Multi-Granular Rationale-Guided Molecular LLM for Property Prediction

# Multi-Granular Rationale-Guided Molecular LLM for Property Prediction
Source: [https://arxiv.org/html/2608.10480](https://arxiv.org/html/2608.10480)
###### Abstract

Large language models \(LLMs\) are widely applied across chemical tasks, such as molecular property prediction, which underpins drug discovery\. Molecular LLMs represent a molecule through several modalities, notably a 1D SMILES sequence or a 2D molecular graph\. Both encode molecular information implicitly, so the contribution of individual substructures remains opaque\. Retrieval and augmentation methods add context, but from external sources\. However, the cues chemists reason over are the internal substructures that drive a property up or down\. We propose MR\-MoL, a multi\-granular rationale\-guided molecular LLM that supplies this evidence directly\. A fine\-tuned GNN scores each substructure through masking, and the most influential ones are serialized as a ranked, direction\-tagged rationale that the LLM reads alongside the SMILES sequence and molecular graph\. The rationale spans three levels of granularity: Murcko scaffolds with their side chains, BRICS fragments, and functional groups\. This is, to our knowledge, the first method to expose GNN\-derived attributions to an LLM as evidence for property prediction\. On eight MoleculeNet tasks, MR\-MoL achieves the best overall results among generalist models and narrows the gap to specialist models tuned for each task\. Five diagnostics further confirm that the model reads the rationale rather than merely benefiting from its presence\. Its direction, rank, and substructure each shape the prediction, and its attributions reproduce known structure\-property relationships\.

Code—https://github\.com/skku\-aihclab/MR\-MoL

## Introduction

Molecular property prediction underpins drug discovery, toxicity screening, and solubility design\. Deep learning models for this task have traditionally operated directly on diverse molecular representations: 2D molecular graphs\(Xia et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib45); Wang et al\.[2022](https://arxiv.org/html/2608.10480#bib.bib40); Liu et al\.[2022](https://arxiv.org/html/2608.10480#bib.bib25)\), 1D Simplified Molecular\-Input Line\-Entry System \(SMILES\) sequences\(Ross et al\.[2022](https://arxiv.org/html/2608.10480#bib.bib36); Chithrananda, Grand, and Ramsundar[2020](https://arxiv.org/html/2608.10480#bib.bib5)\), and 3D conformers\(Zhou et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib54); Schütt et al\.[2018](https://arxiv.org/html/2608.10480#bib.bib38)\)\. Such models remain among the strongest predictors on standardized benchmarks such as MoleculeNet\(Wu et al\.[2018](https://arxiv.org/html/2608.10480#bib.bib43)\), but are typically trained or fine\-tuned per task\.

More recently, large language models \(LLMs\) have been adapted to molecular property prediction\. Applying an LLM to a molecule first requires casting its representation into a form the model can consume, and two routes have emerged\. The first route keeps everything textual\. The molecule is serialized as a string and fed to the LLM as text tokens\(Yu et al\.[2024](https://arxiv.org/html/2608.10480#bib.bib48); Fang et al\.[2024](https://arxiv.org/html/2608.10480#bib.bib8)\)\. The second route attaches a graph encoder and aligns it to the LLM through a projector, so that the molecule enters the prompt as continuous embeddings\. Some of these models are fine\-tuned per task\(Liu et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib27); Chen et al\.[2024](https://arxiv.org/html/2608.10480#bib.bib4)\), while others share a single model over many chemical tasks\(Park et al\.[2024](https://arxiv.org/html/2608.10480#bib.bib32); Kim, Lee, and Hwang[2026](https://arxiv.org/html/2608.10480#bib.bib18)\)\.

Yet these molecular LLMs leave a common gap\. Since modalities such as 2D graph and 1D SMILES representations encode molecular information implicitly, the contribution of the molecule’s substructures remains opaque\(Wellawatte et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib42); Jiménez\-Luna, Grisoni, and Schneider[2020](https://arxiv.org/html/2608.10480#bib.bib16); Wu et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib44)\)\. Adding retrieval or knowledge graphs to the prompt can supply extra context at the molecule level\(Xian et al\.[2025](https://arxiv.org/html/2608.10480#bib.bib46); Lee et al\.[2026](https://arxiv.org/html/2608.10480#bib.bib21)\), but none of these channels exposes the importance of the current molecule’s substructures in a form the model can read\. In practice, this is the evidence chemists rely on, namely a functional group or fragment that elevates or suppresses a property\. Motivated by this gap, we equip the LLM with explicit, per\-substructure evidence specific to the input molecule\.

We propose MR\-MoL, aMulti\-granularRationale\-GuidedMolecularLLM\. The idea is simple\. We let a fine\-tuned graph neural network \(GNN\) score each substructure’s contribution to its prediction\(Wu et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib44)\), and give this to the LLM as a rationale, a ranked and direction\-tagged list indicating whether each substructure pushes the GNN’s prediction higher or lower\. We serialize it as text, so the LLM reads it directly alongside SMILES sequences and molecular graph representations\. Specifically, we draw the rationale from three views at different structural granularities: Murcko scaffolds together with their side chains\(Bemis and Murcko[1996](https://arxiv.org/html/2608.10480#bib.bib2)\), retrosynthetically motivated BRICS fragments\(Degen et al\.[2008](https://arxiv.org/html/2608.10480#bib.bib6)\), and chemically named functional groups\.

This design supports two claims\. First, the rationale should improve predictive performance, since structural cues are known to matter in molecular property prediction\(Fang et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib9); Zhang et al\.[2021](https://arxiv.org/html/2608.10480#bib.bib50); Wu et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib44)\)\. Second, the LLM should use rationales as directional and ranked evidence, not as inert filler\. We evaluate the first claim against seven specialist models, comprising five GNN\-based models and two molecular LLMs that are fine\-tuned per task, and five generalist molecular LLMs that share a single model across tasks\. We test the second claim with five diagnostics, four behavioral interventions that edit the rationale and one chemical check on whether its attributions reproduce known structure\-property relationships\.

Our contributions are as follows\.

- •We introduce MR\-MoL, a multi\-task molecular LLM for property prediction\. To our knowledge, it is the first to feed GNN\-derived attributions into the prompt as evidence\.
- •Building on substructure masking, we formulate a rationale that decomposes per\-substructure attribution into three complementary views and exposes it as direction\-tagged, ranked textual evidence\.
- •We benchmark MR\-MoL on eight MoleculeNet tasks with ablations, and design diagnostics that show the rationale acts as effective evidence\.

## Related Work

#### Instruction\-Tuned Molecular Language Models\.

A growing body of molecular LLMs is instruction\-tuned on molecule\-text pairs to follow natural\-language task descriptions\(Wei et al\.[2022](https://arxiv.org/html/2608.10480#bib.bib41); Ouyang et al\.[2022](https://arxiv.org/html/2608.10480#bib.bib31)\)\. LlaSMol\(Yu et al\.[2024](https://arxiv.org/html/2608.10480#bib.bib48)\)and nach0\(Livne et al\.[2024](https://arxiv.org/html/2608.10480#bib.bib28)\)tune LLMs on large chemistry instruction corpora, representing molecules as SMILES sequences\. The molecule reaches the LLM through this textual prompt alone\. The other group attaches a modality\-specific encoder and aligns it to the LLM through a projector\. MolCA\(Liu et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib27)\), LLaMo\(Park et al\.[2024](https://arxiv.org/html/2608.10480#bib.bib32)\), and InstructMol\(Cao et al\.[2025](https://arxiv.org/html/2608.10480#bib.bib3)\)project a molecular graph into the LLM, while others enrich this representation with a 2D image\(Liu et al\.[2024a](https://arxiv.org/html/2608.10480#bib.bib24)\), hierarchical motif\-level graph tokens\(Chen et al\.[2024](https://arxiv.org/html/2608.10480#bib.bib4)\), or a 3D conformer\(Li et al\.[2024](https://arxiv.org/html/2608.10480#bib.bib23); Kim, Lee, and Hwang[2026](https://arxiv.org/html/2608.10480#bib.bib18)\)\. Whether molecular information enters as text or as projected tokens, neither path makes the substructure\-level signal legible to the model\.

#### Molecular GNNs and Explainability\.

GNNs treat a molecule as a graph of atoms and bonds, learning representations by message passing between neighboring atoms\(Scarselli et al\.[2008](https://arxiv.org/html/2608.10480#bib.bib37); Gilmer et al\.[2017](https://arxiv.org/html/2608.10480#bib.bib10)\)\. To learn representations that transfer across tasks, self\-supervised pretraining has become the dominant approach\(Hu\* et al\.[2020](https://arxiv.org/html/2608.10480#bib.bib14); Wang et al\.[2022](https://arxiv.org/html/2608.10480#bib.bib40); Liu et al\.[2022](https://arxiv.org/html/2608.10480#bib.bib25)\)\. Such pretrained GNNs are strong predictors, yet the effect of an individual substructure stays hidden within their representations\. To expose it, a parallel line of work attributes predictions back to subgraphs, with GNNExplainer\(Ying et al\.[2019](https://arxiv.org/html/2608.10480#bib.bib47)\)ranking individual nodes and edges and SubgraphX\(Yuan et al\.[2021](https://arxiv.org/html/2608.10480#bib.bib49)\)scoring whole connected subgraphs\. FollowingWu et al\. \([2023](https://arxiv.org/html/2608.10480#bib.bib44)\), we adopt substructure masking, which scores each fragment by the prediction change caused when it is removed\. Such interpretability is widely viewed as essential for bridging black\-box predictors and human\-understandable chemistry\(Jiménez\-Luna, Grisoni, and Schneider[2020](https://arxiv.org/html/2608.10480#bib.bib16); Wellawatte et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib42)\)\.

