LLMs4All: A Review of Large Language Models Across Academic Disciplines

6.1 New Frontiers in LLMs

• The conclusion also frames new frontiers in LLMs.Source: §6 · p. 6

6.2 Beyond the Brief: LLMs for Arts, Letters, and Law

• The arts, letters, and law synthesis covers shared opportunities, common limitations, and paradigms from history to law.Source: §6 · p. 6

6.3 Signals to Strategy: LLMs for Economics and Business

• The economics-and-business synthesis covers shared opportunities, common limitations, and paradigms from finance to marketing.Source: §6 · p. 6

6.4 Models as Instruments: LLMs for Science and Engineering

• The science-and-engineering synthesis is organized around discipline probes, shared opportunities, common limitations, and LLM paradigms.Source: §6 · p. 6

6.5 Pilot the Present, Plot the Future

• The closing path forward distinguishes the current state from a future path.Source: §6 · p. 6

• The review surveys LLMs across arts, letters, and law; economics and business; and science and engineering, alongside limitations, challenges, and future directions.It is intended to guide researchers and practitioners seeking to apply LLMs in real-world work.Source: §1 · pp. 7–8
• Chapter 2 introduces LLM definitions and history, profiles major model families, and reviews evaluation tasks, benchmarks, and methods.It also aims to provide a performance-at-a-glance synthesis and practical guidance for interpreting results and choosing methods.Source: §1 · p. 7
• The review maps disciplinary tasks to a unified computational input–output framework and surveys applications, limitations, and future directions.It also discusses representative benchmarks.Source: §1 · p. 7
• The humanities-and-law chapter covers research and applications in history, philosophy, political science, arts and architecture, and law.Its topics range from historical narrative and philosophical analysis to legal question answering, drafting, document analysis, and judgment prediction.Source: §1 · pp. 7–8
• The economics-and-business chapter surveys LLM applications in finance, economics, accounting, and marketing, including evaluations and benchmarks.Topics include investment research, economic simulation, auditing, taxation, consumer analysis, and campaign design.Source: §1 · p. 8
• The science-and-engineering chapter spans mathematics, physical sciences, chemistry, life sciences, earth and civil engineering, and computing.The conclusion synthesizes cross-domain opportunities and constraints and outlines a path involving grounded attribution, tool use, oversight, and governance.Source: §1 · p. 8

2.1.1 Definition of LLMs

• LLMs are Transformer-based next-token systems trained through self-supervised pretraining and task-oriented fine-tuning, and can be categorized by domain focus or reasoning capacity.They are typically sized from billions to hundreds of billions of parameters, with scale associated with emergent capabilities [13]; domain-specific examples include BioBERT [17] and SciBERT [18].Source: §2.1.1 · pp. 9–10

2.1.2 History of LLMs

• LLMs evolved from brittle rule-based systems through statistical models and RNNs to Transformers, whose parallel self-attention improved long-range modeling and scalability.Earlier statistical models were limited by fixed context, while standard RNNs processed words sequentially and struggled with long-term dependencies.Source: §2.1.2 · pp. 10–11
• Reasoning advances use chain-of-thought to break complex problems into steps, self-consistency to select among reasoning paths, and rule-based rewards with reinforcement learning in DeepSeek-R1.DeepSeek-R1 has 671 billion parameters and is optimized for mathematics, programming, and logical reasoning.Source: §2.1.2 · pp. 12–13

2.2.1 Overview

• Closed-source models generally offer stronger performance, larger context windows, and higher throughput, while open-source models can reduce costs, latency, and data-confidentiality risks.API use can add latency and substantial expense, particularly for reasoning models; local open-source response times depend on user infrastructure.Source: §2.2.1 · p. 13

2.2.2 GPT-series Models

• GPT’s progression scaled models, introduced in-context learning and RLHF, made ChatGPT conversational, expanded to multimodal inputs, and later emphasized reasoning, controllability, and efficiency.GPT-3 reached 175 billion parameters; GPT-3.5 used RLHF, GPT-4 added image input, and GPT-5 introduced standard and extended-reasoning configurations.Source: §2.2.2 · pp. 15–16

2.2.3 OpenAI Reasoning Models

• OpenAI’s reasoning models use reasoning tokens and reinforcement learning to support structured multi-step problem solving, with configurations that trade speed against accuracy.The passage identifies advanced programming, scientific inquiry, and strategic planning as target domains.Source: §2.2.3 · p. 16

2.2.4 Claude 3 Model Family

• Claude 3’s family is guided by a constitution of behavioral principles intended to make outputs helpful, honest, harmless, empathetic, and accessible.The passage says a principle informed by public feedback aims to reduce stereotype bias and support people with disabilities.Source: §2.2.4 · p. 17

2.2.4 Claude 3 Model Family; 2.2.5 Gemini 2 Model Family; 2.2.6 Gork Model Family

• Claude models offer multimodal capabilities, with Claude 3.7 Sonnet’s adjustable extended thinking for complex tasks; Gemini 2.0 adds native tools and multimodal inputs and outputs, while Grok 3 uses reinforcement learning for extended reasoning.The section reports Claude 3.5 Sonnet as state of the art in coding and Grok 3 as scoring 1402 Elo in Chatbot Arena; Grok 3 APIs were unavailable to the public at the time of writing.Source: §2.2.4 · p. 17, §2.2.5 · p. 17, §2.2.6 · p. 17

2.2.7 GPT-OSS; 2.2.8 Llama 3 Model Family

• GPT-OSS offers open-weight MoE models for tool-using reasoning on an 80 GB GPU or 16 GB consumer devices, while Llama 3 scales from 1B to 405B parameters and uses DPO for preference alignment.GPT-OSS reports near-parity with o4-mini for gpt-oss-120b and performance comparable to o3-mini for gpt-oss-20b; Llama’s authors found DPO less computationally demanding and better-performing than PPO for large-scale models.Source: §2.2.7 · p. 18, §2.2.8 · p. 18

2.2.9 Qwen 2 Model Family; 2.2.10 DeepSeek Model Family

• Qwen 2 combines MoE with support for 29+ languages, while DeepSeek-V3 and R1 use 671B-parameter MoE models with 37B active per token and R1’s RL training cultivates self-verification and error correction.DeepSeek-V3 is reported to match GPT-4o and Claude 3.5 Sonnet; DeepSeek-R1 achieves reasoning performance on par with OpenAI o1 at lower training cost and is released under the MIT License.Source: §2.2.9 · p. 18, §2.2.10 · p. 19

2.3.1 Tasks

• The review groups LLM capabilities into text understanding, text generation, complex reasoning, and knowledge utilization, spanning tasks from sentiment detection and translation to code generation and tool-augmented retrieval.Knowledge utilization includes accessing, retrieving, and applying factual or procedural knowledge from internal memory or external sources.Source: §2.3.1 · pp. 19–20

