we chose them for 2 main reasons: despite their limitations, they remain a common evaluation tool for molecular generation algorithms, but most importantly, they are “fast” to run. Running molecular dynamics simulations at scale for hundreds of systems in parallel, in a GRPO post-training run for an LLM is too computationally heavy, considering a simulation has to be done for each generated rollout in the RL run.

4. Missing commercial models. We acknowledge this limitation. Not evaluating commercial models was chosen to: (1) ensure the results are fully reproducible, and mainly (2) for the cost of the experiments as evaluating a single commercial model comprehensively would be too expensive. For this reason we focused our analysis on open-source LLMs.

5. Tanimoto similarity and the diversity-aware top-k. We do not use the Tanimoto similarity only on ecfp fingerprints, but also investigate other descriptors, such as pharmacophoric fingerprints and avalon fingerprints. Depending on the nature of the fingerprints, a small structural change might not result in high Tanimoto Similarities.

We ran experiments to evaluate the results using ChemBert as an encoder and the cosine similarity for the similarity measure, on a set of open-source models. Overall the results are not very informative: the same dynamics can be observed, but the curves are less distinguishable (figure: [https://anonymous.4open.science/r/iclr-figures-876D/chembert-mtr-main.pdf ]).

Finally, we find using fingerprints (of different kinds, e.g. ecfp, maccs, pharmacophoric, …) has the advantage of sharing the results using tools and similarity kernels that are widely used in the chem-informatic field. We would sincerely appreciate if reviewer DtKD could share an alternative way to measure the diversity, as it could be a great opportunity to enrich our work.

  1. Scope of objectives (3 properties + docking). We agree investigating the optimization of more objectives is an interesting next step, but building diverse multi-property optimization tasks with more than 3 objectives. In particular, this would mean optimizing other types of objectives, which we are investigating for future work (synthesis planning, ADME properties, …). For reference, the PMO dataset for instance focuses mostly on single objectives, and the multi-property optimisation objectives it includes contains at most 4 objectives, with no docking scores.

Other questions:

  1. Ensuring chemical diversity in the training We do not enforce diversity constraints during training, we evaluate it a-posteriori. The evaluation of group-level rewards to ensure diversity in the generated compounds could be investigated, however such methods often prove to be unstable in practice. While this would be a very interesting future direction, we believe it is out of the scope of this paper.

  2. Performances of specialized models Relative to their size, the performance of ChemDFM-R is not particularly low. However the performances of Ether0 are indeed underwhelming, which we attribute to a strong alignment procedure (cf. Supplementary materials F.5)

We thank the reviewer again for their careful reading, and we will incorporate the points raised above — notably the discussion of dataset biases and the comparison of diversity measures — into the revised manuscript.

I would like to thank you for your response. However, the following critical concerns on MOLRGEN are still outstanding.

  1. I would like to point out that SAIR  is constructed from ChEMBL and BindingDB bioactivity records folded using the Boltz-1 model.  I am not convinced of preferring synthetic co-folded pockets  to the experimental  pockets because  any experimental co-crystal structure from PDBbind can be used simply by removing  the native ligand and docking generated molecules. I would like to correct the typo PLINDER (not PINDER) which is protein-ligand https://www.plinder.sh/.

  2. Since the manuscript proposes the benchmark dataset, the diversity analysis and potential artifacts introduced by the Boltz-1 model  are the central quantitative metrics. For example,    ChEMBL/BindingDB overrepresent  the kinases and GPCRs which will also trickle down to the MOLRGEN benchmarks. Hence, these aspects should be discussed from both qualitative and quantitative manners to ensure a solid Benchmark.

  3.  My question is about how the evaluation metric handles structurally similar molecules that have divergent biological/chemical properties, not Tanimoto similarity.  

  4. Good benchmark should have MD simulation as a verifier. However, other quick verifiers such PoseBusters physical validity checks, PAINS/structural alerts, and synthetic accessibility (SA) scores can also be used without any additional computational burden.

  5. The following publications may be useful to the authors:

[1] Anita Rácz et al, "Life beyond the Tanimoto coefficient: similarity measures for interaction fingerprints." Journal of Cheminformatics, 2018
[2] Schottlender, Gustavo, et al. "Beyond Tanimoto: a learned bioactivity similarity index enhances ligand discovery." Frontiers in Bioinformatics, 2025.

