Back to Basics: Improving Molecular Understanding in LLMs via SMILES-Graph Translation

# Back to Basics: Improving Molecular Understanding in LLMs via SMILES–Graph Translation
Source: [https://arxiv.org/html/2607.03007](https://arxiv.org/html/2607.03007)
Wenda Wang Gaoling School of Artificial Intelligence Renmin University of China Beijing, China wangwenda87@ruc\.edu\.cn &Jinjia Feng Gaoling School of Artificial Intelligence Renmin University of China Beijing, China jinjia\_feng@ruc\.edu\.cn &Zhewei Wei Gaoling School of Artificial Intelligence Renmin University of China Beijing, China zhewei@ruc\.edu\.cn

###### Abstract

Recent advances in molecular large language models have led to strong performance on molecular understanding and generation tasks, yet these gains often come without reliable structural grounding\. In particular, existing approaches conflict with the chemistry principle that structure determines function: despite their downstream success, current molecular LLMs perform poorly on basic structure recognition, suggesting that they fail to capture molecular graphs from canonical SMILES\. To remedy this, we propose MolBasic, a structure\-first framework that strengthens structural comprehension via SMILES–Graph translation\. MolBasic is built around a multi\-level structure perception benchmark, where bidirectional SMILES–Graph conversion serves as the core task to align sequential and topological representations\. On top of this foundation, we employ a progressive learning scheme with a standardized Chain\-of\-Thought \(CoT\) to steer models from structure acquisition toward higher\-level molecular reasoning\. Experiments show that MolBasic substantially improves structural understanding and yields robust gains on downstream tasks, including property prediction and objective optimization, supporting our structure\-first paradigm\.

Back to Basics: Improving Molecular Understanding in LLMs via SMILES–Graph Translation

Wenda WangGaoling School of ArtificialIntelligenceRenmin University of ChinaBeijing, Chinawangwenda87@ruc\.edu\.cnJinjia FengGaoling School of ArtificialIntelligenceRenmin University of ChinaBeijing, Chinajinjia\_feng@ruc\.edu\.cnZhewei WeiGaoling School of ArtificialIntelligenceRenmin University of ChinaBeijing, Chinazhewei@ruc\.edu\.cn

## 1Introduction

Understanding molecular characteristics and properties through structural analysis, and designing molecules for specific objectives, has long been a central focus of research in biochemistry\. This effort is of fundamental importance for advancing drug mechanism studies and the design of novel drugs\. With the continuous improvement of Large Language Models \(LLMs\) in learning and reasoning, more studies are exploring how to leverage the domain knowledge embedded in these models for molecular understanding tasks\. Compared to manual analysis, LLMs offer clear advantages in terms of cost and efficiency, making them an ideal tool for this purpose\. Therefore, developing an “AI chemist” capable of structural understanding and reasoning of molecular compounds is crucial\.

![Refer to caption](https://arxiv.org/html/2607.03007v1/latex/figures/radar.png)Figure 1:Preliminary evaluation of generalist and molecular specialist LLMs on basic structure comprehension tasks \(atom counting and bond counting\) using a pilot dataset sampled from PubChemKim et al\. \([2023](https://arxiv.org/html/2607.03007#bib.bib19)\)\. Results reveal that current LLMs fail to capture fundamental graph structure information from SMILES\.![Refer to caption](https://arxiv.org/html/2607.03007v1/latex/figures/overview.png)Figure 2:Comparison between the previous paradigm and our MolBasic framework for molecular understanding\.Generalist LLMs, such as GPTFloridi and Chiriatti \([2020](https://arxiv.org/html/2607.03007#bib.bib9)\); OpenAI \([2023](https://arxiv.org/html/2607.03007#bib.bib27)\), QwenBai et al\. \([2023](https://arxiv.org/html/2607.03007#bib.bib1)\), T5Raffel et al\. \([2020](https://arxiv.org/html/2607.03007#bib.bib30)\), InternLMTeam \([2023](https://arxiv.org/html/2607.03007#bib.bib34)\), VicunaZheng et al\. \([2023](https://arxiv.org/html/2607.03007#bib.bib46)\), and ChatGLMGLM et al\. \([2024](https://arxiv.org/html/2607.03007#bib.bib10)\), acquire knowledge across diverse domains during pre\-training and thus possess certain molecular understanding capabilities\. However, their knowledge is primarily derived from medical or biological texts related to molecules, which grants them basic knowledge retrieval abilities, rather than genuine structural comprehension\. Of course, models with strong reasoning and mathematical capabilities, such as GPT\-5Singh et al\. \([2025](https://arxiv.org/html/2607.03007#bib.bib31)\), can achieve molecular understanding through tool invocation and mathematical computation\. Molecular specialist LLMs, on the other hand, are the mainstream approach for LLM\-based molecular understanding\. Existing research predominantly focuses on two directions: \(i\) designing molecular representations that facilitate model comprehension—for instance, HIGHTChen et al\. \([2024a](https://arxiv.org/html/2607.03007#bib.bib5)\)employs multi\-level molecular representations at atom and motif levels to highlight subgraph\-level features, while InstructMolCao et al\. \([2025](https://arxiv.org/html/2607.03007#bib.bib4)\)and MoMuSu et al\. \([2022](https://arxiv.org/html/2607.03007#bib.bib32)\)use graph encoders to embed molecular graphs into input tokens; \(ii\) designing domain\-specific post\-training tasks on base models—for instance, MolT5Edwards et al\. \([2022](https://arxiv.org/html/2607.03007#bib.bib7)\)replaces corrupted spans between SMILES and text, AtomasZhang et al\. \([2025](https://arxiv.org/html/2607.03007#bib.bib44)\)and MoleculeSTMLiu et al\. \([2023b](https://arxiv.org/html/2607.03007#bib.bib26)\)align SMILES\-text token representations, while GIMLETZhao et al\. \([2023](https://arxiv.org/html/2607.03007#bib.bib45)\)and ChemLLMZhang et al\. \([2024](https://arxiv.org/html/2607.03007#bib.bib43)\)focus on chemical question answering\.

Current molecular LLMs appear to have achieved substantial progress, demonstrating strong performance on common molecular understanding tasks including molecular captioning, property prediction, reaction prediction, and \(conditional\) molecular generation\. However, our preliminary experiments \(shown in[Figure˜1](https://arxiv.org/html/2607.03007#S1.F1)\) reveal that both generalist and specialist models perform surprisingly poorly on fundamental structural perception tasks, such as heavy atom counting and bond counting\. Although prior SMILES parsing methodsJang et al\. \([2025](https://arxiv.org/html/2607.03007#bib.bib17)\); Hao et al\. \([2026](https://arxiv.org/html/2607.03007#bib.bib13)\)attempt to improve chemical understanding by learning local syntactic patterns or substructure\-level features, they still fall short of establishing explicit equivalence between sequential SMILES and the complete 2D molecular graph\. This indicates that current LLMs fail to capture the graph structure information underlying canonical SMILESWeininger \([1988](https://arxiv.org/html/2607.03007#bib.bib37)\)—information that is crucial for subsequent molecular reasoning and represents the most natural and direct cognition that human chemists have of molecules\. These findings suggest that the current understanding paradigm of molecular LLMs contradicts the widely accepted principle in chemistry:structure determines function\. We argue that molecular LLMs should not learn high\-level tasks immediately after simple pre\-training on chemical texts, as this approach lacks logical interpretability\. Instead, they should first focus on establishing a clear understanding of molecular structure before proceeding to downstream reasoning tasks\.

Building on the progress and problems identified above, we propose MolBasic \(BasicStructureIdentification andComprehension\), a back\-to\-basics framework that grounds molecular understanding in fundamental structure comprehension\. With SMILES–Graph mutual conversion as the core task, MolBasic enables LLMs to establish equivalence between sequential and topological molecular representations, building a solid foundation for higher\-level reasoning\. Our contributions can be summarized as follows:

- •Structure\-first Reasoning: We construct a multi\-level Molecular Structure Comprehension benchmark \(MSC\) comprising eight tasks, with SMILES–Graph mutual conversion as the core task, evaluating LLMs’ fundamental 2D graph perception capabilities for reliable downstream reasoning\.
- •Progressive Framework: We propose a staircase reasoning framework that mirrors the cognitive process of human chemists, progressively advancing from structure to property to optimization\. For each stage, we design Chain\-of\-Thought protocols combining powerful LLMs with expert knowledge to enhance structure\-based reasoning capabilities\.
- •Enhanced Reasoning: Through sufficient experiments across structural property prediction, molecular objective optimization, and few\-shot bioactivity prediction tasks, we demonstrate the superior performance and practical applicability of our structure\-first reasoning and stepwise framework\.

[Figure˜2](https://arxiv.org/html/2607.03007#S1.F2)illustrates the comparison between our paradigm and previous methods in molecular understanding, emphasizing key distinctions\.

## 2Related Work

### 2\.1Molecular Understanding

Early approaches employed task\-specific architectures such as GNNsKipf \([2016](https://arxiv.org/html/2607.03007#bib.bib20)\); Veličković et al\. \([2017](https://arxiv.org/html/2607.03007#bib.bib35)\); Ying et al\. \([2021](https://arxiv.org/html/2607.03007#bib.bib41)\)for property prediction and sequence modelsSutskever et al\. \([2014](https://arxiv.org/html/2607.03007#bib.bib33)\); Irwin et al\. \([2022](https://arxiv.org/html/2607.03007#bib.bib15)\)for molecular optimization\. More recently, molecular LLMs have emerged that generalize across tasks through chemical pretrainingZeng et al\. \([2022](https://arxiv.org/html/2607.03007#bib.bib42)\); Zhao et al\. \([2023](https://arxiv.org/html/2607.03007#bib.bib45)\), instruction tuningFang et al\. \([2023](https://arxiv.org/html/2607.03007#bib.bib8)\); Zhang et al\. \([2024](https://arxiv.org/html/2607.03007#bib.bib43)\); Xian et al\. \([2025](https://arxiv.org/html/2607.03007#bib.bib38)\), or multimodal integrationLi et al\. \([2025](https://arxiv.org/html/2607.03007#bib.bib23)\)\. In contrast, our work focuses on post\-training LLMs to enhance their structure\-aware reasoning abilities, strengthening the internal understanding of molecular structure as a foundation for downstream reasoning, rather than aligning chemical text or augmenting retrieval capabilities\.

### 2\.2Reasoning Enhancement for LLMs

Chain\-of\-Thought \(CoT\) reasoning has evolved from a prompting techniqueWei et al\. \([2022](https://arxiv.org/html/2607.03007#bib.bib36)\); Kojima et al\. \([2022](https://arxiv.org/html/2607.03007#bib.bib21)\)to a core capability of advanced LLMs\. OpenAI o1/o3Jaech et al\. \([2024](https://arxiv.org/html/2607.03007#bib.bib16)\)and DeepSeek\-R1Guo et al\. \([2025](https://arxiv.org/html/2607.03007#bib.bib12)\)incorporate long\-chain CoT during training, while GPT\-5Singh et al\. \([2025](https://arxiv.org/html/2607.03007#bib.bib31)\)further integrates an intelligent router that automatically invokes deep reasoning for complex queries\. Progressive learning, drawing from curriculum learningBengio et al\. \([2009](https://arxiv.org/html/2607.03007#bib.bib3)\), enhances reasoning by organizing training according to task dependencies and difficultyXu et al\. \([2020](https://arxiv.org/html/2607.03007#bib.bib39)\), proving effective in mathematical problem solvingLightman et al\. \([2023](https://arxiv.org/html/2607.03007#bib.bib24)\)and multi\-hop reasoningPress et al\. \([2023](https://arxiv.org/html/2607.03007#bib.bib29)\)\. In our work, we combine CoT of strong reasoning models with expert knowledge, proposing the staircase progressive learning framework, starting from basic tasks and following the reasoning process of human chemists\.