2.3.2 Benchmarks

• MMLU tests accuracy across 57 subjects, BIG-Bench spans 204 tasks, and HumanEval assesses functional correctness on 164 Python problems.The review also describes TruthfulQA as testing truthful answers across 817 questions and GSM8K as a set of 8.5K arithmetic word problems requiring typically 2 to 8 reasoning steps.Source: §2.3.2 · pp. 20–22

2.3 Evaluation on LLMs; 2.3.3 Evaluation Methods

• Evaluation combines task-specific automatic metrics with broader measures of trustworthiness, toxicity, fairness, robustness, calibration, and reasoning quality; human review adds nuance but is costly and subjective, while LLM judges scale faster but risk bias and hallucination.The section positions these evaluation methods and benchmark results as ways to compare model strengths and trade-offs across reasoning, coding, multilingual understanding, and tool use.Source: §2.3 · p. 19, §2.3.3 · pp. 22–23

2.3.4 Performance at a Glance

• GPT-4.5 scores 92.40% on HumanEval and 96.40% on MATH; DeepSeek R1 scores 97.30% on MATH, while Claude 3.7 Sonnet (reasoner) scores 86.10% on Airline tool use and 93.20% on IF-Eval.The review concludes that no single LLM dominates all tasks: GPT-4.5 leads chatbot dialogue, Claude 3.7 Sonnet (reasoner) excels in reasoning and tool use, Claude 3.5 Sonnet leads coding, and DeepSeek R1 excels in math.Source: Table 5 · p. 23, Table 6 · p. 23, §2.3.4 · pp. 23–24

3 LLMs for Arts, Letters, and Law

• The chapter surveys LLM applications in history, philosophy, political science, arts and architecture, and law, including research, creative, analytical, and professional uses.Examples include historical interpretation, political opinion simulation, arts and architectural creation, and legal question answering, drafting, document analysis, and judgment prediction.Source: §3 · p. 25

3.1.1 Overview

• Historical research remains vulnerable to data volume, slow synthesis across perspectives, and labor-intensive methods; LLM capabilities span narrative, quantitative, and cross-disciplinary approaches.The section’s taxonomy distinguishes narrative and interpretive, quantitative and scientific, and comparative and cross-disciplinary history.Source: §3.1.1 · pp. 25–27

3.1.2 Narrative and Interpretive History

• LLMs can generate and analyze narratives, yet GPT-4 and Claude underperform humans in discourse-level reasoning; discourse-structure prompting improves emotional engagement and narrative diversity [185].On historical annotations under Structural Demographic Theory and the Big Cycle model, GPT-4 and Claude 3.5 achieve significantly higher inter-annotator agreement than humans [190].Source: §3.1.2 · pp. 27–28

3.1.3 Quantitative and Scientific History

• In quantitative and scientific history, LLMs can transcribe and summarize primary sources, process historical data, and simulate past psychological responses, though errors and anachronism remain risks.NER and classification support entity and event linking; simulations trained on historical texts may illuminate past mentalities, but available corpora can be elite-skewed and historical accuracy is difficult to benchmark.Source: §3.1.3 · pp. 28–29

3.1.4 Comparative and Cross-Disciplinary History

• Comparative-history applications include analogy generation with self-reflection to mitigate hallucinations [194] and DiscipLink’s human-AI workflow for exploring and synthesizing interdisciplinary literature [195].DiscipLink elicits exploratory questions, broadens queries with domain vocabulary, and supports researcher agency through iterative engagement.Source: §3.1.4 · p. 29

3.1.5 Benchmarks

• HiST-LLM reports 33.6%–46% accuracy across seven models, while AC-EVAL [198] finds Chinese-trained models outperform English-trained models and TimeTravel [197] exposes capability gaps.HiST-LLM performance is stronger for earlier periods and varies by region; AC-EVAL finds chain-of-thought helps larger models, while few-shot learning can add noise. TimeTravel covers 10,250 samples from 266 cultures across 10 historical regions.Source: §3.1.5 · p. 30

3.1.6 Discussion

• LLMs offer powerful new tools for historical research and novel inquiry, but factual hallucinations, deficient temporal reasoning, bias, oversimplified debates, and weak source attribution make critical human oversight important.Source: §3.1.6 · p. 31

3.2.1 Overview

• Philosophy examines fundamental questions about existence, knowledge, value, reason, consciousness, and language through reflection, logical modeling, narrative reasoning, and conceptual analysis.The overview also notes philosophical inquiry draws on human experience and wisdom and religious traditions.Source: §3.2.1 · p. 32
• LLMs can support philosophical literature synthesis, argument analysis, conceptual exploration, and dialogue, while the field’s taxonomy spans normative, analytical, and comparative approaches.The overview also identifies hallucination, oversimplification, bias, and weak source attribution as limitations.Source: §3.2.1 · pp. 33–34

3.2.2 Normative and Interpretive Philosophy

• Studies find that people may judge GPT-4’s ethical advice favorably, while LLM moral dialogue and learning remain shaped by conversational context and limits on conceptual maturity.Research also examines how behavioral cues lead people to attribute consciousness and moral agency to LLMs.Source: §3.2.2 · p. 35
• LLMs may reshape expectations of moral reasoning, but their use in normative domains remains constrained by contextual sensitivity, cultural bias, and limits to machine understanding.The cited discussion emphasizes that fluent ethical output does not itself establish genuine understanding.Source: §3.2.2 · p. 35

3.2.3 Analytical and Logical Philosophy

• LLMs both assist formal argument mapping and prompt renewed scrutiny of understanding, knowledge, and semantic grounding, while opacity and hallucination raise epistemic concerns.Their effectiveness also depends on prompt design and alignment mechanisms.Source: §3.2.3 · pp. 35–36

3.2.4 Comparative and Cross-Disciplinary Philosophy

• Comparative philosophy uses LLMs to connect philosophical paradigms and cultural perspectives, supporting comparative ontology, cultural critique, and interdisciplinary synthesis.The cited work explores consciousness attribution, psychoanalytic interpretations of disrupted metaphors, and how computational models reflect and distort meaning.Source: §3.2.4 · p. 36

3.2.5 Benchmarks

• Project Gutenberg and PhilPapers provide complementary materials for constructing philosophy-oriented evaluations, though neither is a benchmark in the strict sense.Gutenberg supplies public-domain texts used for reasoning and text-analysis tasks; PhilPapers contributes metadata for topic modeling, stance detection, and retrieval-based evaluation.Source: §3.2.5 · pp. 36–37, Table 9 · p. 36

3.2.6 Discussion

• Philosophy research faces risks from ungrounded outputs, oversimplification, Western-centric bias, and limited transparency, accountability, and lived understanding.Future work includes diverse corpora, argument mining, cross-cultural ontology mapping, dialogue systems for teaching, and research on epistemology and AI ethics; LLMs are framed as tools that extend rather than replace reflection.Source: §3.2.6 · pp. 37–38