#### Knowledge\-Augmented Molecular Prediction\.

A separate line of work augments property prediction with auxiliary knowledge\. For instance, LLM\-MPP\(Jin et al\.[2025](https://arxiv.org/html/2608.10480#bib.bib17)\)and MolProphecy\(Zhao et al\.[2025](https://arxiv.org/html/2608.10480#bib.bib52)\)encode LLM\-generated descriptions or chemical knowledge and fuse them with molecular features\. Other approaches focus on retrieving supporting evidence; MolRAG\(Xian et al\.[2025](https://arxiv.org/html/2608.10480#bib.bib46)\)draws on analogous molecules, while CLADD\(Lee et al\.[2026](https://arxiv.org/html/2608.10480#bib.bib21)\)employs collaborative LLM agents that query knowledge bases\. Separately, KANO\(Fang et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib9)\)builds chemical priors into the GNN by coupling graph contrastive learning with a functional group knowledge graph\. Across these methods, the supplied evidence is external or holistic, rather than identifying how specific substructures influence the target property\.

## Method

MR\-MoL takes four inputs: a task instructionII, a 1D SMILES sequenceSS, a 2D molecular graphGG, and a multi\-granular rationaleRR\. The outputYYis the answer to the task, produced as text\.

Two complementary information paths feed the LLM\. Thegraph embedding pathturnsGGinto molecular tokens through a GNN encoder and a projector\. Therationale pathmasks substructures against a fine\-tuned GNN and serializes the most influential ones as ranked items\. The LLM conditions on the instruction, SMILES sequence, molecular tokens, and rationale to produceYY, as shown in Figure[1](https://arxiv.org/html/2608.10480#Sx3.F1)\.

We train MR\-MoL in two stages\. Stage 1 performs molecular graph\-language alignment for the graph embedding path\. Stage 2 performs multi\-task, rationale\-guided instruction tuning for property prediction\.

![Refer to caption](https://arxiv.org/html/2608.10480v1/x1.png)Figure 1:Overview of MR\-MoL\. The graph embedding path encodes a 2D molecular graph through a graph encoder and the projector, yielding molecular tokens\. The rationale path applies substructure masking to a source predictor, scoring substructures by their attribution\. The top\-scoring substructures form the items of a multi\-granular rationale that is provided to the LLM together with the instruction, SMILES, and molecular tokens, and the model generates the answer to the task\.### Molecular Graph\-Language Alignment

The projector module for graph\-language alignment follows the Q\-Former design\(Li et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib22)\), which has been adopted for grounding molecular graphs in LLMs\. We initialize the GNN encoder and Q\-Former with weights pre\-trained on molecule\-text data\(Liu et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib27)\), and then train them with a newly added linear projection during Stage 1 to align the molecular tokens with the embedding space of our LLM\.

GivenG=\(V,E\)G=\(V,E\), the encoderEϕE\_\{\\phi\}with parametersϕ\\phiproduces atom\-level features:

HG=Eϕ​\(G\),HG∈ℝN×dg,H\_\{G\}=E\_\{\\phi\}\(G\),\\qquad H\_\{G\}\\in\\mathbb\{R\}^\{N\\times d\_\{g\}\},\(1\)whereNNis the number of atoms anddgd\_\{g\}is the graph hidden dimension\. The Q\-Former with parametersψ\\psiusesKKlearnable queriesQ∈ℝK×dqQ\\in\\mathbb\{R\}^\{K\\times d\_\{q\}\}and cross\-attends toHGH\_\{G\}, wheredqd\_\{q\}is the Q\-Former hidden dimension:

Z=QFormerψ​\(Q,HG\),Z∈ℝK×dq\.Z=\\mathrm\{QFormer\}\_\{\\psi\}\(Q,H\_\{G\}\),\\qquad Z\\in\\mathbb\{R\}^\{K\\times d\_\{q\}\}\.\(2\)A linear projectionWp∈ℝdq×dℓW\_\{p\}\\in\\mathbb\{R\}^\{d\_\{q\}\\times d\_\{\\ell\}\}then maps the query embeddingsZZto the LLM hidden dimensiondℓd\_\{\\ell\}:

M=Z​Wp,M∈ℝK×dℓ\.M=ZW\_\{p\},\\qquad M\\in\\mathbb\{R\}^\{K\\times d\_\{\\ell\}\}\.\(3\)These molecular tokensMMare inserted into the LLM input sequence\.

### Ranked Multi\-Granular Rationale

The rationale is task\-conditioned textual evidence built from a fine\-tuned GNN\. We decompose the molecule into candidate substructures, score each by its attribution\(Wu et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib44)\), and serialize the top ones as ranked text\.

#### Multi\-granular decomposition\.

For each molecule, we decompose the structure at three granularities, each realized as a view\. TheMurcko viewsplits the molecule into its scaffold and side chains through Murcko decomposition\(Bemis and Murcko[1996](https://arxiv.org/html/2608.10480#bib.bib2)\)\. TheBRICS viewprovides retrosynthetically motivated fragments with attachment information\(Degen et al\.[2008](https://arxiv.org/html/2608.10480#bib.bib6)\)\. Thefunctional group viewlists chemically named local motifs such as amide and carbonyl\. The three views are complementary, spanning a range of structural granularity from molecular scaffolds down to local motifs\.

#### Attribution and ranking\.

Letftf\_\{t\}denote the fine\-tuned source predictor for tasktt, whose outputft​\(G\)f\_\{t\}\(G\)is the positive\-class score for classification or the predicted value for regression\. Let𝒰​\(G\)\\mathcal\{U\}\(G\)be the candidate substructures pooled from the three views, indexed byjj\. For each substructureuj∈𝒰​\(G\)u\_\{j\}\\in\\mathcal\{U\}\(G\), the attribution is

at,j=ft​\(G\)−ft​\(G∖uj\),a\_\{t,j\}=f\_\{t\}\(G\)\-f\_\{t\}\(G\\setminus u\_\{j\}\),\(4\)whereG∖ujG\\setminus u\_\{j\}is the graph withuju\_\{j\}masked\. A positiveat,ja\_\{t,j\}means that removinguju\_\{j\}lowers the prediction, so the substructure drives the prediction up; a negativeat,ja\_\{t,j\}indicates the reverse\. We rank substructures within each molecule by\|at,j\|\|a\_\{t,j\}\|and keep the top five, those producing the largest change in the source prediction\.

#### Serialization\.

The rationaleRRlists the top five substructures as ranked items\. Each item carries three fields, the view type, the substructure, and the effect\. The substructure is written as scaffold or side chain SMILES for the Murcko view, as fragment SMILES with attachment points for the BRICS view, and as the group name for the functional group view\. The effect field tags the item astoward higherortoward lower, applied to the predicted probability of the target property for classification and the predicted value for regression\. This wording is an attribution about the source predictor, not a chemical mechanism claim\. A complete serialized input is given in Appendix B\.

### Training

#### Two\-stage training\.

Both stages cast learning as instruction following over samples in a chat format, where the molecular tokens replace the molecule\-embedding placeholders and the model is trained to generate the targetYY\. Serving different purposes, the two stages differ in the targetYYand the presence of the rationaleRR\. In Stage 1, each sample takes the form\(I,S,G,Y\)\(I,S,G,Y\)and the targetYYis a molecule description\. In Stage 2, each sample takes the form\(I,S,G,R,Y\)\(I,S,G,R,Y\), adding the rationaleRR, andYYis the answer to the property prediction task\. Stage 1 makes the molecular tokens legible to the LLM, a grounding step where the rationale plays no role\. Stage 2 then inherits these grounded tokens and adds the rationale as task\-conditioned evidence for property prediction\.

#### Objective\.

We optimize next\-token prediction, where the prompt tokens are excluded from the loss and only the answer tokens contribute\. WritingY=\(y1,…,yT\)Y=\(y\_\{1\},\\dots,y\_\{T\}\)for theTTanswer tokens andX~\\widetilde\{X\}for the embedded input sequence, the answer likelihood factorizes as

pθ​\(Y∣G,S,I,R\)=∏m=1Tpθ​\(ym∣y<m,X~\),p\_\{\\theta\}\(Y\\mid G,S,I,R\)=\\prod\_\{m=1\}^\{T\}p\_\{\\theta\}\(y\_\{m\}\\mid y\_\{<m\},\\widetilde\{X\}\),\(5\)and the training loss is its negative log\-likelihood,

ℒLM=−∑m=1Tlog⁡pθ​\(ym∣y<m,X~\)\.\\mathcal\{L\}\_\{\\mathrm\{LM\}\}=\-\\sum\_\{m=1\}^\{T\}\\log p\_\{\\theta\}\(y\_\{m\}\\mid y\_\{<m\},\\widetilde\{X\}\)\.\(6\)Stage 1 minimizes the loss over the molecule\-text data, whose targets are descriptions, and Stage 2 over the multi\-task instruction data, whose targets are property prediction answers\. Each stage selects its checkpoint by validation loss\.