## 3Method

![Refer to caption](https://arxiv.org/html/2607.03007v1/latex/figures/opverview_new.png)Figure 3:Overview of the MolBasic framework\.Left:The staircase learning path progresses through three stages: \(1\) structure comprehension at node, edge, and graph levels using the MSC benchmark; \(2\) Structural Property Prediction deriving molecular properties from structural features; \(3\) Objective Molecular Optimization modifying molecules toward desired properties\.Right:The two\-stage training strategy for each task, where Stage 1 \(Answer Learning\) establishes correct output anchors, and Stage 2 \(Reasoning Learning\) incorporates Chain\-of\-Thought supervision distilled from strong reasoning models\.The complete workflow of MolBasic is illustrated in[Figure˜3](https://arxiv.org/html/2607.03007#S3.F3)\. In[Section˜3\.1](https://arxiv.org/html/2607.03007#S3.SS1), we introduce the structure comprehension benchmark designed to enable LLMs to understand molecular structures\.[Section˜3\.2](https://arxiv.org/html/2607.03007#S3.SS2)presents our staircase learning framework for molecular understanding\.[Section˜3\.3](https://arxiv.org/html/2607.03007#S3.SS3)describes the design of reasoning Chain\-of\-Thought for molecular understanding tasks\.[Section˜3\.4](https://arxiv.org/html/2607.03007#S3.SS4)details the comprehensive training strategies\.

### 3\.1Multi\-level Structure Comprehension Benchmark

Canonical SMILES serves as the primary input for molecular LLMs due to its text\-based format that facilitates tokenization\. However, most existing models treat SMILES as plain text sequences, performing global semantic alignment with textual inputs, or at best, matching at the motif or substructure level\. For human chemists, the more critical information encoded in SMILES representations is the molecular graph structure\. Therefore, to enable LLMs to understand and analyze molecular structures more precisely rather than superficially, models must first learn the equivalence between canonical SMILES and 2D molecular topology, and be able to correctly articulate the details of this graph structure, such as node and edge information, which are fundamental to chemical reasoning\.

Based on this motivation, we construct aMulti\-levelStructureComprehension QA benchmark \(MSC\) to equip molecular LLMs with the ability to parse SMILES and understand the SMILES\-graph equivalence\. Our benchmark comprises eight tasks𝒯s​t​r​u​c​t=\{T1,T2,⋯,T8\}\\mathcal\{T\}\_\{struct\}=\\\{T\_\{1\},T\_\{2\},\\cdots,T\_\{8\}\\\}organized across three levels\. At the node level, models are required to identify the number of heavy atoms and the count of specific atom types\. At the edge level, we assess whether models understand molecular connectivity by querying the total number of bonds and counts of specific bond types\. At the graph level, molecular formula conversion serves as a comprehensive evaluation of node and edge understanding, as models must count each atom type and infer the number of hydrogen atoms based on heavy atom connectivity to satisfy chemical valence rules\. Substructure recognition is also essential, as functional groups often serve as the fundamental units of molecular function\. Most importantly, we design bidirectional conversion tasks between canonical SMILES and textual graph adjacency lists, enabling LLMs to directly learn the equivalence between molecular representations and 2D graph structures\. We process the PubChemSTM datasetKim et al\. \([2023](https://arxiv.org/html/2607.03007#bib.bib19)\); Liu et al\. \([2023b](https://arxiv.org/html/2607.03007#bib.bib26)\)and obtain 185,286 valid molecules\. Detailed dataset splits are described in Section 4\.1\.

### 3\.2Staircase Learning Framework

To ensure that higher\-level molecular reasoning is built upon reliable structural understanding, we organize molecular tasks into astaircase learning framework that explicitly models their dependency relationships\.

We view the natural logic of molecular understanding as a DAG\-structured reasoning process rooted in structure comprehension\. Formally, let𝒢=\(𝒱,ℰ\)\\mathcal\{G\}=\(\\mathcal\{V\},\\mathcal\{E\}\)denote a directed acyclic graph \(DAG\) representing the logical relationships among all tasks, where each nodev∈𝒱v\\in\\mathcal\{V\}represents a molecular understanding task, and each directed edge\(vi,vj\)∈ℰ\(v\_\{i\},v\_\{j\}\)\\in\\mathcal\{E\}indicates that taskvjv\_\{j\}depends on the relevant knowledge and reasoning capabilities acquired fromviv\_\{i\}\. The root nodev0v\_\{0\}corresponds to structure comprehension, serving as the foundation for all downstream reasoning\.

Let𝒞\\mathcal\{C\}denote the reasoning capability of the model\. For any taskvvwith prerequisite tasksPa​\(v\)=\{u∈𝒱∣\(u,v\)∈ℰ\}\\text\{Pa\}\(v\)=\\\{u\\in\\mathcal\{V\}\\mid\(u,v\)\\in\\mathcal\{E\}\\\}, the model first acquires capability𝒞Pa​\(v\)\\mathcal\{C\}\_\{\\text\{Pa\}\(v\)\}by learning the prerequisite tasks, then subsequently obtains new capability through learning taskvvas follows:

𝒞v=fv​\(𝒞Pa​\(v\),𝒟v\)\\mathcal\{C\}\_\{v\}=f\_\{v\}\(\\mathcal\{C\}\_\{\\text\{Pa\}\(v\)\},\\mathcal\{D\}\_\{v\}\)\(1\)wherefvf\_\{v\}denotes the learning process on taskvv, and𝒟v\\mathcal\{D\}\_\{v\}is the knowledge data for taskvv\. For the root node,𝒞v0=fv0​\(𝒞0,𝒟structCoT\)\\mathcal\{C\}\_\{v\_\{0\}\}=f\_\{v\_\{0\}\}\(\\mathcal\{C\}\_\{0\},\\mathcal\{D\}^\{\\text\{CoT\}\}\_\{\\text\{struct\}\}\)represents the fundamental capability of structure comprehension, where𝒞0\\mathcal\{C\}\_\{0\}is the initial capability of the base model prior to fine\-tuning\.

In this work, we study linear task sequences𝒫=\(v0,v1,…,vK\)\\mathcal\{P\}=\(v\_\{0\},v\_\{1\},\.\.\.,v\_\{K\}\)where\(vk−1,vk\)∈ℰ\(v\_\{k\-1\},v\_\{k\}\)\\in\\mathcal\{E\}, meaning that the reasoning knowledge required byKKtasks exhibits a linear progressive relationship:

𝒯0≺𝒯1≺⋯≺𝒯k−1≺𝒯k≺⋯≺𝒯K\\mathcal\{T\}\_\{0\}\\prec\\mathcal\{T\}\_\{1\}\\prec\\cdots\\prec\\mathcal\{T\}\_\{k\-1\}\\prec\\mathcal\{T\}\_\{k\}\\prec\\cdots\\prec\\mathcal\{T\}\_\{K\}\(2\)Thekk\-th task is learned by an LLM that has integrated the reasoning capabilities from the precedingk−1k\-1tasks\.

In this paper, we demonstrate the learning and reasoning process of our staircase learning framework through the path𝒫:vstruct→vprop→vopt\\mathcal\{P\}:v\_\{\\text\{struct\}\}\\rightarrow v\_\{\\text\{prop\}\}\\rightarrow v\_\{\\text\{opt\}\}, which chains together a series of representative molecular understanding tasks:

𝒞struct\\displaystyle\\mathcal\{C\}\_\{\\text\{struct\}\}=fstruct​\(𝒞0,𝒟struct\)\\displaystyle=f\_\{\\text\{struct\}\}\(\\mathcal\{C\}\_\{0\},\\mathcal\{D\}\_\{\\text\{struct\}\}\)\(3\)𝒞prop\\displaystyle\\mathcal\{C\}\_\{\\text\{prop\}\}=fprop​\(𝒞struct,𝒟prop\)\\displaystyle=f\_\{\\text\{prop\}\}\(\\mathcal\{C\}\_\{\\text\{struct\}\},\\mathcal\{D\}\_\{\\text\{prop\}\}\)𝒞opt\\displaystyle\\mathcal\{C\}\_\{\\text\{opt\}\}=fopt​\(𝒞prop,𝒟opt\)\\displaystyle=f\_\{\\text\{opt\}\}\(\\mathcal\{C\}\_\{\\text\{prop\}\},\\mathcal\{D\}\_\{\\text\{opt\}\}\)This path aligns with the cognitive process of human chemists and reflects the fundamental chemistry principle thatstructure determines function\.

### 3\.3Chain\-of\-Thought Construction

To improve the accuracy and interpretability of molecular reasoning, we introduce explicit Chain\-of\-Thought supervision thatconstrains intermediate structural analysis during training\.In traditional molecular understanding task training, models typically learn only the mapping from input moleculeSSto final predictionaa, i\.e\., optimizing the probabilityPθ​\(a∣S\)P\_\{\\theta\}\(a\\mid S\)\. Although models can directly learn input\-to\-output mappings, in complex tasks, training without explicit reasoning supervision leads to reduced model accuracy and lack of interpretability\. Therefore, we enhance the knowledge QA dataset by introducing Chain\-of\-Thought \(CoT\) distilled from strong reasoning models\. Specifically, we distill reasoning trajectories from GPT\-5Singh et al\. \([2025](https://arxiv.org/html/2607.03007#bib.bib31)\)to construct the CoT supervision\.

By incorporating the reasoning trajectoryzzas a supervisory signal, we extend the training objective to jointly model the conditional distribution over molecular representation, reasoning process, and results:

Pθ​\(a,z∣S,q\)=Pθ​\(z∣S,q\)⋅Pθ​\(a∣S,z,q\)P\_\{\\theta\}\(a,z\\mid S,q\)=P\_\{\\theta\}\(z\\mid S,q\)\\cdot P\_\{\\theta\}\(a\\mid S,z,q\)\(4\)Compared to supervision through final answers alone, this approach effectively constrains the hypothesis space and reduces ambiguity in the reasoning process, thereby improving reasoning accuracy and stability\. Joint supervision ensures that the loss function considers not only the final output but also the prediction error of the reasoning process\.

We construct CoTs by distilling reasoning trajectories from strong reasoning models and applying manual review to standardize reasoning steps and remove erroneous trajectories\. This ensures reliable, reproducible reasoning\. We highlight two key aspects of the resulting CoT design:

\(i\) Deterministic Intermediate Results for Regression Tasks\. For structure comprehension and property prediction, we design CoTs with verifiable intermediate results\. For instance, atom counting explicitly parses SMILES into atoms and records counts before aggregation; property prediction first identifies functional groups and structural features, then derives values from these intermediates\. This makes each step deterministic and checkable against ground truth\.

\(ii\) Explicit Analysis and Modification Explanation for Optimization Tasks\. For molecular optimization, the CoT follows an “analysis–strategy–verification” pipeline: it analyzes the original structure and properties, identifies groups affecting the target property and proposes modifications, and verifies the optimized molecule achieves the desired improvement\. This encourages learning transferable modification principles rather than memorizing specific edits\.

