The same molecule can be expressed in multiple, syntactically distinct representations, yet it remains unclear whether large language models (LLMs) can recognize and relate these equivalent encodings. We introduce MolReps, a pipeline that generates 9,600 question-answer pairs across ten representations for conversion, normalization, and 2D structure recognition. Post-training on MolReps improves text-only and multimodal models, MolQwen3-4B-Instruct (MQ3-4B-I) and MolQwen3-VL-4B-Instruct (MQ3-VL-4B-I), from 4.01% to 75.13% and from 7.11% to 73.14% (InChIKey match) on the molecule-overlap test set, respectively. On the molecule-disjoint test set, post-training increased validity from 20.35% to 58.12% for MQ3-4B-I and from 17.19% to 59.34% for MQ3-VL-4B-I. However, InChIKey match remained low, decreasing from 3.86% to 3.09% and from 6.44% to 2.29%, respectively. Furthermore, the resulting models achieve a relative 27.84% improvement in average over their base model on MolRepBench, a comprehensive benchmark for evaluating molecular property prediction. These results show the effectiveness of targeted post-training for enhancing molecular representation understanding and its potential to improve downstream molecular tasks.
Figures & tables
Figure 1: Overview of MolReps. (a) Ten representations of one molecule. (b) Six task templates, scored by InChIKey match under molecule-overlap and molecule-disjoint protocols. (c) LoRA post-training of text-only and multimodal Qwen3 models. (d) InChIKey match rises from 4.01% to 75.13% on molecule-overlap. (e) 27.84% average gain on MolRepBench.
Figure 2: Token-count distributions in MolReps under the Qwen3 tokenizer. The top panel shows the total token count per QA instance; the bottom panel counts the number of tokens for input representations.
Valid. (%)
IKM (%)
TS
Conversion Task (text)
MO
LLM
Q3-4B-I
20.15 ±1.31
4.01 ±0.67
0.06 ±0.01
MQ3-4B-I
90.69 ±1.00
75.13 ±1.50
0.81 ±0.01
MLLM
Q3-VL-4B-I
16.91 ±1.30
7.11 ±0.84
0.09 ±0.01
MQ3-VL-4B-I OCSR
25.50 ±1.51
5.01 ±0.78
0.09 ±0.01
MQ3-VL-4B-I Conv.
89.39 ±1.04
70.84 ±1.55
0.78 ±0.01
Table 1: Overall MolReps results for base models and training recipes. MO denotes molecule-overlap evaluation and MD denotes molecule-disjoint evaluation. Valid.: validity; IKM: InChIKey Match; TS: Tanimoto similarity.
Valid. (%)
IKM (%)
TS
Base
Q3-4B-I
Rep. Trans.
15.11 ±2.26
0.00 ±0.00
0.01 ±0.00
IUPAC Und.
27.34 ±5.23
0.00 ±0.00
0.02 ±0.01
CML Und.
11.86 ±3.66
0.00 ±0.00
0.00 ±0.00
Intra Norm.
49.12 ±4.71
30.79 ±4.46
0.39 ±0.04
Cross Norm.
15.21 ±1.84
0.00 ±0.00
0.01 ±0.00
Q3-VL-4B-I
Rep. Trans.
5.59 ±1.46
0.40 ±0.40
0.01 ±0.00
Table 2: MO conversion performance by task type for base models and the selected MolQwen3 text-only and multimodal models. InChIKey Match (IKM) is the primary metric.
IKM (%)
Input\Output Rep.
SMILES
SELFIES
DeepSMILES
InChI
Base
SMILES
–
0.00 ±0.00
0.00 ±0.00
0.00 ±0.00
SELFIES
0.00 ±0.00
–
0.00 ±0.00
0.00 ±0.00
DeepSMILES
2.22 ±2.21
0.00 ±0.00
–
0.00 ±0.00
InChI
0.00 ±0.00
0.00 ±0.00
0.00 ±0.00
–
R. SMILES
30.68 ±5.54
0.00 ±0.00
0.00 ±0.00
0.00 ±0.00
Table 3: MO InChIKey Match (IKM) by input and output representation for base models and the selected MolQwen3 text-only and multimodal models. We report the average result across all instances.
k
pass@k (%)
Δ pass@k (%)
valid@k (%)
AC@k (%)
AW@k (%)
1
67.89
67.89
87.09
67.89
32.11
2
76.11
8.23
87.89
60.46
23.89
4
80.11
4.00
87.89
54.06
19.89
8
84.34
4.23
88.21
46.17
15.66
16
86.74
2.40
88.29
39.09
13.26
Table 4: Rollout performance of MQ3-VL-4B-I on the MO test set. pass@k, valid@k, all-correct@k (AC@k), and all-wrong@k (AW@k) are reported for k=1 to 16 generated samples per input.