## Experiments

### Datasets and Preprocessing

Our data is organized by training stage\. Stage 1 draws molecule\-text pairs from PubChem324k\(Liu et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib27)\), DrugBank\(Knox et al\.[2024](https://arxiv.org/html/2608.10480#bib.bib19)\), Mol\-Instructions\(Fang et al\.[2024](https://arxiv.org/html/2608.10480#bib.bib8)\), and ChEBI\-20\(Edwards, Zhai, and Ji[2021](https://arxiv.org/html/2608.10480#bib.bib7); Hastings et al\.[2012](https://arxiv.org/html/2608.10480#bib.bib12)\)\. To ensure quality, we keep only pairs whose description spans at least 30 non\-whitespace characters, and we drop those duplicated across different molecules\. Molecules shared across sources are also deduplicated\. This filtering yields 89,919 samples\. The instructions for this stage follow Mol\-Instructions\. We hold out the ChEBI\-20 test split as validation data for Stage 1 checkpoint selection\.

Stage 2 covers eight MoleculeNet datasets\(Wu et al\.[2018](https://arxiv.org/html/2608.10480#bib.bib43)\), six classification and two regression, where ClinTox, SIDER, and Tox21 are multi\-label\. These span physiological and biophysical classification, such as toxicity \(ClinTox, Tox21, SIDER\), blood\-brain barrier permeability \(BBBP\), and bioactivity \(BACE, HIV\), together with physicochemical regression on aqueous solubility \(ESOL\) and lipophilicity \(Lipo\)\. Each dataset is divided 8:1:1 into train, validation, and test sets using a scaffold split, and the per\-task sizes are reported in Appendix A\. We build the Stage 2 instructions on SMolInstruct\(Yu et al\.[2024](https://arxiv.org/html/2608.10480#bib.bib48)\), supplementing the uncovered datasets with Claude Opus 4\.6\(Anthropic[2026](https://arxiv.org/html/2608.10480#bib.bib1)\), and evaluate the molecular LLMs with the same instructions\.

All SMILES are canonicalized withRDKit \([2006](https://arxiv.org/html/2608.10480#bib.bib35)\), and molecules shared between stages are removed\.

### Baselines

We compare MR\-MoL against two families of baselines, specialist models and generalist models for property prediction\.

#### Specialist\.

We define specialist models as those fine\-tuned per task\. We include seven such baselines, spanning GNN\-based models and molecular LLMs\. Here a GNN\-based model encodes the molecular graph with a GNN and outputs the property directly\. These models differ in their pretraining strategy\. MolCLR\(Wang et al\.[2022](https://arxiv.org/html/2608.10480#bib.bib40)\)and GraphMVP\(Liu et al\.[2022](https://arxiv.org/html/2608.10480#bib.bib25)\)use contrastive learning, and Mole\-BERT\(Xia et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib45)\)combines it with masked atom modeling\. MGSSL\(Zhang et al\.[2021](https://arxiv.org/html/2608.10480#bib.bib50)\)is motif\-generative, and KANO\(Fang et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib9)\)augments pretraining with a chemical knowledge graph\. Among the molecular LLMs, 3D\-MolT5\(Pei et al\.[2025](https://arxiv.org/html/2608.10480#bib.bib33)\)combines 1D and 3D representations, while HIGHT\(Chen et al\.[2024](https://arxiv.org/html/2608.10480#bib.bib4)\)uses hierarchical graph tokenization\.

#### Generalist\.

We define generalist models as molecular LLMs that are instruction\-tuned and applied across many tasks\. We include five generalist baselines, grouped by how they represent molecules\. LlaSMol\(Yu et al\.[2024](https://arxiv.org/html/2608.10480#bib.bib48)\), ChemDFM\(Zhao et al\.[2024](https://arxiv.org/html/2608.10480#bib.bib53)\), nach0\(Livne et al\.[2024](https://arxiv.org/html/2608.10480#bib.bib28)\), and MolecularGPT\(Liu et al\.[2024b](https://arxiv.org/html/2608.10480#bib.bib26)\)are string\-based models that process molecules as 1D sequences such as SMILES or SELFIES\. In contrast, GIMLET\(Zhao et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib51)\)is a graph\-text model that directly encodes 2D molecular graphs\.

### Implementation Details

We use Llama\-3\.1\-8B\-Instruct\(Grattafiori et al\.[2024](https://arxiv.org/html/2608.10480#bib.bib11)\)as the LLM\. For the graph embedding path, the GNN encoder and Q\-Former are initialized from MolCA’s\(Liu et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib27)\)stage\-1 pre\-trained weights\. The Q\-Former uses eight learnable query tokens and is followed by a projection layer\. In Stage 1, we freeze the LLM and train the GNN encoder and the projector\. In Stage 2, the GNN encoder stays frozen, and we adapt the LLM with Low\-Rank Adaptation \(LoRA\)\(Hu et al\.[2022](https://arxiv.org/html/2608.10480#bib.bib13)\)by training the adapters together with the projector\. For each task, we fine\-tune a separate GNN initialized from Mole\-BERT\(Xia et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib45)\)to generate rationales\. We optimize with AdamW\(Loshchilov and Hutter[2019](https://arxiv.org/html/2608.10480#bib.bib29)\)under a cosine schedule with linear warmup, and provide the full training and optimization details in Appendix A\.

Every component of our method is run three times under the same split for statistical robustness\. This covers both the GNNs that generate the rationales and all MR\-MoL variants in Table[2](https://arxiv.org/html/2608.10480#Sx4.T2)\. Diagnostic experiments perturb only the inference\-time input of MR\-MoL on the test split\. Every other baseline is reproduced in a single run\. We fine\-tune specialist models on our training split, whereas generalist models are evaluated from their publicly available checkpoints\.

### Evaluation Metrics

We report ROC\-AUC for classification and RMSE for regression as the main metrics, following MoleculeNet’s evaluation protocol\(Wu et al\.[2018](https://arxiv.org/html/2608.10480#bib.bib43)\)\. On multi\-label datasets, ROC\-AUC is computed per endpoint and averaged across endpoints\.

Unlike GNN\-based models, which output the score directly, molecular LLMs including MR\-MoL require reading the score from the generated output\. In classification, we apply a two\-way softmax over the logits of the positive and negative answer tokens \(e\.g\., “Yes” and “No”\)\. With logitszYesz\_\{\\mathrm\{Yes\}\}andzNoz\_\{\\mathrm\{No\}\}, the positive\-class score is

P​\(Yes\)=exp⁡\(zYes\)exp⁡\(zYes\)\+exp⁡\(zNo\),P\(\\mathrm\{Yes\}\)=\\frac\{\\exp\(z\_\{\\mathrm\{Yes\}\}\)\}\{\\exp\(z\_\{\\mathrm\{Yes\}\}\)\+\\exp\(z\_\{\\mathrm\{No\}\}\)\},\(7\)which we use for ROC\-AUC\. In regression, we parse the first valid number, and exclude outputs without a parseable number from the computation\.

For the direction diagnostic of MR\-MoL, we additionally report the Matthews correlation coefficient \(MCC\) for classification and the mean error \(ME\) for regression\.

Table 1:Results on eight MoleculeNet property prediction tasks\. Classification is reported as ROC\-AUC and regression as RMSE\. Within the Specialist and Generalist groups, the best score per task is inboldand the overall best across groups isunderlined\. GNN\-based models have no base LLM and are marked –\.Table 2:Ablation of MR\-MoL on molecular graphGGand rationaleRR\. Classification is reported as ROC\-AUC and regression as RMSE\. All variants report mean±\\pmstd\. Best per task in bold\.

## Results and Analysis

We organize our analysis around the two claims of the introduction, improved prediction and effective evidence use\.

### Property Prediction Results

Table[1](https://arxiv.org/html/2608.10480#Sx4.T1)reports the property prediction results, and Table[2](https://arxiv.org/html/2608.10480#Sx4.T2)reports the ablation\. Baselines in Table[1](https://arxiv.org/html/2608.10480#Sx4.T1)are grouped into specialist and generalist families\. For the ablation, we vary the two channels MR\-MoL adds to the LLM, the molecular graphGGand the rationaleRR\.W/oG,RG,Rremoves both channels, leaving only the instruction and SMILES, andw/oGGremoves only the graph\. Neither uses the Stage 1 checkpoint, as the graph path is absent\.W/oRRremoves the rationale\.

#### Comparison with baselines\.

Among the generalist models, MR\-MoL outperforms every baseline on six of the eight tasks, and the margin is often substantial\. It leads by more than1111ROC\-AUC points on BACE and by nearly77on SIDER\. The advantage carries over to regression, where every other generalist exceeds3\.73\.7RMSE on ESOL against our1\.2101\.210\. We attribute this to the structural evidence MR\-MoL receives alongside the 1D and 2D molecule representations\. The two exceptions are HIV and Lipo, where nach0 leads\.

More notably, MR\-MoL closes much of the gap to the specialist models, which are fine\-tuned per task\. It surpasses both molecular LLMs, 3D\-MolT5 and HIGHT, on four of the six classification tasks and on ESOL, and on SIDER it achieves the best result across all baselines\. Elsewhere in classification it trails the strongest specialist by at most4\.54\.5points, with the sole exception of ClinTox, where KANO keeps a clear lead\. In regression the strongest GNN\-based models retain an advantage on both tasks, as generating numbers remains harder for an LLM\. We report output validity, the fraction of parseable answers for molecular LLMs, in Appendix C\.