3.3.1 Overview

• Political science uses LLMs to analyze large bodies of political text, classify sentiment and ideology, simulate survey responses, and accelerate archival research.These tools complement traditional textual, comparative, survey, formal-modeling, and computational methods, which face challenges of scale, integration, and interpreting social meaning.Source: §3.3.1 · p. 38
• Political LLMs face hallucination, temporal and institutional fragility, bias, shallow interpretation, and opaque attribution, so they should augment rather than replace disciplinary judgment.The proposed task taxonomy groups applications into political-text analysis, opinion simulation and forecasting, and generation and framing of political messaging.Source: §3.3.1 · pp. 38–39

3.3.2 Text Analysis for Political Insight

• LLMs support scalable political-text analysis, with studies reporting expert-aligned ideological scaling and strong classification, stance, and sentiment performance, including zero-shot classification surpassing expert coders and crowd workers.Few-shot prompting can also handle complex classification and topic modeling without large labeled datasets. The surveyed studies note that transparency, replicability, and clear concepts remain concerns.Source: §3.3.2 · p. 40

3.3.3 Opinion Simulation and Forecasting

• LLMs can simulate public opinion and political decisions, but representational variation across regions and demographics and bias in underrepresented groups limit generalizability.Role-conditioned prompts can produce diverse opinion distributions [245], and covariates improve estimates of population-level views on global warming; token-probability distributions also offer a method for forecasting U.S. elections [247].Source: §3.3.3 · p. 40

3.3.4 Generation and Framing of Political Messaging

• LLM-generated political messages can persuade more as model size increases, while returns diminish; tailored messages can outperform generic ones, and models can also produce persuasive election disinformation.Hackenburg et al. [248] report logarithmically increasing persuasion across 24 models and 720 U.S.-focused prompts, with diminishing returns beyond a certain scale; Matz et al. find psychological tailoring improves effectiveness. Williams et al. [251] find several models generate high-quality false election narratives that people struggle to distinguish from truth.Source: §3.3.4 · p. 41

3.3.5 Benchmarks

• Political NLP benchmarks cover ideology, rhetoric, factuality, and bias; TruthfulQA tests truthfulness and informativeness, and finds larger models may give more fluent but less truthful answers.TruthfulQA [257] contains 817 questions across 38 categories, including politics, law, and health, and tests responses to false premises and misconceptions.Source: §3.3.5 · p. 42

3.3.6 Discussion

• LLMs offer political science opportunities in discourse analysis, behavioral modeling, civic access, and faster textual coding, but hallucination, bias, shallow reasoning, opacity, and political misuse require safeguards.The discussion recommends domain-specific fine-tuning, causal and temporal modeling, bias auditing, interactive simulations, and disinformation defenses; it argues political AI should support deliberation and accountability rather than manipulation.Source: §3.3.6 · p. 43

3.4.1 Overview

• In arts and architecture, LLMs support creative and analytical workflows, but limitations in originality, cultural context, authorship, and evaluation make them assistants rather than substitutes for human expertise.Applications include co-creating visual and literary works, interpreting creative material, assisting with architectural concepts and briefs, and using multimodal inputs; the overview taxonomy also covers visual, literary, performing, and architectural practice.Source: §3.4.1 · pp. 45–46

3.4.2 Visual Art

• MiniGPT-4, LLaVA, and ArtGPT-4 use different vision-language integration strategies, and ArtGPT-4 demonstrates strong performance on ArtEmis and ArtMM for nuanced emotional and aesthetic qualities.MiniGPT-4 uses a frozen vision encoder and Q-former, LLaVA partially fine-tunes the LLM, and ArtGPT-4 adds trainable image adapters for parameter-efficient fine-tuning.Source: §3.4.2 · p. 46
• GalleryGPT grounds artwork descriptions in visual features, while CognArtive converts images to SVG-based text so language models can perform visual reasoning without dedicated visual encoders.GalleryGPT’s structured supervision is intended to counter visual hallucination driven by memorized textual knowledge.Source: §3.4.2 · p. 46

3.4.3 Literary Art

• Creative dialogue, temperature modulation, and multi-voice prompting support literary generation and critique, while human-centered rubric-based evaluation captures craft, originality, and stylistic fidelity.Interactive feedback and simulated authorial and critical roles enable stylistic refinement and examination of narrative voice and nuance.Source: §3.4.3 · p. 47

3.4.4 Performing Art

• Dramatron structures script generation across story layers, Auto-Drama enables player-driven narratives, and Improbotics uses curated human-in-the-loop dialogue to make an LLM a performance partner.Improbotics supports fluid collaboration and audience engagement, though speech-recognition errors and delayed response selection can disrupt performance.Source: §3.4.4 · pp. 47–48

3.4.5 Architecture

• LLMs support architectural workflows by generating floorplans and early-stage concepts, documenting design decisions, and coordinating tools for renovation and heritage analysis.The approaches include prompt-based ideation and decision-record generation, with human oversight retained for quality assurance.Source: §3.4.5 · p. 48

3.4.6 Benchmarks

• StyleGAN2 + ADA leads ArtBench-10, DRAFT outperforms alternatives for architectural decisions, and WenMind’s best model, ERNIE-4.0, scores only 64.3 on Chinese classical literature tasks.AIST++ also benchmarks music-conditioned 3D dance generation and reports that FACT produces more natural and synchronized movements than prior methods.Source: §3.4.6 · pp. 48–50

3.4.7 Discussion

• Artistic LLMs remain limited by weak intentionality, cultural and historical sensitivity, contested authorship, subjective evaluation, and incomplete integration of spatial, temporal, and embodied creativity.Proposed directions include culturally grounded corpora, tighter multimodal integration, and iterative human–LLM collaboration.Source: §3.4.7 · pp. 50–51

3.5.1 Overview

• The legal overview groups LLM assistance into consultant question answering, document drafting, case analysis, and case prediction, while warning of privacy risks, weak multi-hop reasoning, and unreliable retrieval.The taxonomy presents these systems as support for legal work rather than substitutes for human judgment, particularly where precedent and factual context matter.Source: §3.5.1 · pp. 53–54

3.5.2 Legal Consultant Question Answering

• LLMs can provide scalable, context-aware legal answers and broaden access to legal knowledge, but effective use requires contextual adaptation, updated legal sources, and safety mechanisms.Retrieval-enhanced explanations can improve clarity and reduce factual errors, while domain adaptation supports complex legal queries.Source: §3.5.2 · p. 55

3.5.3 Legal Document Drafting

• LLM-based drafting can generate and refine clauses, check contracts with natural language inference, and support domain-specific needs, improving document accuracy, consistency, and compliance while streamlining workflows.Prompt-driven drafting also raises questions about authorship, provenance, and enforceability.Source: §3.5.3 · pp. 55–56

