PaperlayerView exact revision on arXiv

Source-linked AI summary

A Review of Large Language Models and Autonomous Agents in Chemistry

Mayk Caldas Ramos, Christopher J. Collison, Andrew D. White

arXiv:2407.01603v3cs.LGcs.AIcs.CLphysics.chem-ph

TL;DR

Chemistry lacks a consolidated account of how LLMs and tool-using agents address molecular design, property prediction, synthesis, and automation. This review synthesizes capabilities and designs across chemistry and other scientific domains, finding that agents support diverse research tasks while benchmark quality and interpretability remain challenges.

  • Problem

    Chemistry requires improved approaches for property prediction, molecule generation, synthesis prediction, and automation across a vast chemical space.

    Source: §1.1 · pp. 3–4

  • Method

    The review synthesizes recent LLM and autonomous-agent capabilities, designs, challenges, and future directions across chemistry and other scientific domains.

    Source: §4.5 · p. 31, §5.3 · p. 37, §7 · p. 42

  • Results

    PaperQA achieved 87.9% precision and 69.5% accuracy on LitQA, compared with a 66.8% human accuracy baseline, while agents supported literature and experiment-planning tasks.

    Source: §5.1 · p. 34, §5 · p. 34, §5.3 · p. 37

  • Takeaways & Limitations

    LLM-based agents provide a framework for integrating specialized tools with scientific reasoning across chemistry research workflows.

    Source: §4.5 · p. 31, §7 · p. 42

    Common chemistry benchmarks can be small, erroneous, inconsistent, and poorly representative of real-world chemistry problems.

    Source: §3.1 · p. 12

Abstract

from arXiv · show

Large language models (LLMs) have emerged as powerful tools in chemistry, significantly impacting molecule design, property prediction, and synthesis optimization. This review highlights LLM capabilities in these domains and their potential to accelerate scientific discovery through automation. We also review LLM-based autonomous agents: LLMs with a broader set of tools to interact with their surrounding environment. These agents perform diverse tasks such as paper scraping, interfacing with automated laboratories, and synthesis planning. As agents are an emerging topic, we extend the scope of our review of agents beyond chemistry and discuss across any scientific domains. This review covers the recent history, current capabilities, and design of LLMs and autonomous agents, addressing specific challenges, opportunities, and future directions in chemistry. Key challenges include data quality and integration, model interpretability, and the need for standard benchmarks, while future directions point towards more sophisticated multi-modal agents and enhanced collaboration between agents and experimental methods. Due to the quick pace of this field, a repository has been built to keep track of the latest studies: https://github.com/ur-whitelab/LLMs-in-science.

1 Introduction

AI in chemistry has progressed from early computational and expert-system methods through deep learning to LLMs and autonomous agents, which support molecular design, prediction, synthesis planning, and laboratory automation. The review frames chemical discovery around property prediction, property-directed molecule generation, synthesis prediction, and automation across vast chemical spaces.

1 Introduction

  • AI in chemistry builds on decades of computational modeling, expert systems, QSAR, neural networks, and machine-learning methods including Support Vector Machines and Random Forests.High-throughput screening generated datasets that improved learning from molecular structures and biological activities, while newer methods outperformed traditional statistical approaches.Source: §1 · p. 3
  • Deep learning advanced molecular property prediction, drug discovery, synthesis prediction, and materials identification through RNNs, CNNs, and GNNs.These models captured complex patterns in data and enabled identification of novel materials for energy storage and conversion.Source: §1 · p. 3
  • LLMs adapted chemical languages such as SMILES for synthesis prediction and molecule generation, while autonomous agents extend these models to data interpretation and robotic experimentation.These capabilities could produce larger, consistent experimental datasets and shorten drug and material commercialization timelines.Source: §1 · p. 3

1.1 Challenges in Chemistry

  • The review identifies property prediction, property-directed molecule generation, synthesis prediction, and automation as four connected challenges in chemical discovery.Property prediction determines whether compounds should be synthesized for applications such as indicators, light harvesters, or catalysts, and depends on high-quality data.Source: §1.1 · p. 3, Figure 1 · p. 3
  • Inverse design and de novo molecular generation remain difficult across vast chemical spaces, especially when optimizing properties such as solubility, toxicity, and blood-brain barrier permeability.The review examines the performance of state-of-the-art models for connecting molecular generation with property prediction.Source: §1.1 · pp. 3–4
  • Synthesis planning must identify inexpensive, available, and non-toxic starting materials while finding high-yield routes for molecules with suitable properties among up to 10^180 stable chemicals.Cataloguing failed syntheses and using autonomous chemical resources can expand datasets and accelerate exploration, making automation a major chemical task.Source: §1.1 · p. 4

2 Large Language Models

Large language models have displaced RNNs in sequence modeling through transformer architectures, whose attention-based designs support diverse chemistry tasks but remain constrained by context-window costs. Their development proceeds through staged pretraining, task adaptation, alignment, and increasingly specialized model types, including multimodal systems.