#### Ablation\.

As shown in Table[2](https://arxiv.org/html/2608.10480#Sx4.T2), removing the rationale lowers performance on seven of the eight tasks, with the largest classification drops on ClinTox and BACE, confirming that the rationale carries information the model uses\. Removing the graph also hurts most tasks, though typically by a smaller margin on classification\. The two channels contribute unequally across tasks, with the rationale mattering more on classification and the graph being the larger contributor on the regression tasks\. This pattern indicates that the rationale complements the molecular graph, and that the model draws on whichever channel is more informative for a given task instead of relying on one alone\. The sole exception is HIV, where neither added channel helps and the w/oGG,RRvariant already suffices\. Notably, w/oGG,RRalready surpasses the generalist baselines in Table[1](https://arxiv.org/html/2608.10480#Sx4.T1)on most tasks, since it shares our strong LLM backbone and multi\-task instruction tuning\. The gains of MR\-MoL over this variant thus isolate the contribution of theGGandRRchannels\. Extended metrics, including F1, MCC, and accuracy, appear in Appendix C\.

### Rationale Diagnostics

A rationale that raises performance matters only if the model reads it rather than benefiting from its mere presence\. We probe this from two sides\. Behaviorally, we intervene on the rationale, by flipping a direction, removing an item, or replacing a substructure, and measure how the prediction moves on four datasets \(BACE, BBBP, ESOL, and Lipo\), then trace how the rationale shifts the answer on individual molecules\. Chemically, we ask whether the attributions the rationale carries reproduce known structure\-property relationships\. All diagnostics use a single checkpoint of MR\-MoL, so the unperturbed scores differ slightly from Table[1](https://arxiv.org/html/2608.10480#Sx4.T1)\.

#### Direction sensitivity\.

We flip only the direction tag in each item, so that*toward higher*becomes*toward lower*and the reverse, while leaving the substructure and its rank intact\. A model that reads the tags should move its prediction\. Table[3](https://arxiv.org/html/2608.10480#Sx5.T3)reports the effect\. On both classification datasets the ROC\-AUC falls below chance and the MCC flips sign, so MR\-MoL does not merely lose the signal but follows the inverted tag\. The reversal is not total, as the MCC magnitude shrinks rather than mirroring the original, indicating that MR\-MoL weighs the direction tag together with other signals\. On both regression datasets the RMSE rises while the mean error changes sign, so the direction also steers the numerical prediction\.

Table 3:Direction flipping\. Each rationale item’s direction tag is inverted\. “Original” and “Flipped” report performance before and after the flip\.
#### Rank sensitivity\.

Rationale items are ranked by attribution magnitude\. For each moleculeiiwe remove the rank\-1 item and measure the prediction shift

Δtop,i=\{Pfull​\(Yes\)i−P−top​\(Yes\)iclassification,y^full,i−y^−top,iregression,\\Delta\_\{\\mathrm\{top\},i\}=\\begin\{cases\}P\_\{\\mathrm\{full\}\}\(\\mathrm\{Yes\}\)\_\{i\}\-P\_\{\-\\mathrm\{top\}\}\(\\mathrm\{Yes\}\)\_\{i\}&\\text\{classification\},\\\\ \\hat\{y\}\_\{\\mathrm\{full\},i\}\-\\hat\{y\}\_\{\-\\mathrm\{top\},i\}&\\text\{regression\},\\end\{cases\}\(8\)whereP​\(Yes\)P\(\\mathrm\{Yes\}\)is the positive\-class probability andy^\\hat\{y\}is the predicted numeric value\. The subscriptfull\\mathrm\{full\}denotes the prediction given the entire rationale, and−top\-\\mathrm\{top\}the prediction after removing the rank\-1 item\. We defineΔrand,i\\Delta\_\{\\mathrm\{rand\},i\}in the same way, removing instead one item drawn at random from the remaining items below rank 1\. We report\|Δtop\|\|\\Delta\_\{\\mathrm\{top\}\}\|and\|Δrand\|\|\\Delta\_\{\\mathrm\{rand\}\}\|, the mean of\|Δtop,i\|\|\\Delta\_\{\\mathrm\{top\},i\}\|and\|Δrand,i\|\|\\Delta\_\{\\mathrm\{rand\},i\}\|over all molecules\. The rank\-1 item is the one whose masking most changed the source prediction, so a model that treats rank as importance should react most when it is removed\. As Table[4](https://arxiv.org/html/2608.10480#Sx5.T4)shows,\|Δtop\|\|\\Delta\_\{\\mathrm\{top\}\}\|exceeds\|Δrand\|\|\\Delta\_\{\\mathrm\{rand\}\}\|by1\.61\.6to6\.26\.2times across the four datasets\. A pairedtt\-test confirms the gap is significant on every dataset, with allp<10−5p<10^\{\-5\}\.

Table 4:Rank\-1 versus random lower\-rank removal\. “ratio” is\|Δtop\|/\|Δrand\|\|\\Delta\_\{\\mathrm\{top\}\}\|/\|\\Delta\_\{\\mathrm\{rand\}\}\|;ppis from a pairedtt\-test between the two\.
#### Substructure sensitivity\.

The first two tests vary the direction and the rank, but leave each item’s substructure unchanged\. Here we replace only its substructure with a random one of the same view drawn from the test pool\. Table[5](https://arxiv.org/html/2608.10480#Sx5.T5)confirms a drop on every dataset\. This shuffling lowers ROC\-AUC by22to44points on BACE and BBBP and raises RMSE on ESOL and Lipo\. The model therefore depends on the specific structure each item names\. Together, these three interventions show that MR\-MoL reads every aspect of a rationale, its direction, its rank, and its substructures\.

Table 5:Substructure shuffling\. “Original” is the intact rationale, while “Shuffled” replaces each item’s substructure with a random one\. Classification is reported as ROC\-AUC and regression as RMSE\.Table 6:Mean signed attribution of seven common polar functional groups, computed for each regression dataset\. Parentheses give the percentage of molecules containing the group in which the attribution sign matches the group’s mean\.
#### Chemical validity of attributions\.

For the functional group view, we average each group’s signed attribution over the molecules containing it, separately for ESOL and Lipo, whose targets are aqueous solubility \(log⁡S\\log S\) and lipophilicity \(log⁡D\\log D\)\. Every group raises predicted solubility and lowers predicted lipophilicity \(Table[6](https://arxiv.org/html/2608.10480#Sx5.T6)\), mirroring the textbook inverse relation captured by the general solubility equation\(Jain and Yalkowsky[2001](https://arxiv.org/html/2608.10480#bib.bib15)\)\. This direction holds in 88 to 100 percent of the molecules that contain each group\.

Beyond the groups in Table[6](https://arxiv.org/html/2608.10480#Sx5.T6), the carboxylic acid is a clear example\. It carries the largest negative lipophilicity attribution in our analysis, a mean of−1\.75\-1\.75across 465 molecules in the Lipo dataset\. Its near\-complete ionization at physiological pH sharply lowerslog⁡D\\log D\(Lassalas et al\.[2016](https://arxiv.org/html/2608.10480#bib.bib20)\)\. The same ionization suppresses blood\-brain barrier permeability\(Wager et al\.[2010](https://arxiv.org/html/2608.10480#bib.bib39)\), which the rationale captures in 93 percent of the molecules in the BBBP dataset that contain it\. These attributions align with established chemistry, so the evidence the rationale carries is not only readable but chemically grounded\.

![Refer to caption](https://arxiv.org/html/2608.10480v1/x2.png)Figure 2:A multi\-granular rationale corrects a blood\-brain barrier permeability prediction for nalidixic acid\. The number beside each substructure in the rationale is its attribution magnitude\|at,j\|\|a\_\{t,j\}\|, which ranks the items\. The bottom axis shows the resulting shift ofP​\(Yes\)P\(\\mathrm\{Yes\}\), the predicted probability of the “Yes” label, across the decision boundary\.Table 7:Rationale item interventions\. Cells showP​\(Yes\)P\(\\mathrm\{Yes\}\)for classification and the predicted value for regression\. “No” is the prediction without the rationale, “Full” uses all rationale items, and “Rm\.” removes the top\-attributed item\. “Truth” is the ground\-truth label or value\.
#### Correction of individual predictions\.

Moving from populations to individual molecules, we take a molecule that MR\-MoL predicts incorrectly without the rationale, and add the rationale to its input\. Figure[2](https://arxiv.org/html/2608.10480#Sx5.F2)traces how MR\-MoL resolves one case in blood\-brain barrier permeability \(BBBP\)\. Given nalidixic acid without the rationale, MR\-MoL wrongly predicts that it crosses the barrier\. In the rationale, the items from the BRICS and Murcko views are tagged as raising permeability, while the functional group carboxylic acid is tagged as lowering it with the largest magnitude\. With this evidence, MR\-MoL flips to the correct No\.

The same pattern holds on other tasks\. Table[7](https://arxiv.org/html/2608.10480#Sx5.T7)shows one molecule each for BACE and ESOL, where the prediction without the rationale is wrong and the full rationale corrects it\. In each case the most influential substructure is one a chemist would point to, and the direction the rationale assigns it matches known chemistry\. For Perylene, the fused\-ring scaffold is tagged toward lower solubility, consistent with its large hydrophobic surface and stable crystal packing\(Jain and Yalkowsky[2001](https://arxiv.org/html/2608.10480#bib.bib15)\)\. In the HEA inhibitor, the hydroxyethylamine scaffold is tagged toward higher activity, consistent with its protonated amine forming a salt bridge with a catalytic aspartate\(Maillard et al\.[2007](https://arxiv.org/html/2608.10480#bib.bib30)\)\. As a further check, deleting this item returns the model to its error, showing that the correction hinges on that specific item\.