3.5.4 Legal Document Understanding and Case Analysis

• LLMs automate legal-document summarization, relevance assessment, and fact extraction, enabling more scalable and consistent workflows while requiring domain-specific tuning, safety mechanisms, and responsible deployment.Applications include e-discovery and large-scale review; proposed systems combine graph neural networks with LLM reasoning or use multi-agent filtering for high-stakes retrieval.Source: §3.5.4 · p. 56

3.5.5 Legal Judgment Prediction

• PILOT uses precedent retrieval and temporal modeling, while PLJP grounds predictions in legal factors; structured reasoning and jurisdiction-specific adaptation can improve interpretability and predictive performance.ADAPT decomposes cases into facts, candidate charges, and outcomes; the studies also stress retrieval-augmented generation and jurisdictional customization.Source: §3.5.5 · p. 57

3.5.6 Benchmarks

• Gemini Pro leads LegalBench accuracy, while Gemini Flash Preview offers comparable results with lower latency and API cost; newer legal benchmarks broaden evaluation beyond narrow tasks.LegalBench, LawBench, LexGLUE, LAiW, Swiss-Judgment-Prediction, and LJP-IV span languages, legal traditions, and task formats, but benchmarks still lack procedural simulation, adversarial robustness, and cross-jurisdiction alignment.Source: §3.5.6 · pp. 57, 60, 61, Table 16 · p. 60

3.5.7 Discussion

• Legal LLMs can improve efficiency, access to justice, and complex reasoning, but hallucinations, opaque outputs, jurisdictional gaps, liability, and bias remain significant deployment risks.The review frames LLMs as assistive technologies that amplify legal insight rather than replace legal professionals.Source: §3.5.7 · pp. 61–62
• Research priorities include fact-verified retrieval, jurisdiction- and time-aware modeling, human oversight, and structured rationales to improve reliability, transparency, and accountability.Time-aware PILOT is cited as an example, and human review is especially relevant to drafting and judgment prediction.Source: §3.5.7 · p. 61
• Legal deployment also calls for governance standards, audit trails, liability attribution, and open domain-specific datasets, particularly for underrepresented legal systems.These measures are intended to address legal uncertainty and support more equitable research and applications.Source: §3.5.7 · p. 61

4.1.1 Overview

• Finance research proposes a broad LLM taxonomy, positioning these models as complements to traditional asset-pricing and risk tools that bridge financial language and data.LLMs are presented as especially suited to extracting information from unstructured documents, interpreting or generating reports, and answering domain-specific questions, but not precise numerical computation or high-frequency decisions.Source: §4.1.1 · pp. 64, 66

4.1.2 Trading and Investment

• Trading agents use layered memory and coordinated roles [370], [361], [360], [371]; GPT-3 beats sentiment dictionaries for return prediction, and LLM-enhanced RL beats baseline RL and historical benchmarks [368],, [376].TradingAgents showed superior cumulative returns and Sharpe ratios in simulation; FinCon added within-episode CVaR alerts and cross-episode risk controls [371], [360].Source: §4.1.2 · pp. 66–67

4.1.3 Corporate Finance

• In corporate finance, LLMs can complement models for investment, M&A, and insolvency forecasting,,; report-analysis work applies RAG and calculation tools to XBRL and fine-tuned models to earnings generation, [381].Multimodal approaches were also reported as more accurate for ESG data extraction.Source: §4.1.3 · p. 68

4.1.4 Financial Market Analysis

• LLM stock forecasting moves from headline sentiment toward interpretable, integrated prediction: SEP and LLMFactor explain outputs, while a QLoRA model incorporating firm and macro data beats GPT-4 [384].ChatGPT also predicted short-term returns from headlines without financial fine-tuning and outperformed traditional methods [385].Source: §4.1.4 · p. 69

4.1.5 Financial Intermediation and Risk Management

• Financial intermediation is framed as institutions channeling savers’ funds to borrowers while managing the risks of those transactions.The supplied passage defines the topic but does not describe LLM applications in this subsection.Source: §4.1.5 · p. 69
• Fraud-R1 [389] reveals persistent LLM weaknesses in role-play and cross-lingual fraud scenarios, while sentence-transformer embeddings have improved anomaly-detection classification [392].These findings motivate multilingual, adaptive, context-aware systems for realistic fraud settings.Source: §4.1.5 · p. 70
• Credit-scoring research spans instruction-tuned generalist tools [394], analyst-preferred Bayesian-guided reports, and narrative-based P2P lending scores [396], but GPT underperforms XGBoost when numerical reasoning is essential [397].The work also raises fairness and bias concerns and emphasizes generalization, explainability, and ethical oversight in credit assessment.Source: §4.1.5 · pp. 70–71

4.1.6 Sustainable Finance

• LLMs may support ESG scoring by extracting and classifying sustainability information [401] [404] or judging company performance, but long reports can expose context brittleness [406].The proposed direction is to synthesize diverse textual sources for more dynamic, context-aware scoring.Source: §4.1.6 · p. 71

4.1.7 Financial Technology

• Financial QA research finds generative LLMs promising for causal questions with few examples, while KG-RAG [413] outperforms traditional RAG in multi-document financial QA.SocialFinanceQA offers a benchmark dataset drawn from finance-focused Reddit questions and answers.Source: §4.1.7 · pp. 71–72
• LLM-powered robo-advisors can tailor recommendations and market analysis to user goals, but human advisors retain an edge in personalization.RAG is discussed as a way to improve the relevance and accuracy of robo-advisor advice.Source: §4.1.7 · pp. 72–73

4.1.8 Benchmarks

• Existing financial benchmarks are increasingly inadequate for assessing LLMs’ broad capabilities and performance in practical settings, motivating multi-task evaluations and more realistic datasets.The section presents a comparison of traditional and LLM-based methods across financial tasks and outlines these two benchmark-development priorities.Source: §4.1.8 · p. 73, Table 18 · p. 73

Multi-Task Benchmarks

• Financial benchmarks report difficulties with forecasting, decision-making, agent reasoning, and complex domain-specific tasks; R-Judge models perform at or below random, while CFinBench’s best accuracy is just over 60%.FinBen also finds strengths in information extraction and sentiment analysis, while FinEval reports weaker performance on dynamic agent tasks and Hirano finds challenges on complex tasks.Source: Multi-Task Benchmarks · pp. 73–75
• UCFE evaluates interactive financial expertise across 17 task types and finds domain-specific models outperform general-purpose counterparts in context-aware, actionable insights.The benchmark draws on a user study of 804 participants and assesses accuracy, adaptability, response depth, and user satisfaction; its findings underscore the importance of user alignment and interactivity.Source: Multi-Task Benchmarks · pp. 74–75