After constructing the molecular reasoning CoT as described above, we merge it into the answer and use𝒟vCoT\\mathcal\{D\}^\{\\text\{CoT\}\}\_\{v\}as shown in[Equation˜5](https://arxiv.org/html/2607.03007#S3.E5)as the knowledge dataset for training:

a^=\(z,a\),𝒟vCoT=\{\(Sj,\(qj,a^j\)\)\}j=1Nv\\hat\{a\}=\(z,a\),\\quad\\mathcal\{D\}^\{\\text\{CoT\}\}\_\{v\}=\\\{\(S^\{j\},\(q^\{j\},\\hat\{a\}^\{j\}\)\)\\\}\_\{j=1\}^\{N\_\{v\}\}\(5\)By augmenting training labels with Chain\-of\-Thought explanations from strong reasoning models, we transform molecular reasoning from a black\-box mapping into a structured prediction problem with explicit supervision\.

Table 1:Performance comparison on multi\-level structure comprehension benchmark\. For MAE, lower is better \(↓\\downarrow\); for ACC, higher is better \(↑\\uparrow\)\.Bestandsecond bestresults are highlighted\.LevelTasksMetricsVicunaChatGLMT5Qwen3\-8BInternLM3MolT5ChemLLMMolBasicNodeHeavy Atom CountingMAE↓\\downarrow28\.80428\.22833\.53511\.89319\.75435\.046\.6961\.223ACC↑\\uparrow1\.21%0\.73%0\.22%6\.09%4\.75%0\.4%14\.81%42\.44%Element\-specific CountingMAE↓\\downarrow8\.5119\.1279\.8822\.9823\.9029\.8333\.7630\.339ACC↑\\uparrow11\.49%12\.22%0\.13%51\.86%40\.98%2\.5%42\.1%73\.01%EdgeTotal Bond CountingMAE↓\\downarrow29\.8526\.50635\.1918\.35230\.26535\.75510\.5811\.158ACC↑\\uparrow1\.08%1\.86%0\.09%3\.07%1\.12%0\.4%3\.41%49%Specific Bond CountingMAE↓\\downarrow9\.810\.55710\.2277\.6748\.1059\.28736\.670\.383ACC↑\\uparrow22\.58%14\.12%2\.81%45\.81%26\.34%3\.3%5\.76%82\.9%GraphFormula ConvertACC↑\\uparrow0%0%0%0%0%0%0%44\.04%Substructure RecognitionACC↑\\uparrow48\.75%46\.03%0\.13%72\.67%66\.23%19\.8%10\.71%98\.14%SMILES to GraphACC↑\\uparrow0%0%0%0%0%0%0\.04%94\.65%Graph to SMILESACC↑\\uparrow0%0%0%0%0%0%0%85\.66%

### 3\.4Training Strategy

We present the multi\-stage training of MolBasic based on LoRA fine\-tuning\. First, following the staircase learning framework proposed in[Section˜3\.2](https://arxiv.org/html/2607.03007#S3.SS2), we sequentially train each task along the task path𝒫:vstruct→vprop→vopt\\mathcal\{P\}:v\_\{\\text\{struct\}\}\\rightarrow v\_\{\\text\{prop\}\}\\rightarrow v\_\{\\text\{opt\}\}on a base model \(Qwen3\-8BYang et al\. \([2025](https://arxiv.org/html/2607.03007#bib.bib40)\)in this work\)\.

For training each individual task, we follow the easy\-to\-hard curriculum learning principle and divide the training into two stages\. Stage 1 focuses on learning the answeraaonly, dedicated to domain knowledge acquisition\. Stage 2 learns the complete response containing CoT, dedicated to refining reasoning capabilities\. This two\-stage training first establishes correct answer anchors to constrain the reasoning process towards the correct direction, then forms standardized and formalized reasoning processes that further enhance accuracy and interpretability through explicit and verifiable intermediate steps\.

Both stages employ LoRA\-based next token prediction with cross\-entropy loss as shown in[Equation˜6](https://arxiv.org/html/2607.03007#S3.E6)and \([7](https://arxiv.org/html/2607.03007#S3.E7)\)\. The training algorithm of MolBasic is presented in Appendix[A](https://arxiv.org/html/2607.03007#A1)\.

ℒstage1\\displaystyle\\mathcal\{L\}\_\{\\text\{stage1\}\}=−∑t=1\|a\|log⁡Pθ​\(at∣S,q,a<t\)\\displaystyle=\-\\sum\_\{t=1\}^\{\|a\|\}\\log P\_\{\\theta\}\(a\_\{t\}\\mid S,q,a\_\{<t\}\)\(6\)ℒstage2\\displaystyle\\mathcal\{L\}\_\{\\text\{stage2\}\}=−∑t=1\|a^\|log⁡Pθ​\(a^t∣S,q,a^<t\)\\displaystyle=\-\\sum\_\{t=1\}^\{\|\\hat\{a\}\|\}\\log P\_\{\\theta\}\(\\hat\{a\}\_\{t\}\\mid S,q,\\hat\{a\}\_\{<t\}\)\(7\)

## 4Experiments

In this section, we present experimental results of MolBasic\. We first describe training and evaluation details on our Multi\-level structure comprehension \(MSC\) benchmark in[Section˜4\.1](https://arxiv.org/html/2607.03007#S4.SS1)\. Then in[Section˜4\.2](https://arxiv.org/html/2607.03007#S4.SS2), we demonstrate how the acquired structural perception capabilities transfer to downstream molecular understanding tasks through progressive reasoning, including structural property prediction \([Section˜4\.2\.1](https://arxiv.org/html/2607.03007#S4.SS2.SSS1)\), objective molecular optimization \([Section˜4\.2\.2](https://arxiv.org/html/2607.03007#S4.SS2.SSS2)\), and few\-shot transfer to bioactivity prediction tasks \([Appendix˜E](https://arxiv.org/html/2607.03007#A5)\)\. Training parameters and training/inference costs for each stage are listed in[Table˜7](https://arxiv.org/html/2607.03007#A2.T7)\([Appendix˜B](https://arxiv.org/html/2607.03007#A2)\)\.

### 4\.1Structure Comprehension

Datasets and Baselines:As introduced in[Section˜3\.1](https://arxiv.org/html/2607.03007#S3.SS1), our MSC benchmark uses the PubChemSTM dataset reported by MoleculeSTMLiu et al\. \([2023b](https://arxiv.org/html/2607.03007#bib.bib26)\), following the same preprocessing pipeline with raw data sourced from PubChemKim et al\. \([2023](https://arxiv.org/html/2607.03007#bib.bib19)\)\. In Stage 1, weights are assigned based on task importance, with higher weights given to the two SMILES\-graph conversion tasks\. In Stage 2, weights are adjusted according to Stage 1 performance, increasing weights for underperforming tasks\. Detailed training set construction strategies for both stages are provided in Appendix[B](https://arxiv.org/html/2607.03007#A2)\. For efficient and unbiased evaluation, we sample 2,316 instances from the full test set to construct a benchmark that preserves the original distribution\. We compare against mainstream general\-purpose and molecular LLMs with similar parameter scales, including Qwen3\-8BYang et al\. \([2025](https://arxiv.org/html/2607.03007#bib.bib40)\), InternLM3\-8BTeam \([2023](https://arxiv.org/html/2607.03007#bib.bib34)\), T5\-LargeRaffel et al\. \([2020](https://arxiv.org/html/2607.03007#bib.bib30)\), Vicuna\-7BZheng et al\. \([2023](https://arxiv.org/html/2607.03007#bib.bib46)\), ChatGLM3\-6BGLM et al\. \([2024](https://arxiv.org/html/2607.03007#bib.bib10)\), MolT5\-LargeEdwards et al\. \([2022](https://arxiv.org/html/2607.03007#bib.bib7)\), and ChemLLM\-7BZhang et al\. \([2024](https://arxiv.org/html/2607.03007#bib.bib43)\)\.

Table 2:Results on structural property prediction\.Bestandsecond bestresults are highlighted among LLM\-based methods\. “–” indicates unreasonable or invalid outputs, implying failure in property prediction\.MethodMWLogPTPSAHBDHBARBQEDPear\.↑\\uparrowMAE↓\\downarrowPear\.↑\\uparrowMAE↓\\downarrowPear\.↑\\uparrowMAE↓\\downarrowPear\.↑\\uparrowMAE↓\\downarrowPear\.↑\\uparrowMAE↓\\downarrowPear\.↑\\uparrowMAE↓\\downarrowPear\.↑\\uparrowMAE↓\\downarrowChemception \(CNN\)0\.8744\.750\.621\.230\.8414\.450\.810\.580\.840\.930\.731\.91\-0\.0040\.40Chemprop \(GNN\)0\.991\.630\.980\.160\.990\.910\.990\.030\.990\.110\.990\.150\.890\.07Qwen\-VL\-7B0\.7858\.070\.141\.480\.1982\.83\-0\.021\.420\.054\.910\.196\.87––InternVL\-v1\.5\-20B0\.5983\.600\.042\.370\.2928\.730\.032\.240\.222\.410\.044\.820\.0030\.40LLaVA\-v1\.5\-7B0\.36115\.70\-0\.0031\.610\.0199\.140\.0043\.740\.042\.930\.033\.85––ChemVLM\-8B0\.8456\.940\.381\.680\.2653\.660\.491\.350\.324\.580\.105\.56\-0\.0030\.37ChemMLLM\-7B0\.9716\.170\.920\.520\.976\.060\.940\.130\.940\.440\.941\.620\.910\.06ChemMLLM\-34B0\.9811\.570\.930\.430\.983\.540\.960\.110\.960\.260\.970\.750\.930\.05MolBasic0\.99613\.980\.970\.280\.992\.710\.990\.050\.990\.160\.980\.360\.940\.06

Results:[Table˜1](https://arxiv.org/html/2607.03007#S3.T1)presents the performance of various models on multi\-level structure comprehension tasks\. MolBasic achieves substantial improvements across all tasks, with the most striking gains on bidirectional SMILES\-graph conversion—a fundamental yet previously overlooked capability where all existing LLMs universally fail with near\-zero accuracy\. Since the graph adjacency list format is uncommon, we provide a simple example in the instruction as an output template \(applied fairly to all models\)\. Conversion accuracy is verified using RDKitLandrum \([2013](https://arxiv.org/html/2607.03007#bib.bib22)\)by checking exact molecular match against the ground truth\.

### 4\.2Downstream Molecular Understanding

After establishing structure comprehension, we evaluate MolBasic on representative high\-level molecular tasks\. Beyond the progressive relationship in[Section˜3\.2](https://arxiv.org/html/2607.03007#S3.SS2), task selection prioritizes structure\-based reasoning over memorization or knowledge retrieval\. We focus on property prediction and molecular optimization, which require structural analysis rather than text generation tasks\.

#### 4\.2\.1Structural Property Prediction

Structural properties refer to molecular attributes determined by the molecular structure\. They can be predicted by analyzing key atoms, bonds, and functional groups, including MW, LogP, TPSA, HBD, HBA, RB, and QED\. The detailed definitions of these properties are provided in Appendix[C](https://arxiv.org/html/2607.03007#A3)\.