## Limitations

We note several limitations that also point to future directions\. First, the rationale inherits the quality of the source predictor and may carry misleading evidence\. Stronger attribution methods or an ensemble of source predictors could reduce this risk, which we leave to future work\. Second, the Murcko, BRICS, and functional group views are not exhaustive, leaving patterns such as macrocycles and stereochemistry\-driven properties uncovered\. Third, the rationale channel covers only tasks where the GNN source predictor yields attributions, namely classification and regression, leaving tasks such as molecule captioning and reaction prediction beyond its current reach\. Extending it to such tasks is another direction for future work\.

## Conclusion

We presented MR\-MoL, a molecular LLM that uses explicit, ranked, GNN\-derived structural evidence as a rationale alongside the molecular graph and SMILES\. Together they let the model predict molecular properties while grounding each prediction in the substructures that drive it\. Our diagnostics confirm that the model acts on the rationale’s content, not its presence alone, and that the cues it carries reflect known chemistry\. Beyond property prediction, this turns model explanations from a diagnostic readout into an input that sharpens the model itself\.

## References

- Anthropic \(2026\)Anthropic\. 2026\.Claude Opus 4\.6 System Card\.https://www\.anthropic\.com/claude\-opus\-4\-6\-system\-card\.Accessed: 2026\-06\-09\.
- Bemis and Murcko \(1996\)Bemis, G\. W\.; and Murcko, M\. A\. 1996\.The properties of known drugs\. 1\. Molecular frameworks\.*Journal of medicinal chemistry*, 39\(15\): 2887–2893\.
- Cao et al\. \(2025\)Cao, H\.; Liu, Z\.; Lu, X\.; Yao, Y\.; and Li, Y\. 2025\.Instructmol: Multi\-modal integration for building a versatile and reliable molecular assistant in drug discovery\.In*Proceedings of the 31st International Conference on Computational Linguistics*, 354–379\.
- Chen et al\. \(2024\)Chen, Y\.; Yao, Q\.; Zhang, J\.; Cheng, J\.; and Bian, Y\. 2024\.Hight: Hierarchical graph tokenization for graph\-language alignment\.arXiv:2406\.14021\.
- Chithrananda, Grand, and Ramsundar \(2020\)Chithrananda, S\.; Grand, G\.; and Ramsundar, B\. 2020\.ChemBERTa: large\-scale self\-supervised pretraining for molecular property prediction\.arXiv:2010\.09885\.
- Degen et al\. \(2008\)Degen, J\.; Wegscheid\-Gerlach, C\.; Zaliani, A\.; and Rarey, M\. 2008\.On the art of compiling and using ’drug\-like’ chemical fragment spaces\.*ChemMedChem*, 3\(10\): 1503\.
- Edwards, Zhai, and Ji \(2021\)Edwards, C\.; Zhai, C\.; and Ji, H\. 2021\.Text2mol: Cross\-modal molecule retrieval with natural language queries\.In*Proceedings of the 2021 conference on empirical methods in natural language processing*, 595–607\.
- Fang et al\. \(2024\)Fang, Y\.; Liang, X\.; Zhang, N\.; Liu, K\.; Huang, R\.; Chen, Z\.; Fan, X\.; and Chen, H\. 2024\.Mol\-Instructions: A Large\-Scale Biomolecular Instruction Dataset for Large Language Models\.In*The Twelfth International Conference on Learning Representations*\.
- Fang et al\. \(2023\)Fang, Y\.; Zhang, Q\.; Zhang, N\.; Chen, Z\.; Zhuang, X\.; Shao, X\.; Fan, X\.; and Chen, H\. 2023\.Knowledge graph\-enhanced molecular contrastive learning with functional prompt\.*Nature Machine Intelligence*, 5\(5\): 542–553\.
- Gilmer et al\. \(2017\)Gilmer, J\.; Schoenholz, S\. S\.; Riley, P\. F\.; Vinyals, O\.; and Dahl, G\. E\. 2017\.Neural message passing for quantum chemistry\.In*International conference on machine learning*, 1263–1272\. Pmlr\.
- Grattafiori et al\. \(2024\)Grattafiori, A\.; Dubey, A\.; Jauhri, A\.; Pandey, A\.; Kadian, A\.; Al\-Dahle, A\.; Letman, A\.; Mathur, A\.; Schelten, A\.; Vaughan, A\.; et al\. 2024\.The llama 3 herd of models\.arXiv:2407\.21783\.
- Hastings et al\. \(2012\)Hastings, J\.; De Matos, P\.; Dekker, A\.; Ennis, M\.; Harsha, B\.; Kale, N\.; Muthukrishnan, V\.; Owen, G\.; Turner, S\.; Williams, M\.; et al\. 2012\.The ChEBI reference database and ontology for biologically relevant chemistry: enhancements for 2013\.*Nucleic acids research*, 41\(D1\): D456–D463\.
- Hu et al\. \(2022\)Hu, E\. J\.; Shen, Y\.; Wallis, P\.; Allen\-Zhu, Z\.; Li, Y\.; Wang, S\.; Wang, L\.; and Chen, W\. 2022\.LoRA: Low\-Rank Adaptation of Large Language Models\.In*International Conference on Learning Representations*\.
- Hu\* et al\. \(2020\)Hu\*, W\.; Liu\*, B\.; Gomes, J\.; Zitnik, M\.; Liang, P\.; Pande, V\.; and Leskovec, J\. 2020\.Strategies for Pre\-training Graph Neural Networks\.In*International Conference on Learning Representations*\.
- Jain and Yalkowsky \(2001\)Jain, N\.; and Yalkowsky, S\. H\. 2001\.Estimation of the aqueous solubility I: Application to organic nonelectrolytes\.*Journal of Pharmaceutical Sciences*, 90\(2\): 234–252\.
- Jiménez\-Luna, Grisoni, and Schneider \(2020\)Jiménez\-Luna, J\.; Grisoni, F\.; and Schneider, G\. 2020\.Drug discovery with explainable artificial intelligence\.*Nature Machine Intelligence*, 2\(10\): 573–584\.
- Jin et al\. \(2025\)Jin, C\.; Guo, S\.; Zhou, S\.; and Guan, J\. 2025\.Effective and explainable molecular property prediction by chain\-of\-thought enabled large language models and multi\-modal molecular information fusion\.*Journal of Chemical Information and Modeling*, 65\(11\): 5438–5455\.
- Kim, Lee, and Hwang \(2026\)Kim, D\.; Lee, W\.; and Hwang, S\. J\. 2026\.Mol\-llama: Towards general understanding of molecules in large molecular language model\.*Advances in Neural Information Processing Systems*, 38: 26921–26960\.
- Knox et al\. \(2024\)Knox, C\.; Wilson, M\.; Klinger, C\. M\.; Franklin, M\.; Oler, E\.; Wilson, A\.; Pon, A\.; Cox, J\.; Chin, N\. E\.; Strawbridge, S\. A\.; et al\. 2024\.DrugBank 6\.0: the DrugBank knowledgebase for 2024\.*Nucleic acids research*, 52\(D1\): D1265–D1275\.
- Lassalas et al\. \(2016\)Lassalas, P\.; Gay, B\.; Lasfargeas, C\.; James, M\. J\.; Tran, V\.; Vijayendran, K\. G\.; Brunden, K\. R\.; Kozlowski, M\. C\.; Thomas, C\. J\.; Smith, I\., Amos B; Huryn, D\. M\.; and Ballatore, C\. 2016\.Structure property relationships of carboxylic acid isosteres\.*Journal of Medicinal Chemistry*, 59\(7\): 3183–3203\.
- Lee et al\. \(2026\)Lee, N\.; De Brouwer, E\.; Hajiramezanali, E\.; Biancalani, T\.; Park, C\.; and Scalia, G\. 2026\.Rag\-enhanced collaborative llm agents for drug discovery\.In*Proceedings of the AAAI Conference on Artificial Intelligence*, volume 40, 561–569\.
- Li et al\. \(2023\)Li, J\.; Li, D\.; Savarese, S\.; and Hoi, S\. 2023\.Blip\-2: Bootstrapping language\-image pre\-training with frozen image encoders and large language models\.In*International conference on machine learning*, 19730–19742\. PMLR\.
- Li et al\. \(2024\)Li, S\.; Liu, Z\.; Luo, Y\.; Wang, X\.; He, X\.; Kawaguchi, K\.; Chua, T\.\-S\.; and Tian, Q\. 2024\.Towards 3d molecule\-text interpretation in language models\.In*International Conference on Learning Representations*, volume 2024, 17352–17371\.
- Liu et al\. \(2024a\)Liu, P\.; Ren, Y\.; Tao, J\.; and Ren, Z\. 2024a\.Git\-mol: A multi\-modal large language model for molecular science with graph, image, and text\.*Computers in biology and medicine*, 171: 108073\.
- Liu et al\. \(2022\)Liu, S\.; Wang, H\.; Liu, W\.; Lasenby, J\.; Guo, H\.; and Tang, J\. 2022\.Pre\-training Molecular Graph Representation with 3D Geometry\.In*International Conference on Learning Representations*\.