Financial Dataset Construction

• Raw financial, cryptocurrency, filing, analyst, news, and social data support tailored trading benchmarks; INVESTORBENCH, TradingAgents, FinMem, and FinCon combine distinct signals and modalities.The examples include OHLCV data, regulatory filings, news, social sentiment, fundamentals, and earnings-call audio.Source: Financial Dataset Construction · pp. 76–77

4.1.9 Discussion

• Finance applications use LLMs to process unstructured data, generate strategic insights, and support decisions across trading, corporate finance, markets, risk management, sustainable finance, and fintech.The review also points to multimodal inputs and natural-language reasoning as capabilities for adaptation across financial subdomains.Source: §4.1.9 · pp. 77–78
• Financial LLMs remain limited by numerical reasoning, opaque decisions, poor robustness to market shifts, long-context demands, and bias in high-stakes applications.The discussion notes that precision-focused tasks may favor econometric or hybrid approaches and that regulated uses require stronger explainability and compliance.Source: §4.1.9 · p. 77
• Proposed finance research priorities include domain-specific and instruction tuning, multimodal multi-agent systems, explainable reasoning, real-time adaptation, and fairness and regulatory compliance.The discussion also calls for ongoing integration with traditional quantitative methods, domain knowledge, and human oversight.Source: §4.1.9 · pp. 77–78

4.2.1 Overview

• The economics overview frames LLMs as extending rather than replacing traditional models, linking language with behavior across four task domains while remaining sensitive to prompts and prone to oversimplification.The domains are behavioral and experimental economics, macroeconomic simulation and agent-based modeling, strategic and game-theoretic interactions, and economic reasoning and knowledge representation.Source: §4.2.1 · pp. 79–80

4.2.2 Behavioral and Experimental Economics

• LLMs can expand behavioral economics experiments as scalable synthetic agents, but elicited behavior remains sensitive to prompts and context rather than replacing human subjects.GPT-3 replicated classic behavioral findings, with prompted social preferences shifting decisions; GPT choices were often consistent with utility maximization but sensitive to framing, while bias patterns varied by model and prompting.Source: §4.2.2 · p. 81

4.2.3 Macroeconomic Simulation and Agent-Based Modeling

• LLM-based agent models simulate macroeconomic behavior with heterogeneous, adaptive agents, including labor and consumption patterns, social influence, and policy responses across income groups.EconAgent produces macro patterns aligned with phenomena such as the Phillips Curve; Woo et al. use generated indices for context-sensitive purchasing, and Hao and Xie model socioeconomic diversity in tax responses.Source: §4.2.3 · pp. 81–82

4.2.4 Strategic and Game-Theoretic Interactions

• In strategic games, LLMs adapt to history and context and can mimic human-like behavior, but do not consistently reach Nash equilibrium and may show greater altruism and cooperation than humans.EconArena shows bounded rationality and model-dependent rule-following; Mei et al. find behavior close to human distributions, while Shapira et al. highlight language style in bargaining and negotiation.Source: §4.2.4 · p. 82

4.2.5 Economic Reasoning and Knowledge Representation

• Step-by-step prompting modestly improves performance, especially on multi-step logic, though gains vary by format; GPT-4 makes frequent errors on non-obvious economic causal links and achieves only moderate accuracy sequencing interrelated events.EconNLI evaluates causal inference between economic events, while EconLogicQA asks models to sequence interrelated events.Source: §4.2.5 · p. 83

4.2.6 Benchmarks

• GLEE, EconLogicQA, and EconNLI benchmark, respectively, language-based sequential games, multi-event economic reasoning, and causal inference grounded in economic principles.GLEE standardizes economic metrics such as efficiency and fairness across LLM–LLM and human–LLM interactions; EconLogicQA uses human-reviewed questions from business news, while EconNLI tests economic event relations.Source: §4.2.6 · p. 83

4.2.7 Discussion

• Below 0.8 F1 for fine-tuned BERT versus around 0.87 for supervised-fine-tuned LLAMA2-7B on EconNLI, the best-performing traditional encoder-only model was surpassed.The comparison illustrates LLM performance on economic causal reasoning, while the broader discussion cautions that fluent outputs do not guarantee deep economic understanding.Source: §4.2.7 · p. 84
• LLM-driven economics remains vulnerable to hallucinations, prompt sensitivity, bounded strategic depth, contextual misalignment, and ethical concerns about bias and external validity.These limitations affect multi-step reasoning, reproducibility, equilibrium behavior, realistic agent simulation, and interpretation of synthetic-agent findings.Source: §4.2.7 · p. 84
• Recommended directions include domain adaptation, standardized prompting, hybrid models, stronger reasoning methods, and ethical evaluation with validation against human data.The discussion frames LLMs as complements, not substitutes, for economic theory and empirical grounding, and calls for robust methods blending linguistic fluency with economic rigor.Source: §4.2.7 · p. 84

4.3.1 Overview

• Accounting applications target voluminous, evolving financial data through task taxonomies, while LLMs complement rather than replace traditional measurement, verification, and professional expertise.The overview identifies complex reporting and unstructured data as challenges and proposes a taxonomy aligned with disciplinary tasks.Source: §4.3.1 · pp. 85–86

4.3.2 Auditing

• 96% accuracy and an 83% reduction in verification time were reported for a government payroll audit; another firm reported 50–80% efficiency gains in audit preparation, fieldwork, and reporting.The review also describes LLM-assisted auditing for narrative analysis, disclosure extraction, documentation, and risk assessment.Source: §4.3.2 · pp. 87–88

4.3.3 Financial and Managerial Accounting

• LLM-generated disclosure summaries sharpened informativeness and aligned better with market reactions [551], while RPA integrations automate coding and reporting and 10-K language signals correlate with discretionary accounting choices.The review also discusses sustainability-disclosure benchmarking and concerns about biased or greenwashed content.Source: §4.3.3 · pp. 88–89

4.3.4 Taxation

• GPT-4-derived firm-level tax-audit measures linked audits to less tax avoidance with lasting effects, but also reduced investment and increased volatility; generic models can give legally incorrect advice [539].The review notes practitioner interest in generative AI for tax research, return preparation, and client advice.Source: §4.3.4 · p. 89

4.3.5 Benchmarks

• PLAT tests nuanced tax-penalty judgments using Korean court precedents; baseline LLMs struggle, while retrieval augmentation, self-reasoning, and multi-agent role-playing improve performance, unlike SARA’s statute-grounded U.S. tax tasks.PLAT focuses on ambiguous standards and factual contexts; SARA uses statutory reasoning on specified fact patterns.Source: §4.3.5 · p. 90