2 Large Language Models

  • Transformers replaced RNNs as the dominant LLM architecture because they avoid recurrent gradient failures in long sequences and underpin state-of-the-art chemistry sequence modeling.RNN gradients can vanish or explode as sequence length increases, while transformers, introduced by Vaswani et al. 71, dominate current chemistry sequence modeling.Source: §2 · p. 4

2.1 The Transformer

  • Transformers tokenize, embed, positionally encode, and attention-weight inputs before encoder-decoder processing generates contextually aligned outputs.Positional encoding supplies sequence order, while self-attention compares queries, keys, and values to integrate contextual information; encoder residual connections help information flow and mitigate vanishing gradients.Source: §2.1 · pp. 4–6
  • Transformer context windows are limited by quadratic attention cost, restricting effective inputs to a few thousand tokens and impairing long-range coherence in LLM-based agents.These limitations affect applications requiring deep contextual understanding; tokenization converts text into integer-valued vocabulary fragments before processing.Source: §2.1 · p. 4

2.2 Model training

  • LLMs are commonly pretrained self-supervised into base models, prompt-guided or supervised-fine-tuned into instruct models, and further optimized for preferences or downstream scientific tasks.Self-supervised learning extracts structure from vast unlabeled corpora, while prompt engineering preserves generality but can be sensitive to wording and assumed domain knowledge.Source: §2.2 · pp. 6–7
  • Alignment methods adjust LLM outputs toward human preferences, safety, and task goals using instruction tuning, RLHF and related reinforcement-learning methods, or reward-free preference optimization.RL models token generation as actions with rewards, whereas DPO, RRHF, and PRO compare completions directly; the literature remains active on method superiority.Source: §2.2 · pp. 6–7

2.3 Model types

  • Encoder-only, decoder-only, and encoder-decoder models specialize respectively in representation and property prediction, autoregressive generation, and sequence-to-sequence transformations such as reaction prediction.Figure 2 illustrates these architectures in chemistry, including molecular-property prediction, novel peptide generation, and reactant-to-product mapping.Source: Figure 2 · p. 5, §2.3 · p. 7

2.3.1 Encoder-only Models

  • BERT’s bidirectional masked-language pretraining enriched encoder representations for sentence- and token-level tasks, while RoBERTa improved understanding and generalization through larger-scale training and altered objectives.BERT also used Next Sentence Prediction; RoBERTa removed that task, dynamically changed masks, and used different hyperparameters.Source: §2.3.1 · pp. 7–8

2.3.2 Decoder-only Models

  • GPT introduced decoder-only causal generation, BART combined bidirectional encoding with autoregressive decoding, and later multi-task and multimodal models extended LLMs across tasks and modalities.T5 reframed tasks as text-to-text, Flan-T5 used instruction fine-tuning with chain-of-thought prompts, Fuyu handled images, and next-GPT supported text, audio, image, and video.Source: §2.3.2 · p. 8, §2.3.3 · p. 8, §2.3.4 · p. 8

3 LLMs for Chemistry and Biochemistry

This section surveys LLMs for chemistry and biochemistry, covering chemical-language and natural-language models across molecular design, property prediction, and synthesis prediction. It emphasizes trustworthy datasets and robust benchmarks as prerequisites for generalizable models, while highlighting cooperation between molecular and text-based LLMs as a path toward automation.

Model-task alignment

  • LLM architectures align with chemical tasks: encoder-only models excel at property prediction, decoder-only models at inverse design, and encoder-decoder models at synthesis prediction, while decoder-only models increasingly support property prediction.This expansion followed improvements in decoder-only models and the reformulation of regression as text completion.Source: §3 · p. 9

Scope and evolution

  • The section surveys LLMs that process SMILES, InChI, and chemistry-relevant natural-language descriptions, with primary attention to chemical representations and literature-based data extraction.The review presents a chronological perspective on model design, functionality, and value, summarized in Figure 4.Source: §3 · p. 9, Figure 4 · p. 10