We fine\-tune and evaluate our model on 100K molecules sampled from the PubChem dataset\. Baseline methods include task\-specific deep learning models based on traditional architectures \(ChemceptionGoh et al\. \([2017](https://arxiv.org/html/2607.03007#bib.bib11)\), ChemPropHeid et al\. \([2023](https://arxiv.org/html/2607.03007#bib.bib14)\)\), as well as representative \(multi\-modal\) large language models, including Qwen\-VLBai et al\. \([2025](https://arxiv.org/html/2607.03007#bib.bib2)\), InternVLChen et al\. \([2024b](https://arxiv.org/html/2607.03007#bib.bib6)\), LLaVALiu et al\. \([2023a](https://arxiv.org/html/2607.03007#bib.bib25)\), and ChemVLMLi et al\. \([2025](https://arxiv.org/html/2607.03007#bib.bib23)\)\.[Table˜2](https://arxiv.org/html/2607.03007#S4.T2)reports the prediction errors and Pearson correlation coefficients between the predicted values and ground\-truth labels for structural property prediction tasks\.

As shown in[Table˜2](https://arxiv.org/html/2607.03007#S4.T2), MolBasic consistently achieves the best performance across almost all properties, significantly outperforming other LLM\-based methods, while achieving performance comparable to task\-specific models\. These results demonstrate that once an LLM is equipped with strong structure comprehension ability, it can predict structure\-related molecular properties more accurately and reliably\.

#### 4\.2\.2Objective Molecular Optimization

Based on the model’s understanding of structural properties learned in the previous stage, we equip MolBasic with the ability to optimize molecular structures toward a desired objective, such as improving drug\-likeness or enhancing solubility\. Specifically, we take the task of increasing LogP as a representative example, and use 175K molecules from the TDC datasetJin et al\. \([2018](https://arxiv.org/html/2607.03007#bib.bib18)\)\.

Table 3:Performance comparison on molecular optimization\. For all 3 metrics, higher is better \(↑\\uparrow\)\.Bestandsecond bestresults are highlighted among LLM\-based methods\.ModelIncreased LogP \(↑\\uparrow\)Diversity \(↑\\uparrow\)Validity \(↑\\uparrow\)Seq2Seq1\.950\.7980\.5%ChemFormer3\.030\.85100%Qwen\-VL\-7B1\.500\.954\.0%InternVL\-v1\.5\-20B0\.770\.9048\.0%LLaVA\-v1\.5\-7B1\.720\.9637\.5%GPT\-4o1\.970\.8699\.0%ChemVLM\-8B0\.670\.8792\.5%MolBasic2\.270\.8698\.7%

Table 4:Ablation studies\. We examine four ablation settings against the full MolBasic pipeline to isolate the contribution of each design choice\. “Skipped” indicates that the corresponding task stage is bypassed under that setting and thus not evaluated\.SettingStructure ComprehensionProperty Prediction \(MAE↓\\downarrow\)Molecular OptimizationAtomMAE↓\\downarrowAtomACC↑\\uparrowBondMAE↓\\downarrowBondACC↑\\uparrowForm\.ACC↑\\uparrowMWLogPTPSAHBDHBARBQEDΔ\\DeltaLogP↑\\uparrowDiv\.↑\\uparrowVal\.↑\\uparrowMolBasic1\.2242\.41\.1649\.044\.013\.980\.282\.710\.050\.160\.360\.062\.270\.8698\.7%\(a\) Is structure comprehension a necessary foundation?w/o StructSkipped42\.290\.6414\.420\.621\.190\.640\.101\.800\.8697\.6%Direct OptSkippedSkipped1\.410\.8594\.2%\(b\) Does the staircase ordering matter, beyond total data volume?Mixed SFTSkipped122\.871\.8535\.490\.891\.872\.820\.250\.980\.8795\.3%\(c\) What role does CoT supervision play?w/o CoT1\.5116\.41\.6111\.222\.135\.670\.5311\.120\.590\.710\.530\.082\.050\.8599\.5%

For baseline methods, we include traditional sequence\-based models, Seq2SeqSutskever et al\. \([2014](https://arxiv.org/html/2607.03007#bib.bib33)\)and ChemFormerIrwin et al\. \([2022](https://arxiv.org/html/2607.03007#bib.bib15)\), as well as representative LLM\-based approaches, incorporating GPT\-4oOpenAI \([2024](https://arxiv.org/html/2607.03007#bib.bib28)\)for comparison\.[Table˜3](https://arxiv.org/html/2607.03007#S4.T3)reports the performance of our MolBasic on the molecular optimization task with LogP increase as the objective, compared to other methods\.

As shown in[Table˜3](https://arxiv.org/html/2607.03007#S4.T3), MolBasic achieves the best LogP improvement among LLM\-based models while maintaining high validity and diversity\. As the highest\-level task in our stepwise reasoning framework, we present example responses with explanation for the optimization task in Appendix[L](https://arxiv.org/html/2607.03007#A12), demonstrating the interpretability of our approach\.

### 4\.3Ablation Study

In this section, we validate the key design choices of MolBasic through ablation experiments organized around three questions: \(i\) whether structure comprehension serves as a necessary foundation for downstream tasks, \(ii\) whether the staircase learning order itself drives performance gains beyond simply increasing training data, and \(iii\) what role Chain\-of\-Thought supervision plays in the framework\. All results are consolidated in Table[4](https://arxiv.org/html/2607.03007#S4.T4)\.

\(a\) Is structure comprehension a necessary foundation?A central claim of MolBasic is that structural perception acquired in the first stage provides essential capabilities for all subsequent tasks\. To test this, we construct two settings that progressively strip away prerequisite stages:w/o Struct, which skips structure comprehension and directly trains the base model on property prediction followed by optimization, andDirect Opt, which trains the base model on optimization alone without any prerequisite\. As shown in Table[4](https://arxiv.org/html/2607.03007#S4.T4)\(a\), performance degrades monotonically as more stages are removed\. The degradation from MolBasic to w/o Struct demonstrates that even when the model receives the same downstream fine\-tuning, the absence of prior structural understanding leads to clearly inferior property prediction\. The further decline from w/o Struct to Direct Opt confirms that each stage in the staircase contributes foundational capabilities that cannot be compensated by task\-specific training alone\.

\(b\) Does the staircase ordering matter?Although the results in Table[4](https://arxiv.org/html/2607.03007#S4.T4)\(a\) already show that skipping prerequisite reasoning stages in the staircase process negatively affects performance, we design theMixed SFTsetting to disentangle this effect from the total training data volume\. We merge the training sets of both downstream tasks \(property prediction and molecular optimization\) and apply the same two\-stage curriculum learning strategy \(answer learning followed by reasoning learning\), ensuring that the total number of supervised signals is matched\. The only difference is the removal of progressive task ordering\. As Table[4](https://arxiv.org/html/2607.03007#S4.T4)\(b\) shows, Mixed SFT leads to reduced performance across metrics, confirming that staircase ordering is essential—when reasoning processes from different task levels are learned simultaneously, they interfere with each other, preventing the model from establishing the structured knowledge hierarchy\.

\(c\) What role does CoT supervision play?Finally, we examine the contribution of Chain\-of\-Thought reasoning supervision by comparing full MolBasic with aw/o CoTvariant that preserves progressive training order but removes Stage 2 reasoning learning at all stages, retaining only answer\-level supervision from Stage 1\. As Table[4](https://arxiv.org/html/2607.03007#S4.T4)\(c\) shows, removing CoT supervision leads to consistent performance drops across structure comprehension and downstream tasks\. Structure comprehension tasks, such as atom counting and formula conversion, require multi\-step reasoning, and CoT provides explicit guidance for these intermediate steps\. Consequently, the absence of CoT reduces the model’s reasoning capabilities, since it lacks explicit supervision on intermediate reasoning steps\. This results in lower accuracy on property prediction tasks and decreased effectiveness in molecular optimization, highlighting the importance of CoT in guiding multi\-step molecular reasoning\. We note that GPT\-5, serving as the source of CoT, was also evaluated directly on these tasks \(see Appendix\); results show that MolBasic becomes competitive with GPT\-5 as the reasoning process unfolds\.

Additional evaluations and analyses, covering external benchmarks, GPT\-5 comparisons, and statistical analysis, are provided in Appendices[D](https://arxiv.org/html/2607.03007#A4)–[H](https://arxiv.org/html/2607.03007#A8)\.

## 5Conclusion

In this work, we revisit a fundamental but overlooked problem in molecular LLMs: the lack of basic structure comprehension\. We demonstrate that without accurately perceiving molecular graphs, higher\-level reasoning tasks are inherently unreliable\. To address this, we equip LLMs with explicit structural comprehension ability, and then introduce Chain\-of\-Thought supervision and a progressive learning framework to transfer this capability to downstream tasks\. Extensive experiments confirm that restoring this basic structural understanding is the key to achieving accurate, interpretable, and generalizable molecular reasoning\.

## Limitations

One limitation of the current MolBasic framework lies in its staircase reasoning design\. In this work, molecular reasoning is modeled as a chain\-like process that progresses from structure comprehension to property prediction and then to molecular optimization\. As downstream tasks become farther from the initial structure comprehension stage, the direct influence of structural perception may gradually weaken\. This is reflected in our results, where the improvement on molecular optimization is less pronounced than that on property prediction\. A promising future direction is to design a more flexible reasoning process that more closely resembles how human chemists analyze molecules, allowing LLMs to explicitly use molecular structure as the foundation for each molecular understanding task\. However, this is non\-trivial: simply adding structure comprehension data to every downstream task is unlikely to be sufficient, as our ablation study shows that directly mixing tasks can introduce interference\. Designing a principled structure\-grounded reasoning framework for diverse molecular tasks remains an important direction for future work\.

## Ethical Considerations

This work focuses on AI\-assisted molecular understanding and beneficial molecular optimization\. In our experiments, optimization objectives are limited to standard drug\-relevant properties, such as improving LogP, QED, and synthetic accessibility, which are commonly used to assess molecular quality and developability\. We do not optimize toward toxic, harmful, or unsafe molecular objectives\. During CoT post\-training, we also observed that the base model retained safety\-aware behavior for potentially hazardous molecules, sometimes refusing to provide related information\. This suggests that the safety alignment of the base model can serve as an additional safeguard against unsafe molecular generation\. For future real\-world applications, generated molecules should still be subject to standard safety screening, toxicity filtering, and human expert review before deployment\.