- Liu et al\. \(2024b\)Liu, Y\.; Ding, S\.; Zhou, S\.; Fan, W\.; and Tan, Q\. 2024b\.MolecularGPT: Open Large Language Model \(LLM\) for Few\-Shot Molecular Property Prediction\.arXiv:2406\.12950\.
- Liu et al\. \(2023\)Liu, Z\.; Li, S\.; Luo, Y\.; Fei, H\.; Cao, Y\.; Kawaguchi, K\.; Wang, X\.; and Chua, T\.\-S\. 2023\.Molca: Molecular graph\-language modeling with cross\-modal projector and uni\-modal adapter\.In*Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing*, 15623–15638\.
- Livne et al\. \(2024\)Livne, M\.; Miftahutdinov, Z\.; Tutubalina, E\.; Kuznetsov, M\.; Polykovskiy, D\.; Brundyn, A\.; Jhunjhunwala, A\.; Costa, A\.; Aliper, A\.; Aspuru\-Guzik, A\.; et al\. 2024\.nach0: multimodal natural and chemical languages foundation model\.*Chemical Science*, 15\(22\): 8380–8389\.
- Loshchilov and Hutter \(2019\)Loshchilov, I\.; and Hutter, F\. 2019\.Decoupled Weight Decay Regularization\.In*International Conference on Learning Representations*\.
- Maillard et al\. \(2007\)Maillard, M\. C\.; Hom, R\. K\.; Benson, T\. E\.; Moon, J\. B\.; Mamo, S\.; Bienkowski, M\.; Tomasselli, A\. G\.; Woods, D\. D\.; Prince, D\. B\.; Paddock, D\. J\.; et al\. 2007\.Design, synthesis, and crystal structure of hydroxyethyl secondary amine\-based peptidomimetic inhibitors of humanβ\\beta\-secretase\.*Journal of medicinal chemistry*, 50\(4\): 776–781\.
- Ouyang et al\. \(2022\)Ouyang, L\.; Wu, J\.; Jiang, X\.; Almeida, D\.; Wainwright, C\.; Mishkin, P\.; Zhang, C\.; Agarwal, S\.; Slama, K\.; Ray, A\.; et al\. 2022\.Training language models to follow instructions with human feedback\.*Advances in neural information processing systems*, 35: 27730–27744\.
- Park et al\. \(2024\)Park, J\.; Bae, M\.; Ko, D\.; and Kim, H\. J\. 2024\.Llamo: Large language model\-based molecular graph assistant\.*Advances in Neural Information Processing Systems*, 37: 131972–132000\.
- Pei et al\. \(2025\)Pei, Q\.; Yan, R\.; Gao, K\.; Zhu, J\.; and Wu, L\. 2025\.3D\-MolT5: Leveraging Discrete Structural Information for Molecule\-Text Modeling\.In*International Conference on Learning Representations*\.
- Ramsundar et al\. \(2019\)Ramsundar, B\.; Eastman, P\.; Walters, P\.; Pande, V\.; Leswing, K\.; and Wu, Z\. 2019\.*Deep Learning for the Life Sciences*\.O’Reilly Media\.https://www\.amazon\.com/Deep\-Learning\-Life\-Sciences\-Microscopy/dp/1492039837\.
- RDKit \(2006\)RDKit\. 2006\.RDKit: Open\-source cheminformatics\.https://www\.rdkit\.org\.Accessed: 2026\-05\-06\.
- Ross et al\. \(2022\)Ross, J\.; Belgodere, B\.; Chenthamarakshan, V\.; Padhi, I\.; Mroueh, Y\.; and Das, P\. 2022\.Large\-scale chemical language representations capture molecular structure and properties\.*Nature Machine Intelligence*, 4\(12\): 1256–1264\.
- Scarselli et al\. \(2008\)Scarselli, F\.; Gori, M\.; Tsoi, A\. C\.; Hagenbuchner, M\.; and Monfardini, G\. 2008\.The graph neural network model\.*IEEE transactions on neural networks*, 20\(1\): 61–80\.
- Schütt et al\. \(2018\)Schütt, K\. T\.; Sauceda, H\. E\.; Kindermans, P\.\-J\.; Tkatchenko, A\.; and Müller, K\.\-R\. 2018\.Schnet–a deep learning architecture for molecules and materials\.*The Journal of chemical physics*, 148\(24\)\.
- Wager et al\. \(2010\)Wager, T\. T\.; Hou, X\.; Verhoest, P\. R\.; and Villalobos, A\. 2010\.Moving beyond rules: the development of a central nervous system multiparameter optimization \(CNS MPO\) approach to enable alignment of druglike properties\.*ACS Chemical Neuroscience*, 1\(6\): 435–449\.
- Wang et al\. \(2022\)Wang, Y\.; Wang, J\.; Cao, Z\.; and Barati Farimani, A\. 2022\.Molecular contrastive learning of representations via graph neural networks\.*Nature Machine Intelligence*, 4\(3\): 279–287\.
- Wei et al\. \(2022\)Wei, J\.; Bosma, M\.; Zhao, V\.; Guu, K\.; Yu, A\. W\.; Lester, B\.; Du, N\.; Dai, A\. M\.; and Le, Q\. V\. 2022\.Finetuned Language Models are Zero\-Shot Learners\.In*International Conference on Learning Representations*\.
- Wellawatte et al\. \(2023\)Wellawatte, G\. P\.; Gandhi, H\. A\.; Seshadri, A\.; and White, A\. D\. 2023\.A perspective on explanations of molecular prediction models\.*Journal of Chemical Theory and Computation*, 19\(8\): 2149–2160\.
- Wu et al\. \(2018\)Wu, Z\.; Ramsundar, B\.; Feinberg, E\. N\.; Gomes, J\.; Geniesse, C\.; Pappu, A\. S\.; Leswing, K\.; and Pande, V\. 2018\.MoleculeNet: a benchmark for molecular machine learning\.*Chemical science*, 9\(2\): 513–530\.
- Wu et al\. \(2023\)Wu, Z\.; Wang, J\.; Du, H\.; Jiang, D\.; Kang, Y\.; Li, D\.; Pan, P\.; Deng, Y\.; Cao, D\.; Hsieh, C\.\-Y\.; et al\. 2023\.Chemistry\-intuitive explanation of graph neural networks for molecular property prediction with substructure masking\.*Nature communications*, 14\(1\): 2585\.
- Xia et al\. \(2023\)Xia, J\.; Zhao, C\.; Hu, B\.; Gao, Z\.; Tan, C\.; Liu, Y\.; Li, S\.; and Li, S\. Z\. 2023\.Mole\-bert: Rethinking pre\-training graph neural networks for molecules\.In*The Eleventh International Conference on Learning Representations*\.
- Xian et al\. \(2025\)Xian, Z\.; Gu, J\.; Li, L\.; and Liang, S\. 2025\.Molrag: unlocking the power of large language models for molecular property prediction\.In*Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\)*, 15513–15531\.
- Ying et al\. \(2019\)Ying, Z\.; Bourgeois, D\.; You, J\.; Zitnik, M\.; and Leskovec, J\. 2019\.Gnnexplainer: Generating explanations for graph neural networks\.*Advances in neural information processing systems*, 32\.
- Yu et al\. \(2024\)Yu, B\.; Baker, F\. N\.; Chen, Z\.; Ning, X\.; and Sun, H\. 2024\.LlaSMol: Advancing Large Language Models for Chemistry with a Large\-Scale, Comprehensive, High\-Quality Instruction Tuning Dataset\.In*First Conference on Language Modeling*\.
- Yuan et al\. \(2021\)Yuan, H\.; Yu, H\.; Wang, J\.; Li, K\.; and Ji, S\. 2021\.On explainability of graph neural networks via subgraph explorations\.In*International conference on machine learning*, 12241–12252\. PMLR\.
- Zhang et al\. \(2021\)Zhang, Z\.; Liu, Q\.; Wang, H\.; Lu, C\.; and Lee, C\.\-K\. 2021\.Motif\-based graph self\-supervised learning for molecular property prediction\.*Advances in Neural Information Processing Systems*, 34: 15870–15882\.
- Zhao et al\. \(2023\)Zhao, H\.; Liu, S\.; Ma, C\.; Xu, H\.; Fu, J\.; Deng, Z\.; Kong, L\.; and Liu, Q\. 2023\.GIMLET: A unified graph\-text model for instruction\-based molecule zero\-shot learning\.*Advances in Neural Information Processing Systems*, 36: 5850–5887\.
- Zhao et al\. \(2025\)Zhao, J\.; Zhou, Q\.; Wang, T\.; Fan, Y\.; Yang, Q\.; Jiao, L\.; Liu, C\.; Guo, Z\.; Lu, Q\.; Zhou, F\.; et al\. 2025\.MolProphecy: Bridging medicinal chemists’ knowledge and molecular pre\-trained models via a multi\-modal framework\.*Journal of Advanced Research*\.
- Zhao et al\. \(2024\)Zhao, Z\.; Ma, D\.; Chen, L\.; Sun, L\.; Li, Z\.; Xia, Y\.; Xu, H\.; Zhu, Z\.; Zhu, S\.; Fan, S\.; et al\. 2024\.Chemdfm: A large language foundation model for chemistry\.In*Neurips 2024 Workshop Foundation Models for Science: Progress, Opportunities, and Challenges*\.
- Zhou et al\. \(2023\)Zhou, G\.; Gao, Z\.; Ding, Q\.; Zheng, H\.; Xu, H\.; Wei, Z\.; Zhang, L\.; and Ke, G\. 2023\.Uni\-Mol: A Universal 3D Molecular Representation Learning Framework\.In*The Eleventh International Conference on Learning Representations*\.