Automation and agents

  • Mol-based and text-based LLMs can work together to advance chemistry automation, providing a foundation for broader autonomous-agent applications in scientific discovery.Figure 3 summarizes current LLM capabilities, while Figure 4 maps their chronological development in chemistry and biology.Source: §3 · p. 11

Data and benchmarks

  • Trustworthy, well-curated, and diverse datasets plus robust benchmarks are critical because narrow or poor-quality resources can limit generalization and obscure models’ real-world potential.Natural-language models are expected to increasingly help curate datasets and improve data quality.Source: §3 · p. 11

3.1 Molecular Representations, Datasets, and Benchmarks

Chemical language models depend on choosing suitable molecular representations, high-quality datasets, and benchmarks that reflect real-world chemistry. The section emphasizes that data quality, realistic evaluation, and formats aligned with molecular structure remain major bottlenecks.

Molecular Representations

  • Molecules can be represented as graphs, 3D point clouds, quantitative descriptors, or strings such as IUPAC names, SMILES, DeepSMILES, SELFIES, and InChI; this review focuses on strings for language models.Molecular descriptions also include structural formulas, electrostatic potentials, solubility, reactivity, and spectral data, which must be converted into numerical forms for machine learning.Source: §3.1 · p. 11

Datasets

  • High-quality, realistically grounded data—not sheer dataset size—is a central bottleneck, as curated reaction data can support state-of-the-art retrosynthesis despite being small.The largest chemical corpus is nearly two orders of magnitude smaller than LLaMA2’s training data, while synthesized-compound data can be over five orders of magnitude smaller; Mol-instructions provides ∼2M quality-controlled instructions. Segler et al. 181 achieved state-of-the-art retrosynthesis using the small, human-curated Reaxys dataset.Source: §3.1 · pp. 11–12

Benchmarks

  • MoleculeNet, the most commonly used labeled chemistry dataset, is small, error-prone, inconsistent, and insufficiently relevant to many real-world chemistry problems.A large public effort to generate and share high-quality data has been advocated for drug-discovery machine learning.188Source: §3.1 · p. 12
  • Benchmarks should use real-compound measurements and appropriate molecular formats, because 1D SMILES inadequately represent the 3D conformations needed for QM7, QM8, and QM9 quantum-property prediction.Suggested alternatives include ADME data curated by Fang et al. 189, the Therapeutic Data Commons (TDC)190,191, and TDC-2.192, which align more closely with drug-development outcomes.Source: §3.1 · p. 12
  • Existing messy public data may be curated into realistic benchmarks, while extracting information from chemistry papers offers a potential source of high-quality, truth-grounded data at scale.Related benchmarks include LAB-Bench197 for biology and MatText, MatSci-NLP, and MaScQA for materials science.Source: §3.1 · p. 12

3.2 Property Prediction and Encoder-only Mol-LLMs

Encoder-only Mol-LLMs convert molecular sequences into representations for property prediction and synthesis classification. Their performance benefits from large-scale pretraining, multitask or contrastive learning, and representations that capture molecular structure, while limited labeled data remains a central challenge.

3.2.1 Property Prediction

  • Property prediction is central to chemistry because molecular structure determines applications, while LLMs can reduce the computational demands of quantum-mechanical and molecular-dynamics methods.The section surveys encoder-only scientific LLMs and identifies limited labeled data as a continuing constraint, motivating augmentation and contrastive-learning approaches such as INTransformer.Source: Table 1 · p. 12, §3.2.1 · pp. 13–14, §3.2.2 · pp. 15–16

3.2.2 Encoder-only Mol-LLMs

  • Encoder-only models convert SMILES strings into latent representations that reflect chemical properties for downstream prediction and classification.BERT-based models also support synthesis classification, including reaction-fingerprint generation without separating reactants from reagents; one model achieved 98.2% accuracy versus 95.2% for an encoder-decoder model.Source: §3.2.2 · p. 14
  • Multitask learning improved generalization and interpretability, with MTL-BERT outperforming 2012 state-of-the-art methods across 60 molecular datasets from ADMETlab and MoleculeNet.MTL-BERT used large-scale pretraining and multitask learning on unlabeled ChEMBL SMILES to address limited property-labeled data and weak cross-dataset correlations.Source: §3.2.2 · p. 14
  • Mol-BERT outperformed existing sequence and graph-based methods by at least 2% in ROC-AUC on Tox21, SIDER, and ClinTox benchmarks.Mol-BERT pretrained on four million unlabeled drug SMILES from ZINC15 and ChEMBL27, treating Morgan fingerprint fragments as tokens for masked-language-model pretraining.Source: §3.2.2 · pp. 14–15
  • SolvBERT matched DMPNN for solvation free-energy prediction and outperformed GROVER on experimentally evaluated solubility for new solute-solvent combinations.Its clustering feature in pretraining, visualized with TMAP, helped the SMILES-based model capture dynamic and spatial solvation interactions.Source: §3.2.2 · p. 15