4.3.6 Discussion

• Up to 25% in costs saved and up to 40% in productivity gains are reported for Deloitte’s Zora AI in finance operations.The review presents this as a real-world example of LLMs as collaborators in accounting.Source: §4.3.6 · p. 90
• Hallucinations and weak domain-specific reasoning constrain high-stakes accounting use, alongside risks to independence, privacy, confidentiality, accountability, interpretability, and reproducibility [557].The review calls for robust validation and transparent deployment strategies.Source: §4.3.6 · p. 90
• The review prioritizes accounting-specific fine-tuning, human-centered systems, verifiable outputs, domain benchmarks, and ethical governance; it frames LLMs as enablers rather than replacements for accounting methods and judgment.It identifies human expertise as central in high-risk auditing and taxation tasks.Source: §4.3.6 · p. 91

4.4.1 Overview

• LLMs support marketing tasks involving unstructured language, but do not replace established theories or empirical methods and cannot reliably provide causal inference without structured data and external validation.The proposed taxonomy organizes tasks across marketing research, while the discussion notes limits in randomized trials and ethical judgment.Source: §4.4.1 · pp. 92–93

4.4.2 Consumer Insights and Behavior Analysis

• LLMs can analyze consumer text at scale for sentiment, themes, and patterns, addressing the scale and nuance challenges of reviews, chats, and social posts.The review also discusses synthetic respondents, noting their value for pretesting or pilots but that they are not always interchangeable with human data.Source: §4.4.2 · p. 94

4.4.3 Content Creation and Campaign Design

• LLMs accelerate content generation and personalization, shifting campaign work toward human-AI orchestration rather than replacing human creativity.The review describes a hybrid practice in which marketers orchestrate AI-assisted workflows.Source: §4.4.3 · pp. 94–95

4.4.4 Market Intelligence and Trend Analysis

• LLMs can synthesize unstructured market data for faster, forward-looking intelligence, supplementing traditional methods that may be slow and difficult to scale.The review characterizes this as a hybrid approach that retains the foundational rigor of traditional methods.Source: §4.4.4 · pp. 95–96

4.4.5 Benchmarks

• Dedicated marketing LLM benchmarks remain scarce, while public datasets offer exploratory and evaluation resources spanning commerce, search trends, news, and advertising.Examples include GA4 event exports, Google Trends, GDELT, and the Ads Transparency Report.Source: §4.4.5 · pp. 96–97, Table 25 · p. 96

4.4.6 Discussion

• 90% of marketers reported using generative AI in a 2024 AMA survey, and over 85% reported productivity gains.The survey also reports 71% weekly use, nearly 20% daily use, and half reporting improvements in both creative output quality and quantity.Source: §4.4.6 · p. 97
• Bias, hallucinations, weak causal inference, prompt sensitivity, reproducibility, and ethical concerns constrain LLM use; proposed responses include human oversight, domain adaptation, standardized evaluation, interpretability, and ethical guidance.The conclusion frames future marketing as combining LLM agility, scale, and personalization with empirical rigor, creativity, and ethical responsibility.Source: §4.4.6 · p. 97

5 LLMs for Science and Engineering

• The chapter covers mathematics; physics and mechanical engineering; chemistry and chemical engineering; life sciences and bioengineering; earth sciences and civil engineering; and computer science and electrical engineering.Topics range from proof support, molecular reasoning, and genomic analysis to geospatial tasks, code generation, verification, and synthesis.Source: §5 · p. 98

5.1.1 Overview

• The survey organizes LLM applications into proof assistance, theoretical exploration and pattern recognition, and mathematics education, against research-access barriers and concerns about reasoning, abstraction, errors, and hallucinations.It identifies lengthy research efforts, difficult cross-specialty collaboration, and the need for years of foundational study, while noting traditional education’s emphasis on procedural fluency.Source: §5.1.1 · pp. 98–100

5.1.2 Mathematical Proof Assistant

• Proof-assistant work reports accuracy gains with larger sample budgets, while other mathematical advances include reducing 4 × 4 matrix multiplication steps from 49 to 48 and progress on packing and kissing-number problems.DeepSeekMath rises from 27.5% with 128 samples to 52.0% under cumulative sampling; DeepSeek-Prover reaches 50.0% at 65,536 samples.Source: §5.1.2 · p. 101, Table 27 · p. 101

5.1.3 Theoretical Exploration and Pattern Recognition

• LLMs can support theoretical exploration by generating conjectures and mathematical statements, with FunSearch pairing a pretrained LLM and systematic evaluator to discover constructions and heuristics for open problems.The section also describes work on conjecture generation in Isabelle/HOL and combining transformer and diffusion models to forecast complex dynamical systems.Source: §5.1.3 · p. 102

5.1.4 Mathematical Education

• In a reported study, students receiving LLM-generated explanations outperform those seeing only correct answers, with the greatest learning gains when they try problems before consulting explanations.The section presents personalized explanations as a potential aid for learning, while cautioning that LLMs can produce unreliable or incorrect solutions and struggle with visual and spatial reasoning.Source: §5.1.4 · p. 102, Table 28 · p. 102

5.1.5 Benchmarks

• The benchmark overview groups mathematics datasets into competition-level, education-focused, math word problem, and mathematical reasoning categories, including MATH, AIME, and GSM8K.MATH includes 12,500 competition problems with step-by-step solutions and seven domains; GSM8K contains 8,500 grade-school word problems commonly requiring 2 to 8 solution steps.Source: §5.1.5 · pp. 102–103, Table 29 · p. 103
• Across benchmarks, DeepSeek R1 scores 97.30 in MATH, GPT-o1 scores 97.90% in MATH 500, and Grok 3 achieves 93.30% on both AIME years; most models approach ceiling performance on GSM8K.The section notes that benchmark performance varies: GSM8K distinguishes fewer models, while AIME differentiates advanced problem-solving capabilities; DeepSeek R1 is described as offering competitive performance at lower token cost than GPT-o1.Source: §5.1.5 · p. 103, Table 30 · p. 104

5.1.6 Discussion

• The discussion presents LLMs as tools to augment mathematicians and educators, while identifying weak reasoning, hallucinations, and visual-spatial difficulties as limits requiring human verification and oversight.It cautions that over-reliance in education could hinder fundamental skills and critical thinking, and that novelty and breakthrough discovery may require substantial human guidance.Source: §5.1.6 · pp. 104–105
• Future work should improve mathematical reasoning and reliability and evaluate educational tools and integration strategies that support learning without compromising conceptual understanding or fundamental skills.Proposed directions include symbolic computation, reasoning-focused datasets, self-verification, external validators, and hybrid learning environments.Source: §5.1.6 · pp. 104–105

5.2.1.1 Introduction to Physics

• Physics studies matter, energy, and their interactions across scales through theory and experiment, spanning fundamental laws, microscopic matter, and cosmic phenomena.The detection of gravitational waves confirmed a theoretical prediction and established gravitational wave astronomy.Source: §5.2.1.1 · pp. 105–106