## Appendix

This appendix collects additional detail and analysis that support the main paper\. References to the main paper are by section name, and each appendix section below is labeled by a letter for reference from the main text\.

## Appendix AExperimental Setup

### A\.1Dataset Statistics

The eight tasks are MoleculeNet datasets\(Wu et al\.[2018](https://arxiv.org/html/2608.10480#bib.bib43)\), six binary classification and two regression\. BACE asks whether a molecule inhibits humanβ\\beta\-secretase 1, and BBBP whether it penetrates the blood\-brain barrier\. HIV asks whether a molecule inhibits viral replication\. ClinTox has two endpoints, clinical\-trial toxicity and FDA approval\. Tox21 has 12 toxicity assays, seven nuclear\-receptor and five stress\-response pathways \(NR\-AR, NR\-AR\-LBD, NR\-AhR, NR\-Aromatase, NR\-ER, NR\-ER\-LBD, NR\-PPAR\-γ\\gamma, SR\-ARE, SR\-ATAD5, SR\-HSE, SR\-MMP, SR\-p53\)\. SIDER has 27 side\-effect classes from the MedDRA system\-organ ontology\. The regression tasks are ESOL for aqueous solubility and Lipo for the octanol\-water distribution coefficient\.

Table[8](https://arxiv.org/html/2608.10480#A1.T8)reports the per\-task split sizes\. Every task uses an 8:1:1 scaffold split, following the standard MoleculeNet protocol\(Wu et al\.[2018](https://arxiv.org/html/2608.10480#bib.bib43)\)\. For the three multi\-label tasks, namely ClinTox, SIDER, and Tox21, we expand each molecule into one query per labeled endpoint, since the model answers a single property at a time\. The reported counts are therefore queries, and exceed the number of molecules\. The molecule counts, in train/valid/test order, are ClinTox 1,149/148/145, SIDER 1,141/143/143, and Tox21 6,264/783/776\. Tox21 departs slightly from the 8:1:1 ratio because unlabeled endpoints are dropped before expansion\. The single\-label tasks BACE, BBBP, HIV, ESOL, and Lipo have matching query and molecule counts\.

Table 8:Stage 2 dataset statistics\. “Type” is the prediction task type and “End\.” is the number of endpoints\. Counts are queries\.
### A\.2Reproducibility Details

We train in two stages, and Table[9](https://arxiv.org/html/2608.10480#A1.T9)lists the full hyperparameter setting\. The base LLM is Llama\-3\.1\-8B\-Instruct\. Stage 1 aligns the graph path with the LLM embedding space by training the graph encoder and the projector under a three\-group learning rate, while the LLM stays frozen\. Stage 2 freezes the graph encoder and the LLM, and adapts the LLM with LoRA\(Hu et al\.[2022](https://arxiv.org/html/2608.10480#bib.bib13)\)jointly with the projector, where the adapters target the attention projections\{q,k,v,o\}proj\\\{q,k,v,o\\\}\_\{\\text\{proj\}\}\. The projector holds a Q\-Former with eight query tokens followed by a linear projection to the LLM\. The effective batch multiplies the per\-device batch by the gradient\-accumulation steps, giving2×32=642\\times 32=64in Stage 1 and2×16=322\\times 16=32in Stage 2\. We select each stage’s checkpoint by the lowest held\-out validation loss\.

Table 9:Training configuration for the two stages\. The Stage 1 learning rate lists the rates for the graph encoder, the Q\-Former, and the linear projection in turn, and the Stage 2 learning rate lists the rates for the LoRA adapters and the projector in turn\.The randomly initialized projection receives the largest Stage 1 rate, while the pretrained graph encoder receives the smallest\. The graph encoder is a five\-layer, 300\-dimensional GIN\. The Q\-Former follows the BLIP\-2 design\(Li et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib22)\)with a 12\-layer SciBERT backbone and six cross\-attention layers, all initialized from MolCA\(Liu et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib27)\)\. The projection maps the 768\-dimensional Q\-Former output to the 4096\-dimensional LLM hidden size\.

#### Stage 1 checkpoint selection\.

We score every Stage 1 checkpoint by next\-token prediction loss and perplexity on the standard ChEBI\-20 test split of3,3003\{,\}300molecules, which we hold out as validation data\. Table[10](https://arxiv.org/html/2608.10480#A1.T10)reports the curve\. The loss plateaus near epoch 10, so we initialize Stage 2 from that checkpoint\.

Table 10:Stage 1 validation loss across epochs\. NTP is the next\-token prediction loss and PPL its perplexity\. The selected checkpoint is in bold\.
#### Run variability and checkpoint selection\.

We repeat the full Stage 2 training three times with random seeds and report the mean and standard deviation, while each baseline is run once\. The three runs converge to close validation losses\. All diagnostics in the main paper use the run with the lowest validation loss\.

#### Source predictor\.

Each task’s rationale comes from a Mole\-BERT\(Xia et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib45)\)GNN fine\-tuned end to end on that task, following the original fine\-tuning setup\. We use Adam with learning rate1​e−31\\mathrm\{e\}\{\-\}3, no weight decay, batch size3232, dropout0\.50\.5, mean pooling, and100100epochs, with no learning\-rate schedule\. We train each source predictor three times per task, and the rationale uses the checkpoint with the lowest validation loss\. The attribution of each item is the masking difference defined in the main paper, and we keep the top five items per molecule\.

#### Hyperparameter selection\.

We tune the Stage 2 hyperparameters and select the final configuration by validation loss, as summarized in Table[11](https://arxiv.org/html/2608.10480#A1.T11)\.

Table 11:Stage 2 hyperparameter search\. “Attention” is the attention projections\{q,k,v,o\}proj\\\{q,k,v,o\\\}\_\{\\text\{proj\}\}and “MLP” the feed\-forward projections\{gate,up,down\}proj\\\{\\text\{gate\},\\text\{up\},\\text\{down\}\\\}\_\{\\text\{proj\}\}\.
#### Regression targets\.

For ESOL and Lipo we load the MoleculeNet datasets through DeepChem\(Ramsundar et al\.[2019](https://arxiv.org/html/2608.10480#bib.bib34)\), which provides the regression targets already z\-normalized with training\-split statistics\. We keep this normalization when fine\-tuning the molecular LLMs, namely MR\-MoL and the specialist LLMs, so they generate values on a standardized scale\. At evaluation we invert the normalization, and every reported metric is in the original target units\. The GNN\-based baselines and the source predictor regress the raw values directly\.

#### Environment\.

All runs use one NVIDIA RTX 5090 with 32 GB of memory and an Intel Core Ultra 9 285K CPU\. The machine runs Ubuntu 24\.04\.3 LTS with Python 3\.12 and CUDA 12\.8, and the main libraries are PyTorch 2\.11, Transformers 4\.57, PEFT 0\.19, RDKit 2026\.03, and PyTorch Geometric 2\.7\. Stage 1 takes about1\.51\.5hours per epoch, and Stage 2 about six hours per epoch\.

### A\.3Prompt Format and Chat Template

Both stages format each sample as a single Llama\-3\.1 chat exchange\. Figure[3](https://arxiv.org/html/2608.10480#A1.F3)shows the skeleton\. The system turn sets the assistant role and, in Stage 2, documents the structure of the rationale\. The user turn carries the instruction, the molecular tokens, the SMILES sequence, and the rationale\. The assistant turn carries the target, a molecule description in Stage 1 and a property answer in Stage 2\. The graph path inserts eight molecule\-embedding placeholders between the molecule delimiters, one per Q\-Former query, and these are overwritten by the projected molecular embeddings\.

```
<|start_header_id|>system<|end_header_id|>
{system_prompt}<|eot_id|>

<|start_header_id|>user<|end_header_id|>
{instruction}
Molecule: <|start_mol_id|> ... <|end_mol_id|>
SMILES: {smiles}
{rationale}<|eot_id|>

<|start_header_id|>assistant<|end_header_id|>
{answer}<|eot_id|>
```

Figure 3:Chat template skeleton for both stages\. The\.\.\.between the molecule delimiters stands for eight molecule\-embedding placeholders\.The Stage 2 system prompt is as follows\.

> “You are a scientific assistant\. You will be provided with an instruction, molecular embeddings, the corresponding SMILES string, and an optional GNN\-derived structural rationale\. - •The rationale is a ranked list of structural items from three decomposition views, namely Murcko, functional group, and BRICS\. - •Items are ranked within the same molecule by the magnitude of the GNN prediction change caused by masking each item\. Ranks are not globally calibrated and are not compared across molecules\. - •The effect field describes the direction of the GNN prediction attributed to the substructure, not a ground\-truth chemical claim\. - •For BRICS view, the raw BRICS string and attachments are decomposition metadata and connection labels\. When a rationale is provided, treat it as auxiliary evidence, not ground truth\.”

The Stage 1 system prompt keeps only the role sentence, without the rationale description\.

## Appendix BRationale Construction and Analysis

### B\.1Source Predictor Choice

We initialize the source predictor from Mole\-BERT\(Xia et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib45)\)\. Three reasons motivate this choice\. First, Mole\-BERT is pretrained to reconstruct masked atoms, so partially masked graphs stay in distribution\. This is exactly the regime in which masking attribution operates\(Wu et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib44)\), since each attribution compares the prediction on the full graph with that on a masked one\. Second, its GIN backbone matches the graph encoder of our graph path\(Liu et al\.[2023](https://arxiv.org/html/2608.10480#bib.bib27)\), which keeps the two graph modules architecturally aligned\. Third, it is a public and reproducible baseline with a fully specified recipe\.