3.3 Property Directed Inverse Design and Decoder-only mol-LLMs

Property-directed inverse design uses decoder-only LLMs to generate valid, synthesizable molecules directly from desired properties, expanding exploration beyond natural products. Decoder-only mol-LLMs support generation, prediction, optimization, and increasingly target-specific design through chemical and biological conditioning.

3.3 Property Directed Inverse Design and Decoder-only mol-LLMs

  • Decoder-only GPT-like models learn molecular sequence patterns from large chemical datasets, generate novel structures under property constraints, and can be efficiently fine-tuned with limited domain-specific data.142,270Their scalability enables rapid exploration of broad chemical spaces for de novo chemistry applications.Source: §3.3 · p. 16

3.3.1 Property Directed Inverse Design

  • Inverse design directly generates candidate molecules meeting predefined properties, retaining valid and synthesizable structures while expanding exploration beyond natural-product derivatives.242,289–291This avoids iterative stepwise modification and accelerates movement from target properties toward viable therapeutic candidates.Source: §3.3.1 · pp. 17–18

3.3.2 Decoder-only Mol-LLMs

  • Decoder-only mol-LLMs advanced from SMILES generation and property prediction to target-conditioned design, with MolGPT demonstrating strong validity, uniqueness, FCD, and KL-divergence performance while balancing novelty and chemical similarity.284,302,303Taiga uses autoregressive generation and REINFORCE for property optimization, whereas cMolGPT conditions generation on protein–binder embeddings; cMolGPT’s QSAR predictions achieved a Pearson correlation above 0.75 but remained concentrated in the training subspace.46,101Source: §3.3.2 · pp. 18–19
  • LlaSMol used parameter-efficient fine-tuning, including LoRA, to achieve state-of-the-art property-prediction performance on MoleculeNet benchmarks while reducing the parameters required for adaptation.273,318–320The broader trend integrates chemical representations, biological context, active learning, contrastive learning, and specialized tools to improve molecular-design precision and interpretability.Source: §3.3.2 · pp. 19–20

3.4 Synthesis Prediction and Encoder-decoder Mol-LLMs

Encoder-decoder Mol-LLMs frame synthesis prediction and molecular generation as sequence translation, progressing from recurrent and template-based systems toward flexible transformer-based and self-correcting approaches. This evolution supports more accurate, diverse, and potentially autonomous chemical design.

3.4 Synthesis Prediction and Encoder-decoder Mol-LLMs

  • Encoder-decoder architectures translate molecular sequences for reaction-outcome prediction and synthesis-pathway generation, with SMILES Transformer demonstrating self-supervised molecular representation learning.The review uses synthesis prediction as its main context while noting broader molecular applications.Source: §3.4 · p. 20

3.4.1 Synthesis Prediction

  • Retrosynthesis is challenging because multiple reactants, mechanisms, and conditions interact, but LLMs can couple molecule selection with easier predicted routes and enable compounds chosen for both properties and synthesizability.Forward synthesis builds targets from simple substances, whereas retrosynthesis works backward from targets; the discodermolide example required 36 steps, a 24-step longest linear sequence, and 3.2% yield.Source: §3.4.1 · p. 20

3.4.2 Encoder-decoder mol-LLMs

  • Synthesis prediction evolved from GRU/LSTM models through template-based and semi-template approaches to template-free transformers, with Segler and Waller 341 reaching 95% top-10 retrosynthesis accuracy and 97% reaction-prediction accuracy.Template-based systems capture reaction transformations but depend on known templates, limiting novel-reaction prediction and requiring manual updates; semi-template methods broaden flexibility through synthons and leaving groups.Source: §3.4.2 · pp. 21–22
  • Transformer and BART-based models improved synthesis and molecular generation through template-free learning, reaction-aware metrics, transfer learning, prompt-guided diversity, Retro-BLEU, MOLGEN prefix tuning, and chemical feedback.Molecular Transformer outperformed prior algorithms, SCROP achieved 59.0% accuracy, Chemformer reached state-of-the-art sequence-to-sequence and discriminative results, and MOLGEN pretrained on 100 million molecules.Source: §3.4.2 · pp. 22–23

3.5 Multi-Modal LLMs

Multi-modal LLMs integrate textual, structural, graphical, visual, and other chemical data representations to improve molecular prediction, retrieval, interpretation, and generation. Despite progress across multimodal architectures, chemical dataset construction and evaluation remain challenging, particularly for underexplored spectral inputs and molecule-text tasks.