## References

- Bai et al\. \(2023\)Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, and 1 others\. 2023\.Qwen technical report\.*arXiv preprint arXiv:2309\.16609*\.
- Bai et al\. \(2025\)Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, and 1 others\. 2025\.Qwen2\. 5\-vl technical report\.*arXiv preprint arXiv:2502\.13923*\.
- Bengio et al\. \(2009\)Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston\. 2009\.Curriculum learning\.In*Proceedings of the 26th annual international conference on machine learning*, pages 41–48\.
- Cao et al\. \(2025\)He Cao, Zijing Liu, Xingyu Lu, Yuan Yao, and Yu Li\. 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*, pages 354–379\.
- Chen et al\. \(2024a\)Yongqiang Chen, Quanming Yao, Juzheng Zhang, James Cheng, and Yatao Bian\. 2024a\.Improving molecule\-language alignment with hierarchical graph tokenization\.
- Chen et al\. \(2024b\)Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, and 1 others\. 2024b\.Internvl: Scaling up vision foundation models and aligning for generic visual\-linguistic tasks\.In*Proceedings of the IEEE/CVF conference on computer vision and pattern recognition*, pages 24185–24198\.
- Edwards et al\. \(2022\)Carl Edwards, Tuan Lai, Kevin Ros, Garrett Honke, Kyunghyun Cho, and Heng Ji\. 2022\.Translation between molecules and natural language\.*arXiv preprint arXiv:2204\.11817*\.
- Fang et al\. \(2023\)Yin Fang, Xiaozhuan Liang, Ningyu Zhang, Kangwei Liu, Rui Huang, Zhuo Chen, Xiaohui Fan, and Huajun Chen\. 2023\.Mol\-instructions: A large\-scale biomolecular instruction dataset for large language models\.*arXiv preprint arXiv:2306\.08018*\.
- Floridi and Chiriatti \(2020\)Luciano Floridi and Massimo Chiriatti\. 2020\.Gpt\-3: Its nature, scope, limits, and consequences\.*Minds and machines*, 30\(4\):681–694\.
- GLM et al\. \(2024\)Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, and 1 others\. 2024\.Chatglm: A family of large language models from glm\-130b to glm\-4 all tools\.*arXiv preprint arXiv:2406\.12793*\.
- Goh et al\. \(2017\)Garrett B Goh, Charles Siegel, Abhinav Vishnu, Nathan O Hodas, and Nathan Baker\. 2017\.Chemception: a deep neural network with minimal chemistry knowledge matches the performance of expert\-developed qsar/qspr models\.*arXiv preprint arXiv:1706\.06689*\.
- Guo et al\. \(2025\)Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others\. 2025\.Deepseek\-r1: Incentivizing reasoning capability in llms via reinforcement learning\.*arXiv preprint arXiv:2501\.12948*\.
- Hao et al\. \(2026\)Li Hao, He Cao, Bin Feng, Daniel Shao, Robert Tang, Zhiyuan Yan, Yonghong Tian, Li Yuan, and Yu Li\. 2026\.Beyond chemical qa: Evaluating llm’s chemical reasoning with modular chemical operations\.*Advances in Neural Information Processing Systems*, 38\.
- Heid et al\. \(2023\)Esther Heid, Kevin P Greenman, Yunsie Chung, Shih\-Cheng Li, David E Graff, Florence H Vermeire, Haoyang Wu, William H Green, and Charles J McGill\. 2023\.Chemprop: a machine learning package for chemical property prediction\.*Journal of Chemical Information and Modeling*, 64\(1\):9–17\.
- Irwin et al\. \(2022\)Ross Irwin, Spyridon Dimitriadis, Jiazhen He, and Esben Jannik Bjerrum\. 2022\.Chemformer: a pre\-trained transformer for computational chemistry\.*Machine Learning: Science and Technology*, 3\(1\):015022\.
- Jaech et al\. \(2024\)Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El\-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, and 1 others\. 2024\.Openai o1 system card\.*arXiv preprint arXiv:2412\.16720*\.
- Jang et al\. \(2025\)Yunhui Jang, Jaehyung Kim, and Sungsoo Ahn\. 2025\.Improving chemical understanding of llms via smiles parsing\.In*Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing*, pages 15694–15709\.
- Jin et al\. \(2018\)Wengong Jin, Kevin Yang, Regina Barzilay, and Tommi Jaakkola\. 2018\.Learning multimodal graph\-to\-graph translation for molecular optimization\.*arXiv preprint arXiv:1812\.01070*\.
- Kim et al\. \(2023\)Sunghwan Kim, Jie Chen, Tiejun Cheng, Asta Gindulyte, Jia He, Siqian He, Qingliang Li, Benjamin A Shoemaker, Paul A Thiessen, Bo Yu, and 1 others\. 2023\.Pubchem 2023 update\.*Nucleic acids research*, 51\(D1\):D1373–D1380\.
- Kipf \(2016\)TN Kipf\. 2016\.Semi\-supervised classification with graph convolutional networks\.*arXiv preprint arXiv:1609\.02907*\.
- Kojima et al\. \(2022\)Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa\. 2022\.Large language models are zero\-shot reasoners\.*Advances in neural information processing systems*, 35:22199–22213\.
- Landrum \(2013\)Greg Landrum\. 2013\.Rdkit documentation\.*Release*, 1\(1\-79\):4\.
- Li et al\. \(2025\)Junxian Li, Di Zhang, Xunzhi Wang, Zeying Hao, Jingdi Lei, Qian Tan, Cai Zhou, Wei Liu, Yaotian Yang, Xinrui Xiong, and 1 others\. 2025\.Chemvlm: Exploring the power of multimodal large language models in chemistry area\.In*Proceedings of the AAAI Conference on Artificial Intelligence*, volume 39, pages 415–423\.
- Lightman et al\. \(2023\)Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe\. 2023\.Let’s verify step by step\.In*The Twelfth International Conference on Learning Representations*\.
- Liu et al\. \(2023a\)Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee\. 2023a\.Visual instruction tuning\.In*NeurIPS*\.
- Liu et al\. \(2023b\)Shengchao Liu, Weili Nie, Chengpeng Wang, Jiarui Lu, Zhuoran Qiao, Ling Liu, Jian Tang, Chaowei Xiao, and Animashree Anandkumar\. 2023b\.Multi\-modal molecule structure–text model for text\-based retrieval and editing\.*Nature Machine Intelligence*, 5\(12\):1447–1457\.
- OpenAI \(2023\)OpenAI\. 2023\.Gpt\-4 technical report\.*arXiv preprint arXiv:2303\.08774*\.
- OpenAI \(2024\)OpenAI\. 2024\.[Introducing gpt\-4o](https://openai.com/blog/introducing-gpt-4o)\.
- Press et al\. \(2023\)Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A Smith, and Mike Lewis\. 2023\.Measuring and narrowing the compositionality gap in language models\.In*Findings of the Association for Computational Linguistics: EMNLP 2023*, pages 5687–5711\.
- Raffel et al\. \(2020\)Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J\. Liu\. 2020\.Exploring the limits of transfer learning with a unified text\-to\-text transformer\.*Journal of Machine Learning Research*, 21\(140\):1–67\.
- Singh et al\. \(2025\)Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El\-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, and 1 others\. 2025\.Openai gpt\-5 system card\.*arXiv preprint arXiv:2601\.03267*\.
- Su et al\. \(2022\)Bing Su, Dazhao Du, Zhao Yang, Yujie Zhou, Jiangmeng Li, Anyi Rao, Hao Sun, Zhiwu Lu, and Ji\-Rong Wen\. 2022\.A molecular multimodal foundation model associating molecule graphs with natural language\.*arXiv preprint arXiv:2209\.05481*\.
- Sutskever et al\. \(2014\)Ilya Sutskever, Oriol Vinyals, and Quoc V Le\. 2014\.Sequence to sequence learning with neural networks\.*Advances in neural information processing systems*, 27\.
- Team \(2023\)InternLM Team\. 2023\.Internlm: A multilingual language model with progressively enhanced capabilities\.[https://github\.com/InternLM/InternLM](https://github.com/InternLM/InternLM)\.
- Veličković et al\. \(2017\)Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio\. 2017\.Graph attention networks\.*arXiv preprint arXiv:1710\.10903*\.
- Wei et al\. \(2022\)Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others\. 2022\.Chain\-of\-thought prompting elicits reasoning in large language models\.*Advances in neural information processing systems*, 35:24824–24837\.
- Weininger \(1988\)David Weininger\. 1988\.Smiles, a chemical language and information system\. 1\. introduction to methodology and encoding rules\.*Journal of chemical information and computer sciences*, 28\(1\):31–36\.
- Xian et al\. \(2025\)Ziting Xian, Jiawei Gu, Lingbo Li, and Shangsong Liang\. 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 \(Volume 1: Long Papers\)*, pages 15513–15531\.
- Xu et al\. \(2020\)Benfeng Xu, Licheng Zhang, Zhendong Mao, Quan Wang, Hongtao Xie, and Yongdong Zhang\. 2020\.Curriculum learning for natural language understanding\.In*Proceedings of the 58th annual meeting of the association for computational linguistics*, pages 6095–6104\.
- Yang et al\. \(2025\)An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others\. 2025\.Qwen3 technical report\.*arXiv preprint arXiv:2505\.09388*\.
- Ying et al\. \(2021\)Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, and Tie\-Yan Liu\. 2021\.Do transformers really perform badly for graph representation?*Advances in neural information processing systems*, 34:28877–28888\.
- Zeng et al\. \(2022\)Zheni Zeng, Yuan Yao, Zhiyuan Liu, and Maosong Sun\. 2022\.A deep\-learning system bridging molecule structure and biomedical text with comprehension comparable to human professionals\.*Nature communications*, 13\(1\):862\.
- Zhang et al\. \(2024\)Di Zhang, Wei Liu, Qian Tan, Jingdan Chen, Hang Yan, Yuliang Yan, Jiatong Li, Weiran Huang, Xiangyu Yue, Wanli Ouyang, and 1 others\. 2024\.Chemllm: A chemical large language model\.*arXiv preprint arXiv:2402\.06852*\.
- Zhang et al\. \(2025\)Yikun Zhang, Geyan Ye, Chaohao Yuan, Bo Han, Long\-Kai Huang, Jianhua Yao, Wei Liu, and Yu Rong\. 2025\.Atomas: Hierarchical adaptive alignment on molecule\-text for unified molecule understanding and generation\.In*The Thirteenth International Conference on Learning Representations*\.
- Zhao et al\. \(2023\)Haiteng Zhao, Shengchao Liu, Ma Chang, Hannan Xu, Jie Fu, Zhihong Deng, Lingpeng Kong, and Qi Liu\. 2023\.Gimlet: A unified graph\-text model for instruction\-based molecule zero\-shot learning\.*Advances in neural information processing systems*, 36:5850–5887\.
- Zheng et al\. \(2023\)Lianmin Zheng, Wei\-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P\. Xing, Hao Zhang, Joseph E\. Gonzalez, and Ion Stoica\. 2023\.Judging llm\-as\-a\-judge with mt\-bench and chatbot arena\.

## Appendices

## Appendix AAlgorithm

Algorithm[1](https://arxiv.org/html/2607.03007#alg1)summarizes the complete training procedure ofMolBasic, which follows the staircase learning framework and adopts a two\-stage optimization strategy for each task\. Starting from a base language model𝒞0\\mathcal\{C\}\_\{0\}, the algorithm progressively equips the model with increasingly advanced molecular reasoning capabilities along a predefined task path𝒫=\(vstruct,vprop,vopt\)\\mathcal\{P\}=\(v\_\{\\text\{struct\}\},v\_\{\\text\{prop\}\},v\_\{\\text\{opt\}\}\)\.