Reviewer_w1unOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper provides a benchmark and online computational reward framework for de novo molecular generation by LLMs. It has about 4,500 protein-pocket targets with molecular descriptors as objectives. Generated SMILES strings are evaluated using docking simulations and RDKit-computed descriptors, with multiple rewards geometrically averaged. The diversity-aware top-k metric selects high-reward molecules subject to similarity constraints.

The paper benchmarks models and fine-tunes Mistral-Small-4 with GRPO, showing that fine-tuning improves scores and decreases diversity.

Critical strengths

It is useful to have a shared, reproducible de novo molecule generation benchmark with multiple objectives consolidated into a single reward.

The diversity-aware evaluation captures the common problem of diversity loss during RL fine-tuning. The experiments demonstrate the usefulness of this metric.

Critical weaknesses

The paper identifies its target collection, verifier, and LLM evaluation and training experiments as contributions. However, I am not convinced that the benchmark's value over existing generation and scoring frameworks is sufficiently demonstrated.

REINVENT 4, mentioned by the paper, supports both unconstrained generation and molecule/scaffold-conditioned modes. MolScore (https://link.springer.com/article/10.1186/s13321-024-00861-w) supports optional and configurable reference-similarity objectives. These show that neither omitting an input molecule nor computing rewards for new molecules is independently novel.

Standardizing targets, property constraints, scoring, and evaluation splits is useful independently of the model. Natural-language prompts provide a convenient LLM interface, but their scientific value should be distinguished from that of the underlying task specification. Fixing prompt wording can support a controlled comparison of LLMs, while alternative prompts or input encodings may yield better molecular-design performance. The LLM-specific formulation alone does not establish added value over model-neutral generation and scoring frameworks. I think it is more important that the framework specify the input information rather than fixing the prompt for a class of models.

The contribution of the expanded target collection and evaluation protocol over existing frameworks needs stronger support.

Questions for authors

Which aspects of MolRGen require an LLM-specific interface, and which define a model-neutral molecular-design task? Can alternative models or prompting strategies use the same target and property specifications and be evaluated by the same verifier? Please explain the scientific value added by the natural-language formulation beyond providing a convenient interface, and distinguish the underlying task constraints from the chosen prompt wording.

Please also clarify the design choice and the consequences of excluding reference molecules. A reference can add constraints when the output must preserve a structure or satisfy a similarity threshold, allowing evaluation of generation under those constraints. How does the chosen reference-free setting affect the capabilities being tested?

2 comments

We thank reviewer w1un for their thorough review of our work. We would like to clarify the position of our paper, and we hope and will address the concerns raised by the reviewer.

Clarification of the paper's purpose. We believe the objective of our work could be made clearer, and we will improve it in the revised paper. The goal of this paper is to evaluate the knowledge and skills of LLMs in de novo drug discovery, not to compare them directly with dedicated molecular generation methods. As discussed in our comparison with REINVENT, an eye-to-eye comparison with such methods is challenging, and these algorithms already perform very well for molecular generation. Our aim is to probe the skill and knowledge of LLMs in this domain.

MolScore. MolScore is indeed an important reference that we missed, and we will add it to the revised manuscript. We note, however, that it does not directly support the natural-language processing and asynchronous handling of our scorer.

Model-neutral vs. LLM-specific formulation. Our verifier supports any prompt or setting for the LLMs. In addition, the code we provide includes tools that automatically generate task prompts from a given molecular generation task's data, so the same target and property specifications can be used across alternative models and prompting strategies and evaluated by the same verifier.

Excluding reference molecules. We deliberately excluded reference molecules because this is precisely the gap we identified in the evaluation of LLMs for molecular generation. Providing a reference can help guide the process and make the task substantially easier, but it also limits exploration of the chemical space — and chemical-space exploration is exactly the ability we want to measure.

We thank the reviewer again for their constructive feedback, which will help us make the paper's purpose and contributions clearer in the revised manuscript.

Thank you for your response. I appreciate your clarification that the same task specifications and verifier support alternative models and prompting strategies.

My remaining concern is the significance of the contribution. Existing resources already support reference-free generation, computational scoring, and molecular-design evaluation. While the expanded target collection and integration for LLM evaluation and post-training can be useful, I am not certain that the added value is sufficient.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.