Input modalities

  • Chemical multimodal modeling spans 1D, 2D, and 3D representations, while spectral inputs such as NMR, IR, and mass spectrometry remain underexplored.Molecular data may appear as descriptors, SMILES, graphs, images, point clouds, grids, geometry files, or text-based PDB files, with modality distinctions based on conveyed information.Source: §3.5 · p. 23

Molecule-text integration

  • Text2Mol [331] aligned molecular representations with natural-language descriptions in a shared embedding space, adding cross-modal attention for explainable retrieval and reranking.Its paired molecule-text dataset addressed cross-lingual molecule retrieval and aimed to improve retrieval metrics for learning from chemical literature.Source: §3.5 · p. 23
  • MolT5 [330] jointly used SMILES and textual descriptions for molecular captioning and property-to-structure generation, but chemical descriptions and paired datasets require substantial domain expertise.Molecules may be described through therapeutic effects, applications, structures, or degradation pathways, making high-quality molecule-text datasets difficult to construct.Source: §3.5 · p. 24

Multimodal architectures

  • Multimodal systems combine chemical and linguistic data through separate modules, retrieval, feedback, conversation, or joint pretraining for biochemical prediction, drug editing, cross-domain generalization, and zero-shot generation.Examples include CLAMP [363], BioTranslator [364], ChatDrug [365], Joint Multi-domain Pre-training [328], and MolXPT [366].Source: §3.5 · p. 24
  • Models integrating graphs, images, text, physicochemical features, docking, or diffusion support targeted and multi-constraint molecule generation and improve prediction for complex properties.Examples include GIT-Mol [367], MolTC [368], PremuNet [369], 3M-Diffusion [370], DockingGA [371], TSMMG [372], TGM-DLM [373], and MULTIMODAL-MOLFORMER [374].Source: §3.5 · p. 24

Overall impact

  • Multimodal integration improves accuracy, interpretability, and scalability for molecular property prediction and generation, supporting drug discovery and material design.Unified linguistic, chemical, and graphical frameworks enable more informed predictions and innovative molecular structures.Source: §3.5 · p. 24

3.6 Textual Scientific LLMs

Textual scientific LLMs extract structured knowledge from unstructured scientific texts and generate domain-specific answers, labels, and new text. Their effectiveness depends on high-quality accessible data, domain-specific training, and the balance between specialization and broader generalization.

3.6 Textual Scientific LLMs

  • Scientific LLM progress remains constrained by inaccessible or non-machine-readable datasets, making open data initiatives and standardized formats important for expanding applications.High-quality and diverse datasets are also required for reliable extraction, classification, and generation.Source: §3.6 · pp. 24–25

3.6.1 Text Classification

  • LLMs enable scientific text classification and extraction, converting unstructured literature into structured information for chemistry, biology, medicine, and materials research.Applications include named entity recognition, relation extraction, document classification, and information extraction from papers and table images.Source: §3.6.1 · pp. 25–26
  • GPT-4 achieved 87.95% correctness versus 51.08% for ChemDataExtractor in band-gap extraction without training data, while domain-specific models improved specialized classification and extraction performance.ChemBERT improved precision by approximately 6% for product extraction and 5% for product-role labeling; MatSciBERT improved subject classification accuracy by 3% over SciBERT.Source: §3.6.1 · p. 25
  • Larger models did not always perform better: BioMegatron’s 345M-parameter model with a 50k-token vocabulary consistently outperformed its 800M and 1.2B counterparts on NER and relation extraction.This result indicates that domain-specific training can matter more than model size for specialized tasks, although narrower datasets may limit broader generalization.Source: §3.6.1 · pp. 25–26

3.6.2 Text Generation

  • Decoder-based models extend scientific LLMs from information extraction to generative question answering, labeling, and text production, with domain-specific models often outperforming general baselines on specialized benchmarks.HoneyBee excelled within its specialized dataset but generalized poorly elsewhere, whereas BioGPT outperformed BioBERT and PubMedBERT particularly in relation extraction and document classification; PMC-LLaMA outperformed LLaMA-2 and ChatGPT despite being approximately ten times smaller.Source: §3.6.2 · pp. 26–27

3.7 The use of ChatGPT in Chemistry

ChatGPT shows useful but uneven chemistry capabilities, with accuracy constrained by data quality, input precision, and limited higher-order reasoning. Chemistry automation can improve productivity by integrating LLM-driven decision-making with robotic synthesis, testing, and laboratory workflows, although substantial limitations remain.