5.2.1.2 Introduction to Mechanical Engineering

• Mechanical engineering applies physical principles to design, analyze, manufacture, and control systems through engineering science, product realization, and interdisciplinary integration.Its applications range from transportation and energy to robotics, biomedical systems, and intelligent production.Source: §5.2.1.2 · pp. 106–107

5.2.1.3 Current Challenges

• LLMs cannot derive physical laws, enforce conservation, accurately simulate coupled systems, validate models, replace physical testing, or reliably resolve material and safety constraints.Their limits include nonlinear, multiphysics PDEs; high-fidelity simulation inputs and validation; hands-on prototyping; manufacturing trade-offs; and risk or regulatory judgments.Source: §5.2.1.3 · p. 108

5.2.1.4 Taxonomy

• A task-driven taxonomy groups LLM uses into documentation, design, simulation, experimental interpretation, and STEM learning, distinguishing primary, supporting, and expert-dependent work.This functional organization is based on computational characteristics and data modalities rather than traditional disciplinary boundaries.Source: §5.2.1.4 · p. 110

5.2.2 Textual and Documentation-Centric Tasks

• LLMs can extract and summarize information from technical documents, standards, reports, and scientific literature, serving as interfaces to machine-readable engineering assets despite precision and context challenges.This role can smooth information flow across documentation, modeling, and decision-making.Source: §5.2.1.3 · p. 109, §5.2.2 · pp. 110–111

5.2.3 Design Ideation and Parametric Drafting Tasks

• LLMs can turn natural-language prompts into design concepts and parametric CAD drafts, accelerating exploration while leaving engineering expertise, simulation, manufacturability, and traceability essential.Prompt-driven systems can suggest materials and design plans or generate and iteratively modify parametric CAD sketches.Source: §5.2.3 · p. 111

5.2.4 Simulation-support and Modeling Interface Tasks

• LLMs can generate simulation code, configure simulation interfaces, and explain results, supporting early prototyping despite limits in physical correctness, multiphysics coupling, and boundary or temporal reasoning.Their assistance includes translating text descriptions into executable simulations and summarizing numerical outputs for users.Source: §5.2.4 · pp. 111–112

5.2.5 Experimental Interpretation and Multimodal Lab Tasks

• LLMs can help structure and interpret multimodal experimental data; combining textual and structural data with graph neural networks has improved material-property prediction accuracy.Laboratory datasets may include logs, numerical measurements, images, and sensor outputs.Source: §5.2.5 · p. 112
• LLMs can explain quantum-mechanical data and support multimodal robotic-task training, but interpretation accuracy and reliability remain concerns with noisy or incomplete experimental data.GenSim2 generates training data by processing and producing text, images, and other media.Source: §5.2.5 · p. 112

5.2.6 STEM Learning and Interactive Reasoning Tasks

• In STEM learning, LLMs provide adaptive teacher-student feedback, grade responses with partial credit, and power tutoring environments for inquiry and critical thinking, while alignment and reliability remain challenges.Research also explores multimodal data to support diverse learning styles.Source: §5.2.6 · p. 112

5.2.7 Benchmarks

• Engineering benchmarks and datasets remain limited in scale, accessibility, and language alignment, while CAD representations are hard for LLMs to use and code-based datasets are small and error-sensitive.A simulator-in-the-loop approach can generate language, simulation, and output triplets at scale, use solver feedback for error correction, and support supervised training and reinforcement learning with feedback.Source: §5.2.7 · pp. 113–114
• BlenderLLM [774] achieves state-of-the-art CADBench results over GPT-4-Turbo and Claude-3.5-Sonnet across spatial, attribute, and instruction metrics, with a low syntax error rate.The paper attributes this result to instruction-to-code training and self-augmentation.Source: §5.2.7 · pp. 113–114, Table 32 · p. 114
• FEABench results show Claude 3.5 Sonnet and GPT-4o perform strongly on interface and feature extraction, while all models struggle to recall physical properties and structured feature attributes.The findings indicate that recovering high-level simulation inputs is more reliable than capturing numerical structure and physical laws.Source: §5.2.7 · p. 114, Table 33 · p. 115

5.2.8 Discussion

• LLMs are emerging as intermediaries for CAD, FEA, material selection, and simulation workflows, simplifying early design and parameter exploration through natural-language commands, code generation, retrieval, and multimodal tools.The section describes their role in design–simulate–validate pipelines.Source: §5.2.8 · p. 115
• Current LLMs struggle with coupled nonlinear physics, numerical stability, unit consistency, physical constraints, and boundary conditions, making unvalidated use in mission-critical engineering risky.Even small errors in design parameters or simulation configurations can cause system failure, safety risks, or structural inefficiencies.Source: §5.2.8 · p. 115
• Promising directions include solver-augmented data generation, modular CAD tasks, multimodal tool-using agents, unified benchmarks, and physical validation, combined with symbolic reasoning, geometric logic, and expert feedback.These integrations are presented as a path from language-based assistance toward trustworthy co-creation in engineering design and modeling.Source: §5.2.8 · pp. 115–116

5.3.1 Overview

• Chemistry studies matter and its transformations, while analysis and characterization identify substances and determine their composition, structure, and morphology using experimental methods and instruments.Examples include spectroscopy, chromatography, and mass spectrometry.Source: §5.3.1 · p. 116

5.3.1.2 Introduction to Chemical Engineering

• Chemical engineering translates laboratory processes into commercial production and covers process design and control, equipment engineering, environmental sustainability, and scale-up.Its work bridges chemistry and industrial production, with pilot testing and process simulation used to assess implementation.Source: §5.3.1.2 · p. 118

5.3.1.3 Contribution of Chemistry and Chemical Engineering

• Chemistry contributes to public health through pharmaceuticals, advances materials innovation, helps reduce pollution, and supports energy transition through chemical technologies.Examples include targeted therapies, advanced materials, clean technologies, and batteries and fuel cells.Source: §5.3.1.3 · pp. 118–119

5.3.1.4 Challenges in the Era of LLMs

• LLMs cannot replace time-consuming experimental validation, may be unreliable where training data are sparse, struggle with non-smooth molecular patterns, and can produce unsafe outputs or enable misuse.Small molecular-structure changes can produce substantial property changes, a difficulty compounded by limited molecular datasets; hazardous or illegal chemical outputs remain a safety concern.Source: §5.3.1.4 · p. 120
• LLMs can narrow molecular search, generate 3D conformations, automate experiments, and support reaction prediction and cross-domain learning.Generative models address the vast chemical search space; geometry-aware LLMs support conformation generation, while autonomous agents can plan and execute experimental workflows. LLMs are also presented as useful for complex reaction prediction and multi-task learning.Source: §5.3.1.4 · pp. 120–121

5.3.1.5 Taxonomy

• The taxonomy links established chemistry tasks—including property prediction, inverse design, reaction prediction, synthesis-condition optimization, and text mining—to a unified input–output modality framework.The framework is intended to support model development, benchmarking, and comparison across tasks.Source: §5.3.1.5 · p. 121