Algorithm 1MolBasic Training Algorithm0:Base model

𝒞0\\mathcal\{C\}\_\{0\}, task path

𝒫=\(vstruct,vprop,vopt\)\\mathcal\{P\}=\(v\_\{\\text\{struct\}\},v\_\{\\text\{prop\}\},v\_\{\\text\{opt\}\}\), datasets

\{𝒟v,𝒟vCoT\}v∈𝒫\\\{\\mathcal\{D\}\_\{v\},\\mathcal\{D\}^\{\\text\{CoT\}\}\_\{v\}\\\}\_\{v\\in\\mathcal\{P\}\}
0:Fine\-tuned model with capability

𝒞opt\\mathcal\{C\}\_\{\\text\{opt\}\}
1:Initialize LoRA parameters

θ\\theta
2:

𝒞prev←𝒞0\\mathcal\{C\}\_\{\\text\{prev\}\}\\leftarrow\\mathcal\{C\}\_\{0\}
3:foreach task

vkv\_\{k\}in

𝒫\\mathcal\{P\}do

4:// Stage 1: Answer Learning

5:foreach

\(S,\(q,a\)\)\(S,\(q,a\)\)in

𝒟vk\\mathcal\{D\}\_\{v\_\{k\}\}do

6:Compute

ℒstage1=−∑t=1\|a\|log⁡Pθ​\(at∣S,q,a<t\)\\mathcal\{L\}\_\{\\text\{stage1\}\}=\-\\sum\_\{t=1\}^\{\|a\|\}\\log P\_\{\\theta\}\(a\_\{t\}\\mid S,q,a\_\{<t\}\)
7:Update

θ\\thetavia gradient descent

8:endfor

9:// Stage 2: Reasoning Learning

10:foreach

\(S,\(q,a^\)\)\(S,\(q,\\hat\{a\}\)\)in

𝒟vkCoT\\mathcal\{D\}^\{\\text\{CoT\}\}\_\{v\_\{k\}\}do

11:Compute

ℒstage2=−∑t=1\|a^\|log⁡Pθ​\(a^t∣S,q,a^<t\)\\mathcal\{L\}\_\{\\text\{stage2\}\}=\-\\sum\_\{t=1\}^\{\|\\hat\{a\}\|\}\\log P\_\{\\theta\}\(\\hat\{a\}\_\{t\}\\mid S,q,\\hat\{a\}\_\{<t\}\)
12:Update

θ\\thetavia gradient descent

13:endfor

14:

𝒞vk←fvk​\(𝒞prev,𝒟vkCoT\)\\mathcal\{C\}\_\{v\_\{k\}\}\\leftarrow f\_\{v\_\{k\}\}\(\\mathcal\{C\}\_\{\\text\{prev\}\},\\mathcal\{D\}^\{\\text\{CoT\}\}\_\{v\_\{k\}\}\)
15:

𝒞prev←𝒞vk\\mathcal\{C\}\_\{\\text\{prev\}\}\\leftarrow\\mathcal\{C\}\_\{v\_\{k\}\}
16:endfor

17:returnModel with capability

𝒞opt\\mathcal\{C\}\_\{\\text\{opt\}\}

For each taskvk∈𝒫v\_\{k\}\\in\\mathcal\{P\}, training is performed in two consecutive stages\.Stage 1 \(Answer Learning\)focuses on acquiring task\-specific domain knowledge by supervising only the final answers\. In this stage, the model is optimized to correctly generate task outputs given the molecular inputSSand queryqq, establishing reliable output anchors and preventing early\-stage reasoning drift\.

Stage 2 \(Reasoning Learning\)further refines the model by incorporating Chain\-of\-Thought \(CoT\) supervision\. Instead of predicting only the final answer, the model is trained to generate the complete reasoning trajectorya^=\(z,a\)\\hat\{a\}=\(z,a\), wherezzrepresents intermediate reasoning steps\. This stage explicitly constrains the reasoning process, reduces ambiguity in intermediate representations, and improves both accuracy and interpretability\.

After completing both stages for taskvkv\_\{k\}, the updated model capability𝒞vk\\mathcal\{C\}\_\{v\_\{k\}\}is obtained and passed as initialization to the next task in the staircase\. In this manner, structural comprehension learned at earlier stages is preserved and reused when learning higher\-level tasks such as structural property prediction and molecular optimization\. After all tasks in𝒫\\mathcal\{P\}are completed, the algorithm returns a model with capability𝒞opt\\mathcal\{C\}\_\{\\text\{opt\}\}, which integrates structure\-aware reasoning across all stages\.

## Appendix BData Preparation and Training Details

We prepare the training data from the PubChemSTM datasetLiu et al\. \([2023b](https://arxiv.org/html/2607.03007#bib.bib26)\), covering all eight structure comprehension tasks \(see Table[6](https://arxiv.org/html/2607.03007#A2.T6)\): atom counting, total bond counting, specific bond counting, element counting, substructure recognition, SMILES to graph conversion, graph to SMILES conversion, and formula generation\. For data cleaning, we first remove duplicate samples across all tasks\. For the two SMILES\-Graph conversion tasks, we further filter out samples whose graph representation exceeds 16K characters, which empirically corresponds to exceeding the 2048\-token context limit under our tokenizer\. After preprocessing, the dataset contains 1,181,816 valid samples expanded across eight tasks, corresponding to 185,286 unique molecules\. The expanded samples are split into training, validation, and test sets\.

We summarize the dataset sources and sizes used for all molecular understanding tasks in Table[5](https://arxiv.org/html/2607.03007#A2.T5)\.

Table 5:Dataset statistics for different molecular understanding tasks\. Values in parentheses indicate the expanded sample counts after constructing MSC tasks\.TaskDatasetTotalTrainingValidTestStructure comprehension\(expand by MSC\)PubChemSTM185,286\(1,181,816\)129,699\(737,828\)37,059\(296,012\)18,528Property predictionPubChem100,00095,000–5,000Molecular optimizationTDC175,173157,673–17,500Table 6:Multi\-level structure comprehension tasks\.LevelTaskTask TypeNodeAtom CountingNumerical RegressionElement CountingNumerical RegressionEdgeTotal Bond CountingNumerical RegressionSpecific Bond CountingNumerical RegressionGraphFormula ConversionSequence GenerationSubstructure RecognitionBinary ClassificationGraph to SMILESSequence GenerationSMILES to GraphSequence Generation

Stage 1 \(Answer Learning\): To emphasize the importance of SMILES\-Graph mutual conversion as the core task, we apply a weighted sampling strategy\. The two conversion tasks are assigned a weight of 2\.0, while all other tasks have a weight of 1\.0, resulting in approximately doubled sampling frequency for the conversion tasks\.

Stage 2 \(Reasoning Learning\): We construct a smaller\-scale dataset by sampling 1/10 from the Stage 1 validation set\. Task weights are adjusted based on Stage 1 performance to prioritize underperforming tasks: atom counting \(3\.0\), total bond counting \(3\.0\), formula generation \(3\.0\), SMILES to graph \(3\.0\), graph to SMILES \(3\.0\), element counting \(1\.5\), specific bond counting \(1\.5\), and substructure recognition \(0\.5\)\. Importantly, Chain\-of\-Thought reasoning is only applied to the three tasks that showed suboptimal performance in Stage 1—atom counting, total bond counting, and formula generation—while other tasks retain the answer\-only training format from Stage 1\. The Stage 1 validation set is used exclusively for Stage 2 training and is not involved in model evaluation, ensuring no data leakage\.

We provide the detailed training configurations for each stage in[Table˜7](https://arxiv.org/html/2607.03007#A2.T7)\. Our progressive training follows a sequential pipeline: Multi\-level Structure Comprehension \(MSC\) → Structural Property Prediction \(Prop\) → Molecular Optimization \(Optim\), where each task consists of two stages—Stage 1 \(Answer Learning\) and Stage 2 \(Reasoning Learning\)\. The LoRA adapter from each stage serves as the initialization for the subsequent stage, enabling knowledge accumulation throughout the training process\. All stages share the same LoRA architecture \(rank=64, alpha=128\) and learning rate \(1e\-4\) with cosine scheduling, while batch sizes and training epochs are adjusted based on dataset sizes and task complexity\.

We also report training and inference efficiency\. The MSC\-S1 stage, which processes the largest dataset, requires approximately 6 days on 4×\\times48 GB RTX 4090 GPUs\. All subsequent stages are significantly more efficient, each completing within a few hours on 2×\\times80 GB A800 GPUs\. For inference, all inference times are measured with batch size 1, except Optim\-S2\. For Optim\-S2, full CoT generation takes about 84s per sample withmax\_new\_tokens=1024at batch size 1; the reported 7\.089s is obtained by batched inference\. This demonstrates that our CoT reasoning approach incurs minimal additional computational cost in practical deployment scenarios\.

Artifact use:We use publicly available datasets, models, and tools for research purposes and follow their original terms of use\.

Table 7:Training configurations across different stages\. MSC: Multi\-level Structure Comprehension, Prop: Structural Property Prediction, Optim: Molecular Optimization\. S1: Answer Learning, S2: Reasoning Learning\.ParameterMSC\-S1MSC\-S2Prop\-S1Prop\-S2Optim\-S1Optim\-S2Base ModelQwen3\-8B\+MSC\-S1\+MSC\-S2\+Prop\-S1\+Prop\-S2\+Optim\-S1LoRA Rank64LoRA Alpha128LoRA Dropout0\.05Learning Rate1e\-4Effective Batch Size321632128128128Epochs152522LR SchedulercosineWarmup Ratio0\.1Max Seq Length2048DeepSpeed\-ZeRO\-2\-\-\-\-Steps518818501336166511081108FLOPs2\.32e192\.86e183\.39e182\.06e193\.20e181\.00e19Samples/sec3\.274\.315\.364\.755\.505\.04Steps/sec0\.0260\.0670\.0420\.0180\.0210\.020Infer\. Time \(s\)0\.3010\.3115\.1394\.9776\.8637\.089
## Appendix CDefinitions of Molecular Properties

In this section, we provide concise definitions of the structural properties used in our experiments, all of which are directly derived from molecular structure\.

- •Molecular Weight \(MW\):The sum of the atomic weights of all atoms in a molecule, reflecting its overall molecular size\.
- •Octanol–Water Partition Coefficient \(LogP\):The logarithm of the ratio of a compound’s concentration in octanol to that in water, measuring its hydrophobicity\.
- •Topological Polar Surface Area \(TPSA\):The surface area contributed by polar atoms \(typically oxygen and nitrogen and their attached hydrogens\), indicating a molecule’s polarity and permeability\.
- •Hydrogen Bond Donors \(HBD\):The number of functional groups in a molecule capable of donating a hydrogen atom to form a hydrogen bond\.
- •Hydrogen Bond Acceptors \(HBA\):The number of atoms in a molecule capable of accepting a hydrogen bond through lone pairs\.
- •Number of Rotatable Bonds \(RB\):The count of non\-ring, single bonds between heavy atoms that allow free rotation, characterizing molecular flexibility\.
- •Quantitative Estimate of Drug\-likeness \(QED\):A composite score that quantifies drug\-likeness by integrating multiple physicochemical properties using a desirability function\.

## Appendix DEvaluation on External Benchmarks

To verify that the gains of MolBasic are not limited to our proposed Molecular Structure Comprehension benchmark, we further evaluate it on two external molecular understanding benchmarks for fair comparison\. First, the CleanMol benchmarkJang et al\. \([2025](https://arxiv.org/html/2607.03007#bib.bib17)\)focuses on SMILES parsing tasks, including functional\-group recognition, ring counting, chain\-length prediction, and canonical SMILES generation\. As shown in Table[8](https://arxiv.org/html/2607.03007#A4.T8), MolBasic achieves the best accuracy on most tasks, substantially outperforming both few\-shot general LLMs and most SFT baselines, demonstrating strong general structural parsing ability\.

Second, ChemCoTBenchHao et al\. \([2026](https://arxiv.org/html/2607.03007#bib.bib13)\)provides a broader evaluation of molecular understanding, including functional\-group counting, ring counting, Murcko scaffold extraction, ring\-system scaffold extraction, and SMILES equivalence judgment\. As shown in Table[9](https://arxiv.org/html/2607.03007#A4.T9), MolBasic achieves the best performance on most metrics and remains competitive on ring counting\. These results further confirm that MolBasic learns transferable structural understanding beyond our proposed benchmark\.