3.7 The use of ChatGPT in Chemistry

  • 27% accuracy for compound name–SMILES conversion and about 31% mean relative error for octanol–water partition coefficients illustrate ChatGPT’s uneven chemistry performance.It achieved 58% hit rate for coordination-compound structural information, 100% for polymer water solubility, and 60% for molecular point groups, with strongest results on widely recognized topics.Source: §3.7 · p. 27
  • ChatGPT has been applied to chemical education, MOF synthesis and optimization, and computational materials tasks including identifying crystal space groups, generating simulation inputs, and refining analyses.Deb et al. 403 also emphasized code-generation and accessibility for non-experts, particularly in catalyst development for CO2 capture.Source: §3.7 · p. 27
  • ChatGPT’s outputs depend on precise, detailed inputs, while reproducible evaluation methods remain underdeveloped and complex reasoning likely requires additional chemical tools or agents.The review relates this limitation to Bloom’s Taxonomy, noting that LLMs and autonomous agents support lower-level tasks but do not yet replicate higher-order cognitive skills comparable to human expertise.Source: §3.7 · p. 27

3.7.1 Automation

  • AI-driven automation combines molecular generation or prediction with robotic synthesis, analysis, and testing to increase efficiency, reduce errors, and address the scale of chemical space.Examples include flow-based synthesis robots, the Chemputer, on-chip synthesis and testing, and the Automated Chemical Design taxonomy for automation and experimental integration.Source: §3.7.1 · p. 28
  • Automated chemical systems remain constrained by multi-step purification, uncertainty quantification, standardized recipes, predefined instructions, and limited ability to originate materials or enable serendipitous discoveries.Organa addresses routine laboratory work through text- and audio-mediated LLM control, instruction reasoning, experimental feedback, and written reporting, while dedicated connected tools may mitigate broader limitations.Source: §3.7.1 · p. 28

4 LLM-based Autonomous Agents

LLM-based autonomous agents are systems that continuously perceive environmental observations, make decisions, and execute actions to complete tasks. Their architecture combines trainable agent modules with environmental tools, including memory, planning, reasoning, profiling, perception, and action selection.

4 LLM-based Autonomous Agents

  • An LLM-based autonomous agent continuously receives environmental observations and executes actions to complete a provided task, with trainable decision-making components distinguished from non-trainable environmental tools.The review’s architecture includes an LLM, agent code, decision process, and modules such as memory, reasoning, planning, and profiling; the environment includes tasks, APIs, laboratories, datasets, and external code.Source: §4 · pp. 28–29

4.1 Memory Module

  • Memory stores and recalls interaction history across sensory, short-term, and long-term forms, but limited context windows can cause information loss and motivate summarization, vector compression, and external databases.Long-term memory can be episodic for previous experiences or semantic for general world knowledge; short-term memory includes task interactions and in-context learning.Source: §4.1 · pp. 29–30

4.2 Planning and Reasoning Modules

  • Planning generates action sequences while reasoning draws conclusions through logical steps; retrieval, Chain-of-Thought, feedback, multi-path methods, verification, and multi-agent collaboration extend these capabilities.Chain-of-Thought breaks complex problems into sequential intermediate steps, while ReAct and Reflexion reason after actions; Chain-of-Agents improved performance by up to 10% against an RAG baseline.Source: §4.2 · p. 30

4.3 Profiling Module

  • Profiling assigns agents roles or psychological characteristics through the LLM or agent code, enabling persona-based multi-agent systems, although handcrafted profiles provide more behavioral control than automated profiling.Profiling can also create agent populations reflecting demographic distributions, while automated profile generation may be better suited to producing large numbers of agents.Source: §4.3 · pp. 30–31

4.4 Perception

  • Perception converts multimodal text, image, and auditory observations into representations that LLMs can process, with multimodal and any-to-any models simplifying perception in autonomous systems.Examples include SAM, GPT4-V, LLaVa, Fuyu8B, BuboGPT, Next-GPT, and Macaw-LLM.Source: §4.4 · p. 31

4.5 Tools

  • Tools constitute the environment’s action space, where agents propose, evaluate, select, and execute actions; tool-use training, routers, specialized neuro-symbolic modules, and self-generated tools support increasingly complex and open-world tasks.MRKL selects specialized modules without requiring separate domain-specific datasets or costly fine-tuning, while newer systems can create tools for unseen problems and continually improve.Source: §4.5 · p. 31

5 LLM-Based Autonomous Agents in Scientific Research

This section examines LLM-based autonomous agents that use memory, planning, reasoning, tool use, and multimodal perception across scientific research tasks, including literature review, chemical innovation, experiment planning, cheminformatics, hypothesis generation, and material discovery. It also considers their increasing autonomy and accessibility alongside the continued need for human oversight, safety, and validation.