Figure 3: MolRepBench results for base models and MolQwen3 family. Results are grouped by the Comprehension, Discrimination, and Generation main task categories (top) and breakdown by each benchmark with which main task they belong to (bottom).
Task
Source
Metrics
Comprehension
Atom Counting (B1)
ChEBI-20
Exact match
Group identification (B2)
ChEBI-20
Macro-F1
Property estimation (B3)
ChEBI-20
ρ , Exact match
Molecule Retrieval (B4)
ChEBI-20
Top-1
Discrimination
Table 5: MolRepBench tasks used for external evaluation of molecular understanding from text representations. The benchmark comprises eight tasks in the Comprehension, Discrimination, and Generation categories.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Representation
Source
Tool
1D: linear notations
Canonical SMILES
–
rdkit
Randomized SMILES
canonical SMILES
rdkit
SELFIES
canonical SMILES
selfies
Randomized SELFIES
rand. SMILES
selfies
DeepSMILES
canonical SMILES
deepsmiles
Appendix
Table 6: The ten representations collected for each molecule, the source each is generated from, and the generation tool. “Graph” denotes the RDKit molecular graph.
Task
Input → Output
#/mol
Cross-representation translation
Std. translation
r→r′ ; r,r′∈Rstd , r′=r
12
IUPAC understanding
IUPAC name →Rstd
4
CML understanding
CML →Rstd
4
Normalization
Intra-rep. norm.
Rrand→ same-family std.
3N
Appendix
Table 7: The six MolReps task templates. Rstd = {canonical SMILES, SELFIES, DeepSMILES, InChI}; Rrand = {randomized SMILES, SELFIES, DeepSMILES}; N is the number of randomized samples per molecule ( N=2 in MolReps).
Split
# Src. Mol.
# QA Pairs
QA-level split (molecule-overlap)
Training set
≈ 200 †
7680
Validation set
≈ 200 †
960
Test set
≈ 200 †
960
Subtotal
200
9600
Molecule-level hold-out (molecule-disjoint)
Appendix
Table 8: MolReps dataset splits. † The MO splits are disjoint at the QA-pair level but share approximately the same 200 source molecules; the MD held-out test set contains 20 molecules unseen during training.
Figure 4: Distribution of task types over all 9,600 MolReps QA pairs. Rep. Trans.: standardized representation translation; IUPAC Und.: IUPAC name understanding; CML Und.: CML understanding; Intra Norm.: intra-representation normalization; Cross Norm.: cross-representation normalization; OCSR: 2D structure recognition. See Section 3.2 for details.
Hyperparameter
SFT
GRPO
Learning rate
1e-4
1e-5
Batch size
16
32
Maximum epochs
10
10
Optimizer
AdamW
AdamW
KL coefficient β
–
0
Group size G
–
8
Appendix
Table 9: Training hyperparameters for SFT and GRPO experiments.
Figure 5: SFT training curves under different recipes on molecule OCSR task (left) and textual conversion task (right). Training and validation loss are tracked over training steps.
Figure 6: Targeted-SFT training curves of loss (left) and mean token accuracy (right). The 336-sample training set combines recoverable cases (fail at pass@1 and succeed by pass@8) with a small number of pass@1 successes and pass@8 failures.
Figure 7: Targeted-GRPO training curves of reward (top) and ratio of zero-std reward in each batch (bottom). GRPO is initialized from MQ3-VL-4B-I and uses the ternary validity/InChIKey reward.
Table 10: Ten MolReps representations for one molecule. A single sampled randomized form is displayed for each randomized representation; the CML document is abbreviated, and the 2D image is rendered with rdkit.Draw .
Example
Content
1: R. DeepSMILES → SELFIES
User: Clean up the randomized DeepSMILES O=CCCCCCCCCCCCCCCC)))))))))))))))[O-] and give me the SELFIES. Just provide the SELFIES without any explanation.
User: Take [C@@]C)[C@H]CCC=C[C@H]6[C@@H]CC%10))CC)C)))))C)))))O (a noisy DeepSMILES) and express it as proper SELFIES. Return only the SELFIES representation.
User: This DeepSMILES CO)[C@H]O[C@H]O)[C@H][C@@H]O[C@H]O[C@H]C)[C@H]O)C[C@H]6O))))))))[C@@H]6O)))O[C@@H][C@@H][C@H][C@H][C@@H]CO))O6))O))O))O may not be canonical. Normalize it and output as InChI. Please output the InChI representation only.
Table 11: Three MolReps QA examples, all cross-representation normalization tasks with a randomized DeepSMILES (R. DeepSMILES) input. “Assistant” is the reference completion used for supervised post-training and evaluation.
The Chinese University of Hong Kong, Shenzhen · Shanghai Artificial Intelligence Laboratory · Cheriton School of Computer Science, University of Waterloo