### B\.2Rationale Analysis

This section examines what the rationales contain in practice, covering how items distribute across the three views, how the views overlap, the attribution statistics, and a serialized example\. For each candidate substructure we compute the masking attribution defined in the main paper, rank items within a molecule by absolute attribution, and keep the top five across the three views\.

#### Granularity composition and overlap\.

Table[12](https://arxiv.org/html/2608.10480#A2.T12)reports how the kept items distribute across the three views and how often two items name the same substructure\. BRICS contributes the most items overall at51%51\\%, followed by Murcko at43%43\\%and functional groups at6%6\\%\. ESOL is the exception, where the Murcko view dominates\. Exact duplication across items is uncommon\. Two of the five kept items name an identical substructure in16\.6%16\.6\\%of molecules overall, ranging from0\.9%0\.9\\%on BACE to38\.6%38\.6\\%on ESOL, while three coincide in only0\.5%0\.5\\%\. The views therefore tend to select distinct substructures, so a rationale spans several granularities rather than repeating one fragment\. A rationale holds4\.54\.5items on average, and76%76\\%of rationales reach the full five\.

Table 12:View composition and cross\-view duplication of the kept rationale items\. The first three columns give the share of items from each view\. FG is the functional group view\. “Dup\. 2” and “Dup\. 3” are the percentages of molecules in which exactly two and exactly three of the five kept items name an identical substructure\.![Refer to caption](https://arxiv.org/html/2608.10480v1/x3.png)Figure 4:Distribution of substructure attributionsat,ja\_\{t,j\}per task, on a logarithmic count axis\. Classification \(blue\) is bounded to the probability range\[−1,1\]\[\-1,1\], regression \(orange\) is in target units\.
#### Attribution distribution\.

Figure[4](https://arxiv.org/html/2608.10480#A2.F4)shows the distribution of signed attributions per task, and Table[13](https://arxiv.org/html/2608.10480#A2.T13)gives the summary statistics\. Classification attributions lie in\[−1,1\]\[\-1,1\], since the source predictor outputs a sigmoid probability, so each attribution is a difference of two probabilities\. Regression attributions instead carry the target unit and are unbounded, which gives them far larger standard deviations\. Every distribution peaks sharply at zero with heavy tails on both sides, so a typical substructure barely moves the source prediction while a few move it sharply\. Across tasks the positive share stays between4040and6464percent, so the predictor assigns both prediction\-raising and prediction\-lowering roles rather than defaulting to one direction\.

Table 13:Summary of the attributionsat,ja\_\{t,j\}\.nnis the number of candidate substructures, “Mean” and “Std” the mean and standard deviation of the attributions, “\+%\+\\%” the share of positive attributions, and “Range” the minimum and maximum\.
#### Serialized example\.

Figure[5](https://arxiv.org/html/2608.10480#A2.F5)shows the first three items of one rationale exactly as stored, for one molecule from the ESOL dataset\. The three shown items span all three views, one Murcko, one BRICS, and one functional group entry\. The Murcko scaffold and the BRICS fragment \(ranks 1 and 2\) are tagged toward lower predicted solubility, while the nitro group \(rank 3\) is tagged toward higher\. Because the task is regression, the effect field names the predicted value, here the log aqueous solubility for ESOL\. For a classification task the field instead names a probability, as intoward higher predicted probability of blood\-brain barrier permeabilityfor BBBP\.

```
Ranked structural rationale items for this molecule.

1. type: Murcko
scaffold SMILES: O=C(Cn1ccnc1)NCc1ccccc1
effect: toward lower predicted log aqueous solubility

2. type: BRICS
fragment SMILES: O=[N+]([O-])c1nccn1
raw BRICS: [9*]n1ccnc1[N+](=O)[O-]
attachments: [9]
effect: toward lower predicted log aqueous solubility

3. type: functional group
name: nitro
effect: toward higher predicted log aqueous solubility
```

Figure 5:The first three items of a serialized rationale as fed to the LLM, for one molecule from the ESOL dataset with SMILESO=C\(Cn1ccnc1\[N\+\]\(=O\)\[O\-\]\)NCc1ccccc1\. The full rationale holds five items\.

### B\.3Number of Rationale Items

To test whether supplying multiple items helps, we compare the full five\-item rationale with a Top\-1 rationale that keeps only the highest\-attribution item\. The two settings are trained identically, differing only in the number of rationale items the model receives during Stage 2 instruction tuning\. Table[14](https://arxiv.org/html/2608.10480#A2.T14)reports both on the headline metric\. The five\-item rationale outperforms the Top\-1 setting on most tasks, with the largest gains on BACE and ClinTox, so the lower\-ranked items supply complementary signal\. HIV is the exception, where the single item does slightly better\.

Table 14:Comparison of the Full and Top\-1 rationales on the headline metric, mean±\\pmstd over three runs\. Top\-1 keeps only the highest\-attribution item, and Full keeps up to five\. The better score per task is in bold\.

## Appendix CAdditional Results

### C\.1Extended Metrics

The main paper reports the ablation on ROC\-AUC and RMSE\. Here we report the corresponding threshold metrics for MR\-MoL and its three channel\-ablated variants, the rationale removed \(w/oRR\), the graph removed \(w/oGG\), and both removed \(w/oG,RG,R\)\. Classification reports the F1 score, the Matthews correlation coefficient \(MCC\), and accuracy at the0\.50\.5threshold onP​\(Yes\)P\(\\mathrm\{Yes\}\); regression reports the mean absolute error \(MAE\) and the mean error \(ME\)\.

WithT​PTP,T​NTN,F​PFP, andF​NFNdenoting the four confusion\-matrix counts, the F1 score is the harmonic mean of precision and recall,

F1=2​T​P2​T​P\+F​P\+F​N,\\mathrm\{F1\}=\\frac\{2\\,TP\}\{2\\,TP\+FP\+FN\},\(9\)and the MCC is

MCC=T​P⋅T​N−F​P⋅F​N\(T​P\+F​P\)​\(T​P\+F​N\)​\(T​N\+F​P\)​\(T​N\+F​N\)\.\\mathrm\{MCC\}=\\frac\{TP\\cdot TN\-FP\\cdot FN\}\{\\sqrt\{\(TP\{\+\}FP\)\(TP\{\+\}FN\)\(TN\{\+\}FP\)\(TN\{\+\}FN\)\}\}\.\(10\)F1 ignoresT​NTN, so it rewards positive\-class detection where accuracy does not\. MCC ranges over\[−1,1\]\[\-1,1\]and uses all four counts, so a model that labels every molecule negative scores0rather than the high accuracy such a model attains, which makes MCC the stricter summary on our imbalanced tasks\. The MAE is the mean of\|y^−y\|\|\\hat\{y\}\-y\|and the ME is the signed mean ofy^−y\\hat\{y\}\-y, whose sign reveals systematic over\- or under\-prediction\.

Table[15](https://arxiv.org/html/2608.10480#A3.T15)and Table[16](https://arxiv.org/html/2608.10480#A3.T16)report these metrics\. These threshold metrics expose effects that ROC\-AUC alone does not reveal\. On ClinTox the two variants without a rationale collapse toward the majority class, with F1 near4848and MCC near zero, which mirrors the large ClinTox gain from the rationale\. Accuracy alone is misleading on the heavily imbalanced tasks, where w/oG,RG,Rreaches high accuracy on ClinTox and HIV yet near\-zero MCC, so it labels almost everything negative\. Across tasks MR\-MoL attains the best F1 on all tasks but BBBP and the best MCC on all but BBBP and SIDER, indicating that the added channels improve the balanced decision quality\.

Table 15:Extended classification metrics, mean±\\pmstd over three runs\. The shaded rows group the results by metric, namely F1 score, MCC, and accuracy\. Within each metric block, the best score per task is in bold\.Table 16:Extended regression metrics, mean±\\pmstd over three runs\. The lowest MAE per task is in bold\.
### C\.2Output Validity

Validity is the fraction of generated answers on the test split that are parseable\. Classification validity is1\.01\.0for every molecular LLM by construction, since the answer is read from the positive and negative token logits\. We therefore report only regression validity, where a model can emit prose without a parseable number\. Table[17](https://arxiv.org/html/2608.10480#A3.T17)shows the result\. MR\-MoL and four baselines parse every output, while three generalists miss a small fraction on ESOL\.

Table 17:Regression output validity\.
### C\.3Generalist Baseline Task Coverage

We evaluate the generalist baselines from their released checkpoints with our own instruction on all eight tasks\. Retraining them on our split would turn each released generalist into a task\-specific model and defeat the purpose of comparing against generalists, so we keep them as released\. As a result, each model has seen a different subset of our eight tasks during its own training, and Table[18](https://arxiv.org/html/2608.10480#A3.T18)makes this coverage explicit\. For every model and task, it marks whether the task was trained \(∙\\bullet\), evaluated zero\-shot \(∘\\circ\), or absent \(−\-\) in that model’s own paper\. We report this coverage so that scores can be read in context\. A trained task may benefit from having appeared in the model’s own training data, whereas a zero\-shot or absent task is evaluated without such exposure\. We therefore avoid drawing conclusions from any single cell, and rely on the overall comparison rather than task\-by\-task claims against these baselines\.

Table 18:Task coverage of the generalist baselines\.∙\\bulletmarks a task in the model’s own instruction data,∘\\circa task held out but reported in the model’s paper, and−\-a task absent from that paper\.