5 LLM-Based Autonomous Agents in Scientific Research

  • Agent systems address LLM limitations in stale knowledge, restricted context, hallucination, and direct tool access by integrating retrieval, domain databases, optimization algorithms, laboratory equipment, and multi-agent collaboration.These capabilities support molecular discovery, protein design, genetic research, and experimental workflows, but responsible deployment requires human oversight, safety checks, and attention to dual-use risks.47,48,501,496Source: §5 · p. 33, §5.2 · p. 36
  • SciAgent improved performance by approximately ∼20% across all five SciToolBench domains by fine-tuning retrieval and planning modules to use domain-specific tools.Other agents connected language models to Bayesian or genetic optimization, materials databases, knowledge graphs, structured data, and educational media, demonstrating broad scientific automation capabilities.Source: §5 · pp. 33–34

5.1 Agents for Literature Review

  • PaperQA achieved 87.9% precision and 69.5% accuracy on LitQA, exceeding the 66.8% human accuracy baseline in scientific question answering.502Its retrieval-augmented workflow repeatedly searches and gathers evidence before answering, while WikiCrow reduced irrelevant statements from 16% to 3% and added 5% more correct citations than human-created articles.502,503Source: §5.1 · pp. 34–35

5.2 Agents for Chemical Innovation

  • ChemCrow outperformed GPT-4 in human evaluation, achieving 9.24/10 versus 4.79/10 by combining LLM reasoning with computational and robotic chemical tools.47Chemist-X reached a maximum 98.6% yield in an HTS reaction-design task, while Coscientist semi-autonomously performed Suzuki-Miyaura and Sonogashira couplings and optimized reaction conditions.48,529Source: §5.2 · pp. 35–36

5.3 Agents for Experiments Planning

  • LLM-RDF automated synthesis from literature search through purification, producing 1 gram with 86% yield and >98% purity, while CRISPR-GPT achieved approximately 70% editing efficiency for four genes.495,533BioPlanner generated correct pseudocode for 59 of 100 procedures, whereas CALMS operated laboratory equipment and enabled automated experiments; human-in-the-loop oversight remained central for reliability and safety.505,507Source: §5.3 · pp. 36–37

5.4 Agents for Automating Cheminformatics Tasks

  • Domain-specific prompting improved Gemma-7B drug-likeness accuracy by approximately ∼60%, while ChatMOF reached ∼90% search-and-prediction accuracy and ∼70% generative accuracy.509,539ChatMOF’s genetic algorithm shifted generated surface-area distributions from 3784 m2/g to a 5554 m2/g peak after three generations without further experimental validation; CoVe increased Eunomia’s precision by ∼20%.504Source: §5.4 · pp. 37–38
  • XpertAI343 automates explainable-AI workflows, making technical chemical-data interpretation more accessible to nonexperts.It computes interpretable descriptors and SHAP or LIME explanations, searches scientific literature, and provides data-specific natural-language explanations with visual XAI plots.Source: §5.4 · p. 38

5.5 Agents for Hypothesis Creation

  • SciMON, Gu and Krenn’s framework, and CoQuest use literature or knowledge-graph retrieval, iterative refinement, and human feedback to generate new research questions [546, 547, 506].CoQuest supports breadth-first and depth-first generation through a ReAct439 agent equipped for literature discovery, proposition, refinement, and evaluation.Source: §5.5 · pp. 38–39
  • Among 20 doctoral researchers, 60% preferred CoQuest’s breadth-first strategy for greater control and trustworthiness, whereas depth-first generation scored higher for novelty and surprise.Depth-first iteration can introduce keywords that users had not considered.Source: §5.5 · p. 39
  • ChemReasoner significantly outperformed GPT-4 on reward scores when generating top-five catalysts for adsorbate, biofuel, and CO2-to-methanol conversion tasks.It combines Monte Carlo thought search with domain-specific rewards and atomistic GNN feedback predicting adsorption energies or reaction barriers.Source: §5.5 · p. 39
  • SGA generates and iteratively optimizes hypotheses such as code or molecules through search, gradient-based algorithms, and computational feedback, but its quantum-property molecular results were not validated [494].For molecule design, it targeted a specified HOMO-LUMO gap predicted with UniMol.549.Source: §5.5 · p. 39

6 Challenges and Opportunities

LLMs and autonomous agents could make chemical research more efficient through prediction, tool orchestration, automation, and human–AI collaboration. Realizing this potential requires better data and benchmarks, interpretability, domain integration, safety controls, agent evaluation, and rigorous experimental oversight.