Table 8:SMILES parsing performance on CleanMol benchmark\. All metrics are reported as accuracy\.Task typeModelFGRingChainCanonical5\-shotDeepseek\-V3\-chat0\.89120\.62660\.29760\.1484GPT\-4o0\.87500\.59550\.28570\.1078Galactica\-6\.7B0\.50000\.07320\.15110\.0000SFTLlama3\.1\-8B \(Single\)0\.94140\.86120\.98590\.9356Llama3\.1\-8B \(Multi\)0\.98910\.87070\.98510\.9463Qwen2\.5\-7B \(Single\)0\.98910\.86740\.99070\.7593Qwen2\.5\-7B \(Multi\)0\.99010\.87500\.99020\.9262MolBasic0\.99560\.97740\.99460\.8836Table 9:ChemCoTBench molecular understanding evaluation compared with molecular LLMs\. The evaluation uses MAE for functional\-group and ring counting, Tanimoto similarity for Murcko scaffold extraction, and Accuracy for ring\-system, SMILES equivalence\.ModelsFunc\-GroupScaffoldSMILESFG↓\\downarrowRing↓\\downarrowMurcko↑\\uparrowRing\-sys↑\\uparrowEq\.↑\\uparrowEther0Failed0\.35FailedFailed0\.63BioMedGPT\-7B1\.62\.430\.180\.530\.39BioMistral\-7B1\.01\.850\.040\.330\.50MolBasic0\.730\.650\.430\.650\.67
## Appendix EFew\-shot in Bioactivity Prediction Tasks

Table 10:Few\-shot results on bioactivity prediction tasks\.Bestandsecond bestresults are highlighted among few\-shot methods\.ShotMethodBACEHIVBBBPTox21Few\-shot Methods0\-shotMolRAG0\.5040\.5140\.5460\.534MolBasic0\.5000\.4990\.5140\.5631\-shotMolRAG0\.5940\.5430\.5460\.545MolBasic0\.6190\.6640\.5500\.5612\-shotMolRAG0\.6150\.5680\.5640\.55MolBasic0\.6490\.6540\.6110\.5794\-shotMolRAG0\.6260\.5950\.5720\.566MolBasic0\.6610\.6590\.6330\.573Pre\-training MethodsGIMLET0\.6960\.6620\.5940\.612KVPLM0\.5130\.6120\.6020\.492MoMu0\.6660\.5030\.4980\.576Galactica\-1\.3B0\.5650\.3390\.5390\.495Graph\-based NetworksGCN0\.7360\.7570\.6490\.749GAT0\.6970\.7290\.6650\.754GIN0\.7010\.7530\.6580\.74Graphormer0\.7760\.7450\.7020\.759Beyond fine\-tuning on downstream tasks, we conduct few\-shot experiments on tasks logically related but not seen during training, to demonstrate the transferability of our framework to novel yet relevant tasks after acquiring fundamental capabilities\. We select bioactivity property prediction, closely related to molecular structural properties, and compare with MolRAG, the current state\-of\-the\-art few\-shot method\. As shown in[Table˜10](https://arxiv.org/html/2607.03007#A5.T10), MolBasic achieves better performance on most properties and approaches the performance of pre\-training methods\. We provide descriptions of the baseline categorization used in[Table˜10](https://arxiv.org/html/2607.03007#A5.T10)in Appendix[F](https://arxiv.org/html/2607.03007#A6)\.

Table 11:Comparison of MolBasic and GPT\-5 on structure comprehension tasks\. Metrics are accuracy \(↑\\uparrow\)\.ModelAtomElem\.Total BondSpec\. BondFormulaSubstruct\.S→\\rightarrowGG→\\rightarrowSMolBasic42\.4%73\.01%49%82\.9%44\.0%98\.1%94\.7%85\.7%GPT\-595%97\.5%71\.5%75\.5%80\.7%85\.5%0%1\.38%

Table 12:Comparison of MolBasic and GPT\-5 on property prediction tasks\. Metrics are MAE \(↓\\downarrow\)\.ModelMWLogPTPSAHBDHBARBQEDMolBasic35\.670\.5311\.120\.590\.710\.530\.08GPT\-58\.960\.9412\.950\.050\.271\.601\.13Table 13:Comparison of MolBasic and GPT\-5 on molecular optimization\. Higher is better for all metrics\.ModelΔ\\DeltaLogPDiversityValidityMolBasic2\.270\.8698\.7%GPT\-51\.890\.8795\.5%
## Appendix FFew\-shot Baseline Settings

In this section, we describe the baseline methods used for few\-shot bioactivity prediction and clarify the rationale behind their categorization\. We group baseline methods into three categories according to their training paradigm and supervision regime:*few\-shot methods*,*pre\-training\-based methods*, and*graph\-based networks*\.

Few\-shot Methods\.Few\-shot methods are evaluated under the same data\-scarce setting as MolBasic, where models are provided with only a limited number of labeled examples per task at inference time\. Specifically, we include MolRAG as a representative few\-shot baseline, which enhances large language models with retrieval over molecular databases\. Both MolRAG and MolBasic are evaluated under identicalkk\-shot settings \(k∈\{0,1,2,4\}k\\in\\\{0,1,2,4\\\}\), ensuring a fair comparison in terms of supervision level and data availability\.

Pre\-training\-based Methods\.Pre\-training\-based methods leverage large\-scale labeled or unlabeled molecular datasets to acquire task\-relevant knowledge before evaluation\. Models such as GIMLET, KVPLM, MoMu, and Galactica\-1\.3B fall into this category, as they rely on extensive pre\-training on molecular or molecule–text corpora\. While these methods are not few\-shot by design, we include them as reference points to contextualize the performance of few\-shot approaches against models with substantially more task\-specific prior knowledge\.

Graph\-based Networks\.Graph\-based networks, including GCN, GAT, GIN, and Graphormer, represent task\-specific models trained with full supervision on molecular graphs\. These methods serve as upper\-bound references, as they have direct access to explicit molecular graph structures and are optimized specifically for bioactivity prediction tasks\.

Overall, this categorization highlights the trade\-off between supervision strength and model generality, and situates MolBasic within the challenging few\-shot regime, where strong structural reasoning must be achieved with minimal labeled data\.

## Appendix GGPT\-5 Evaluation on Molecular Tasks

To further contextualize the contribution of Chain\-of\-Thought supervision, we evaluate GPT\-5 directly on structure comprehension, property prediction, and molecular optimization tasks\. Tables[11](https://arxiv.org/html/2607.03007#A5.T11)–[13](https://arxiv.org/html/2607.03007#A5.T13)summarize the direct evaluation of GPT\-5 and its comparison with MolBasic\.

These results show that GPT\-5 provides strong reasoning signals, particularly for numerical and additive properties, validating its use as a source of Chain\-of\-Thought supervision\. MolBasic leverages these distilled signals through progressive, stage\-wise learning and CoT supervision\. As a result, MolBasic progressively internalizes structural reasoning: it first acquires foundational structure comprehension, then transfers these capabilities to property prediction and molecular optimization tasks\. This training pipeline enables MolBasic to achieve competitive performance relative to GPT\-5 itself, especially for structure\-dependent properties such as LogP and QED, while maintaining better performance in molecular optimization\.

## Appendix HAdditional Analysis

### H\.1Error Analysis

We analyze the sources of prediction errors in numerical regression tasks\. As shown in[Figure˜4](https://arxiv.org/html/2607.03007#A8.F4), prediction errors tend to increase as the ground\-truth values become larger, especially for atom and bond counting tasks\. This trend mainly arises from two factors\. First, LLMs are generally less sensitive to precise numerical values, making exact counting difficult\. Second, larger molecules usually correspond to longer and more structurally complex SMILES sequences, containing more rings, branches, and functional groups, which further increases parsing difficulty\. These factors lead to prediction errors that scale with molecular size\.

![Refer to caption](https://arxiv.org/html/2607.03007v1/latex/figures/error.png)Figure 4:Error distribution across four regression tasks in the multi\-level structure comprehension benchmark\. Each panel shows the mean absolute error \(MAE\) as a function of ground\-truth values for the corresponding task: \(a\) Heavy Atom Counting, \(b\) Total Bond Counting, \(c\) Element\-specific Counting, and \(d\) Specific Bond Counting\. The shaded regions represent the standard deviation of errors within each bin\.We further analyze MolBasic’s prediction errors on molecular property prediction tasks\. As shown in[Figure˜5](https://arxiv.org/html/2607.03007#A8.F5), most absolute errors are concentrated near zero, indicating generally accurate predictions, with only a small number of extreme predicted values forming long\-tail distributions\.[Figure˜6](https://arxiv.org/html/2607.03007#A8.F6)further shows that MolBasic remains stable on molecules of normal size across most properties, while performance degradation mainly appears for larger molecules, where longer SMILES strings and more complex structures make structural parsing and property reasoning more difficult\.

![Refer to caption](https://arxiv.org/html/2607.03007v1/latex/figures/property_mae_distribution.png)Figure 5:Distribution of sample\-level absolute errors for seven molecular properties\. Each subplot corresponds to one property\. The solid curve shows the density distribution of absolute errors, and the shaded area indicates the filled density region\. The vertical dashed line marks the mean absolute error \(MAE\) of the corresponding property\.![Refer to caption](https://arxiv.org/html/2607.03007v1/latex/figures/property_mae_length.png)Figure 6:Property prediction errors with respect to molecular size\. Each subplot shows the mean absolute error \(MAE\) of one property across bins of heavy atom counts\. The line with markers represents the average MAE within each bin, and the shaded region denotes the standard deviation of errors in that bin\.
### H\.2Multi\-objective Molecular Optimization

In[Section˜4\.2\.2](https://arxiv.org/html/2607.03007#S4.SS2.SSS2), we use LogP improvement as the representative objective for molecular optimization\. To examine whether optimizing a single property compromises other drug\-relevant properties, we further evaluate the changes in QED and synthetic accessibility \(SA\) under this single\-objective setting\. As shown in Table[14](https://arxiv.org/html/2607.03007#A8.T14), although the model is only instructed to improve LogP, QED decreases only slightly, while SA improves substantially\. This indicates that MolBasic, after acquiring molecular structure and property knowledge, can optimize molecules toward the target objective without severely damaging other important chemical properties\.

We further conduct a multi\-objective fine\-tuning experiment, where the model is explicitly trained to increase LogP and QED while decreasing SA\. The results show that all three objectives are improved simultaneously: LogP remains substantially increased, QED improves from 0\.74 to 0\.79 with the QED≥0\.8\\geq 0\.8ratio increasing from 36\.3% to 57\.3%, and SA decreases from 3\.64 to 2\.60 with the SA<3<3ratio increasing from 23\.5% to 73\.8%\.

Table 14:Multi\-objective molecular optimization results\. HigherΔ\\DeltaLogP and QED indicate better performance, while lower SA indicates better synthetic accessibility\. QED≥0\.8\\geq 0\.8denotes the percentage of generated molecules with high drug\-likeness, and SA<3<3denotes the percentage of easy\-to\-synthesize molecules\.SettingΔ\\DeltaLogP↑\\uparrowQED↑\\uparrowQED≥0\.8\\geq 0\.8↑\\uparrowSA↓\\downarrowSA<3<3↑\\uparrowSingle\-obj\. \(LogP only\)\+2\.270\.72→\\rightarrow0\.7139\.3%→\\rightarrow37\.6%3\.49→\\rightarrow2\.6532\.3%→\\rightarrow73\.9%Multi\-obj\. \(LogP\+QED\+SA\)\+2\.100\.74→\\rightarrow0\.7936\.3%→\\rightarrow57\.3%3\.64→\\rightarrow2\.6023\.5%→\\rightarrow73\.8%
### H\.3Out\-of\-distribution Analysis

We conduct Murcko scaffold overlap analysis on both downstream tasks to examine whether MolBasic relies on scaffold memorization\. For property prediction, 55\.3% of test scaffolds are unseen during training; for molecular optimization, 24\.1% of test samples are OOD\. Table[15](https://arxiv.org/html/2607.03007#A8.T15)compares the performance on the full test set and scaffold\-OOD subsets\.