Data Quality and Availability

  • Reliable chemical progress requires grounded, consistent datasets and practical benchmarks, because hypothetical data can distort molecular priors and MoleculeNet contains errors and weak links to real chemistry; TDC offers a more therapeutic foundation.Curated multimodal datasets from peer-reviewed literature could improve model applicability and transferability.Source: §6 · pp. 39–40

Model Interpretability

  • LLM interpretability remains limited by black-box behavior, but attention analysis, language-based explanations, prompting, and tools such as eXpertAI343 can expose chemical reasoning.Examples include linking reactant atoms to products and assigning H-NMR peaks to molecular hydrogens.Source: §6 · p. 40

Integration with Domain Knowledge and Cross-disciplinary Applications

  • Chemically valid and cross-disciplinary systems require explicit domain knowledge, specialized pretraining, chemical tokens, and cooperating specialist agents spanning chemistry, physics, electrochemistry, and materials.Tool quality and interoperability also determine whether LLM-agent frameworks can handle diverse chemical tasks effectively.Source: §6 · p. 40

Reinforcement Learning

  • Direct reinforcement learning for language-model agents is an emerging goal, with molecular design and scientific-task reward models representing steps toward agent-level optimization.Existing examples apply reinforcement learning to generative molecular design and protein language models rather than directly to language-model agents.Source: §6 · p. 40

Agent Evaluation

  • Agent evaluation lacks robust benchmarks and agreed success criteria, motivating Bloom’s taxonomy404,405 as a framework for measuring increasingly complex reasoning and tool use.A reliable metric could clarify what constitutes a “superhuman” digital chemist and guide research standards.Source: §6 · p. 40

Ethical and Safety Concerns

  • Safe deployment requires addressing bias, misuse, hallucinations, inadequate alignment, and missing specialized risk-control models through automated red-teaming, safety evaluation, and human responsibility for dual-use risks.Human evaluation alone is not scalable, while reliable safety assessments remain necessary for tool-using LLMs.Source: §6 · pp. 40–41

Human-AI Collaboration in Chemical Research

  • Human-in-the-loop collaboration is essential because hallucinations, formatting errors, reasoning failures, and disputed AI-generated structures can undermine practical reliability, as illustrated by A-Lab’s 41 materials in 17 days.Seasoned chemists should validate AI results and ensure predictions account for reagent quality, equipment limitations, and laboratory conditions.Source: §6 · p. 41

Promotion of Impactful Discoveries

  • Agents already automate data analysis, literature review, and parts of experimental design, while a human-guided PaperQA2492 and ChemCrow47 workflow could reduce a reaction investigation from 2–3 months to one week.The proposed hybrid approach combines computational analysis, targeted experiments, automated synthesis, and expert validation to improve efficiency and creativity.Source: §6 · pp. 41–42

7 Conclusions

The review concludes that LLMs and autonomous agents are reshaping chemistry and may accelerate scientific discovery, while transparency, data bias, and computational demands remain important challenges. The authors will maintain a repository tracking relevant studies.

7 Conclusions

  • The authors argue that AI in chemistry is here to stay and will increasingly become a necessary tool supporting greater progress and more jobs.Source: §7 · p. 42
  • LLMs have demonstrated potential to reshape chemical research through advances in molecular simulation, reaction prediction, and materials discovery, with important caveats.Source: §7 · p. 42
  • LLM-based autonomous agents can improve the accuracy and efficiency of chemical research while supporting innovative solutions to complex problems.The review frames agents as a central LLM connected to interchangeable components.Source: §7 · p. 42
  • Further progress depends on addressing model transparency, data biases, and computational demands to support responsible and useful deployment in chemistry.Source: §7 · p. 42
  • The authors will maintain an organized repository of new studies on LLMs and LLM-based agents for scientific purposes.The repository is available at https://github.com/ur-whitelab/LLMs-in-science.Source: §7 · p. 42

Related papers

Loading 2407.01603v3…