Table 15:Performance on the full test set and scaffold\-OOD subsets\. Property prediction metrics are MAE \(↓\\downarrow\), and molecular optimization is measured byΔ\\DeltaLogP \(↑\\uparrow\)\.SubsetMWLogPTPSAHBDHBARBQEDΔ\\DeltaLogPAll13\.980\.282\.710\.050\.160\.360\.062\.27OOD15\.020\.303\.070\.060\.180\.330\.062\.11The results show that MolBasic maintains stable performance on scaffold\-OOD samples\. Compared with the full test set, property prediction errors increase only mildly on OOD molecules, and the LogP improvement in molecular optimization decreases slightly from 2\.27 to 2\.11\. Importantly, neither downstream task exhibits performance collapse under scaffold shift, suggesting that MolBasic learns transferable structure\-property reasoning rather than merely memorizing training scaffolds\.

## Appendix IFuture Work

Regarding capability enhancement, further improvements are needed in LLMs’ ability to comprehensively comprehend fundamental molecular structures\. As our current results indicate, the model still requires strengthening in complex regression tasks and interpretable graph structure conversion\. Future work will focus on developing more sophisticated training strategies to better address these challenging aspects of structural perception\.

Regarding scope expansion, our next direction aims to approximate the reasoning process of an "AI chemist" by integrating LLMs’ structural understanding with efficient retrieval of chemical literature\. This would establish a structure\-reasoning\-driven, knowledge\-assisted paradigm for molecular understanding\. We believe this combination of structural perception and domain knowledge retrieval will further advance molecular LLMs toward more comprehensive and practical chemical reasoning capabilities\.

## Appendix JChain\-of\-Thought Templates

In this section, we present the standardized Chain\-of\-Thought \(CoT\) templates used in MolBasic for different molecular understanding tasks\. These templates guide the model through structured reasoning processes, ensuring consistent and interpretable outputs\. For structure comprehension tasks, we apply CoT only during Stage 2 \(reasoning training\) for three tasks that showed suboptimal performance in Stage 1: atom counting, total bond counting, and formula conversion\. The CoT for these tasks primarily constrains the accuracy of intermediate calculation steps, as illustrated in[Figure˜7](https://arxiv.org/html/2607.03007#A10.F7)\. For structural property prediction, the reasoning follows a four\-step protocol: atom counting, functional group identification, ring system analysis, and property\-specific reasoning \([Figure˜8](https://arxiv.org/html/2607.03007#A10.F8)\)\. For molecular optimization, the CoT systematically analyzes polar and lipophilic groups, proposes modification strategies, and verifies property improvements \([Figure˜9](https://arxiv.org/html/2607.03007#A10.F9)\)\. These standardized templates enable the model to develop robust structural understanding while maintaining transparency in its reasoning process\.

Task: Total Bond CountingInstruction:Count the total number of chemical bonds in the molecular structure\.Input:C=C\(C\(=O\)O\)C1CCC2\(O\)C=CC\(=O\)C\(C\)C2\(C\)C1Chain\-of\-Thought:•Parsed the SMILES into atoms and bonds\.•Identified two carbonyl groups and all substituent connections\.•Enumerated all edges: total = 20\.•Verification:V−1\+cycles=19−1\+2=20V\-1\+\\text\{cycles\}=19\-1\+2=20✓\.Answer:20

Task: Heavy Atom CountingInstruction:Count the total number of heavy atoms \(non\-hydrogen\) in the molecule\.Input:CC\(C\)=CCOc1c\(C\)c\(C\)c\(C\(=O\)O\)c\(O\)c1CChain\-of\-Thought:•First aromatic ringc1ccc\.\.\.c1: 6 ring carbons = 6 atoms\.•Aliphatic chain and branches: 9 C = 9 atoms\.•Oxygen atoms, including carbonyl and ether groups: 4 O = 4 atoms\.•Total =6\+9\+4=196\+9\+4=19heavy atoms\.Answer:19

Task: Formula ConversionInstruction:Generate the standard molecular formula for the molecule\.Input:CC\(C\)=CCOc1cc\(=O\)oc2cccc\(C\)c12Chain\-of\-Thought:•Count carbons from SMILES: 15 total\.•Count heteroatoms: O = 3\.•Determine hydrogens from substitution and valence:–Ring hydrogens: 4 H\.–Aliphatic groups: 12 H\.–Total H = 16\.Answer:C15H16O3

Figure 7:Examples of standardized Chain\-of\-Thought reasoning for structure perception tasks\.Task: Structural Property Prediction \(Example 1\)Input:CC1=CC\(=C\(C=C1\)C\)C\(=O\)CC2\(C3=C\(C=CC\(=C3\)Cl\)NC2=O\)OChain\-of\-Thought:1\. Atoms and counts\(Molecular formula: C18H16ClNO3\):•C: 18, H: 16, N: 1, O: 3, Cl: 1\.•Heavy atoms: 23\.2\. Functional groups identified:•Hydroxyl \(–OH\), ketone \(C=O\), amide \(–C\(=O\)NH–\), lactam \(cyclic amide\), and chlorine \(–Cl\)\.3\. Ring systems:•Total rings: 3, aromatic rings: 2, aliphatic rings: 1, heterocyclic rings: 1\.4\. Property reasoning:•MW: Sum of atomic weights from C18H16ClNO3= 329\.78 g/mol\.•LogP: Moderately lipophilic due to aromatic rings, balanced by polar groups→\\rightarrow3\.37\.•TPSA: Contributions from N and O atoms→\\rightarrow66\.4 Å2\.•HBD: 2, from NH/OH groups\.•HBA: 3, from N/O atoms with lone pairs\.•RB: 3, from non\-ring single bonds excluding terminal bonds\.•QED: Good drug\-likeness→\\rightarrow0\.85\.Answer:MW=329\.78, LogP=3\.37, TPSA=66\.4, HBD=2, HBA=3, RB=3, QED=0\.85

Task: Structural Property Prediction \(Example 2\)Input:CCC1=C\[N\+\]\(=C\(C=C1\)C\)CC2=CC\(=O\)N3C=CSC3=N2Chain\-of\-Thought:1\. Atoms and counts\(Molecular formula: C15H16N3OS\+\): C: 15, H: 16, N: 3, O: 1, S: 1\. Heavy atoms: 20\.2\. Functional groups: No common functional groups identified\.3\. Ring systems: Total rings: 3, aromatic rings: 3, heterocyclic rings: 3\.4\. Property reasoning:•MW: 286\.38 g/molLogP: 1\.96, slightly lipophilicTPSA: 38\.25 Å2\.•HBD: 0HBA: 4RB: 3QED: 0\.69\.Answer:MW=286\.38, LogP=1\.96, TPSA=38\.25, HBD=0, HBA=4, RB=3, QED=0\.69

Figure 8:Examples of standardized Chain\-of\-Thought reasoning for structural property prediction\.Task: Molecular Optimization \(Example 1\)Input:Cc1ccccc1/C=C1/Oc2c\(ccc\(O\)c2C\[NH\+\]2CCC\[C@@H\]\(C\)C2\)C1=OChain\-of\-Thought:1\. Original molecule analysis: C23H25NO\+3\{\}\_\{3\}^\{\+\}, 4 rings, 2 aromatic rings, LogP: 3\.13, HBD: 2, HBA: 3, charge: \+1\.2\. Polar groups that decrease LogP: quaternary ammonium \[N\+\], hydroxyl –OH, ketone C=O, and ether –O–\.3\. Lipophilic groups that increase LogP: 2 phenyl rings and a 23\-carbon skeleton\.4\. Optimization strategy:•Neutralize ionic charge, which is expected to provide a major LogP increase\.•Reduce HBD:2→12\\rightarrow 1\.•Add a piperazine ring with a phenyl substituent\.5\. Verification: LogP:3\.13→4\.643\.13\\rightarrow 4\.64\(Δ=\+1\.51\\Delta=\+1\.51\) ✓\.Output:Cc1ccccc1/C=C1\\Oc2c\(ccc\(O\)c2CN2CCN\(c3ccccc3\)CC2\)C1=O

Task: Molecular Optimization \(Example 2\)Input:C\[NH\+\]\(C\)C\(C\)\(C\)CNC\(=O\)c1cnn\(Cc2ccccc2Cl\)c1Chain\-of\-Thought:1\. Original molecule analysis: C17H24ClN4O\+, 2 aromatic rings, LogP: 1\.24, HBD: 2, HBA: 3, charge: \+1\.2\. Polar groups that decrease LogP: quaternary ammonium \[N\+\], secondary amine –NH–, and amide –CONH–\.3\. Lipophilic groups that increase LogP: phenyl ring, chlorine, and 17\-carbon skeleton\.4\. Optimization strategy:•Remove charged quaternary N, which is expected to provide a major LogP increase\.•Reduce HBD:2→12\\rightarrow 1\.•Add trifluoromethyl –CF3and phenyl ring\.5\. Verification: LogP:1\.24→4\.861\.24\\rightarrow 4\.86\(Δ=\+3\.62\\Delta=\+3\.62\) ✓\.Output:O=C\(Nc1ccccc1C\(F\)\(F\)F\)c1cnn\(Cc2ccccc2Cl\)c1

Figure 9:Examples of standardized Chain\-of\-Thought reasoning for molecular optimization \(LogP improvement\)\.
## Appendix KUse of AI Assistants

We used AI assistants during the preparation of this work for limited auxiliary purposes, including code debugging and grammar\-level writing assistance\. All research ideas, experimental design, analysis, conclusions, and final manuscript decisions were made by the authors\.

In addition, some of the evaluated baseline systems in our experiments are themselves AI assistants or general\-purpose LLMs, such as GPT\-series models and Qwen\-series models\. These models were used only as experimental subjects or baselines for evaluating molecular understanding capabilities, following the experimental protocols described in the paper\.

## Appendix LInterpretable Molecular Optimization Examples

We present representative examples of MolBasic’s responses on the objective molecular optimization task in[Section˜4\.2\.2](https://arxiv.org/html/2607.03007#S4.SS2.SSS2)\. Our framework generates optimized molecules while simultaneously providing interpretable analysis of the optimization rationale, including structural feature identification, modification strategies, and property change verification\.[Figures˜10](https://arxiv.org/html/2607.03007#A12.F10)and[11](https://arxiv.org/html/2607.03007#A12.F11)illustrate three such examples with detailed explanations\. Note that the intermediate property values in the explanations may not be perfectly accurate, as the model itself exhibits certain prediction errors on property estimation tasks \(as shown in[Table˜2](https://arxiv.org/html/2607.03007#S4.T2)\)\. Nevertheless, the overall trends in property changes remain correct and the errors stay within acceptable ranges, which is sufficient for guiding effective molecular modifications\. This demonstrates that MolBasic has developed genuine structural reasoning capabilities for molecular understanding\.

![Refer to caption](https://arxiv.org/html/2607.03007v1/latex/figures/case1.png)Figure 10:Case study 1: Molecular optimization examples with CoT explanations\.![Refer to caption](https://arxiv.org/html/2607.03007v1/latex/figures/case2.png)

![Refer to caption](https://arxiv.org/html/2607.03007v1/latex/figures/case3.png)

Figure 11:Case study 2 and 3: Molecular optimization examples with CoT explanations\.