5.3.2 Chemical Structure Textualization

• Chemical structure textualization maps molecular structures, often encoded as SMILES, into human-readable descriptions that help communicate, search, and integrate chemical information.Examples include generating a compound’s IUPAC name or a plain-language description of an unknown substance.Source: §5.3.1.5 · p. 121, §5.3.2 · p. 124
• MolT5 [882], MolFM [886], GIT-Mol, MolCA [894], and GraphT5 [892] trace progress toward richer and more efficient molecular captions.Approaches evolve from sequence-based translation to multimodal molecular inputs, with tradeoffs in caption fidelity, model size, deployment cost, and pretraining-data coverage.Source: §5.3.2 · p. 124
• Struct2IUPAC converts SMILES to IUPAC names with 98–99% accuracy, rivaling OPSIN, while MolT5 [882] generates concise molecular-feature summaries.The review presents these results as evidence that LLMs can translate between structural representations and human-readable chemistry.Source: §5.3.2 · p. 125
• Naming models can be opaque and need updates as IUPAC standards evolve; captions risk unsupported claims and uncertain levels of detail.The proposed directions include rule-based validation, structured knowledge retrieval, multimodal inputs, and property predictors to improve correctness and reliability.Source: §5.3.2 · p. 125

5.3.3 Chemical Characteristics Prediction

• 77-million-molecule pretraining gave ChemBERTa-2 performance comparable to state-of-the-art MoleculeNet models; SELFormer led several tasks, and MoLFormer reported 60x energy efficiency.SMILES-only SolvBERT also achieved performance competitive with, or superior to, GNN approaches for solvation free energy and solubility.Source: §5.3.3 · p. 127
• T5Chem handles multiple reaction tasks in one text-to-text model; Yield-BERT achieved R2 values exceeding 0.90 on high-throughput coupling datasets.Egret improved R2 on several benchmarks through condition-based contrastive learning; ReactionT5 [912] and ReaMVP explored reaction-level pretraining and multimodal information for yield prediction.Source: §5.3.3 · p. 128

5.3.4 Chemical Structure Prediction and Tuning

• Mol2Mol work spans reaction outcomes, retrosynthesis, molecule optimization, and pathway completion; the Molecular Transformer achieved 95% top-1 accuracy on USPTO.The review notes that SMILES augmentation can reduce overfitting, while GPT-3.5 and GPT-4 lag domain-trained models in zero- or few-shot precise structural prediction [1016].Source: §5.3.4 · pp. 129–130
• LLMs make routine reaction prediction highly reliable, while multi-step reactions, quantitative yields and selectivities, and out-of-domain chemistry remain difficult.Synthesis planning has shifted from recurrent and template-driven systems toward data-trained models; molecule tuning seeks related analogs with improved properties while retaining a core scaffold, but generalization to unseen property combinations remains challenging.Source: §5.3.4 · pp. 131–132

5.3.5 Chemical Text Mapping

• LLMs enable chemical text mining with minimal or no labeled data, combining extraction and numerical prediction in unified workflows, but hallucinations and long-context limits remain.The review frames this work as transforming narrative chemical descriptions into categorical and numerical data; retrieval and reasoning chains can help with long documents and cross-sentence dependencies, while complex ontologies remain a challenge.Source: §5.3.5 · pp. 132–133

5.3.6 Property-Directed Chemical Design

• LLMs support property-directed molecular generation from natural-language requests, but generated candidates still require chemical validation and evaluation.MolT5 and few-shot MolReGPT can translate descriptions into candidate structures, narrowing the search space and supporting multi-objective generation; however, unfamiliar property relationships and invalid molecules remain concerns, and success often requires predictive models or experiments. (MolT5 [882]; MolReGPT [884])Source: §5.3.6 · pp. 134–135

5.3.7 Chemical Knowledge Narration

• Chemical knowledge QA has advanced through instruction-tuned and multimodal models, though answers can be incomplete and knowledge remains bounded by training data.LlaSMol [1041] outperformed GPT-4 on several chemistry benchmarks, while ChemVLM supports chemical OCR and reasoning over image and text cues; the review also identifies chemical QA and education as applications of knowledge narration.Source: §5.3.7 · p. 136
• Fine-tuned models reached approximately 69%–95% exact accuracy across five chemical text-extraction tasks using minimal annotated data.The section also describes LlaSMol, fine-tuned on SMolInstruct’s over three million samples across 14 chemistry tasks [1041], and models that link entities and relations in materials-chemistry text.Source: §5.3.7 · pp. 136–137
• LLMs can standardize procedures, summarize reactions, and convert between SMILES and IUPAC names, but may hallucinate, exceed context limits, or mishandle nested and multistep reaction descriptions.Few-shot and in-context learning can support adaptation with minimal examples, reducing researchers’ time spent editing protocols, summaries, and electronic lab notebooks.Source: §5.3.7 · p. 137

5.3.8 Benchmarks

• Chemistry benchmarks mainly cover property prediction, reaction outcomes or classification, and molecule generation, while multistep synthesis, reaction-condition optimization, and 3D conformer reasoning remain underrepresented.The review notes dataset bias toward drug-like molecules and patent reactions, with inorganic chemistry, environmental pollutants, and negative-result reporting less represented.Source: §5.3.8 · pp. 137, 141

5.3.9 Discussion

• +0.11 average ROC-AUC across MoleculeNet (0.67 to 0.78) followed ChemBERTa-2 scaling from 0.1 M to 10 M molecules, while GCN/GAT baselines plateaued near 0.70 [900].Struct2IUPAC achieved 98.9% correct SMILES-to-IUPAC conversions, halving OPSIN’s residual error (1% vs 3%) on comparable benchmarks.Source: §5.3.9 · pp. 142–143
• 90.4% top-1 accuracy on USPTO product prediction for Molecular Transformer exceeded RetroComposer’s 65.9% by 24 points.The review attributes route-planning gains to transformer self-attention over long-range dependencies between synthesis steps.Source: §5.3.9 · p. 143
• 91.25% overall message-level accuracy and 92.25% field-level accuracy were reported for LLaMA-2-7B structured reaction records, versus 88.9% F-measure for a CHEMDNER-patents pipeline.The model was fine-tuned on 100,000 USPTO reaction processes to generate Open Reaction Database records and identify values such as temperature, time, and yield.Source: §5.3.9 · p. 143
• 100% syntactic validity under SELFIES-based sampling for Adapt-cMolGPT compared with 88–94% for earlier SMILES-based GPT decoders.The review also describes inverse-design tools as reducing trial-and-error cycles from weeks to minutes.Source: §5.3.9 · p. 143
• Chemical LLMs face a validation gap and often lack explicit mechanistic reasoning, limiting trust in predictions and reliability on complex multistep reactions [1082] [923].The review also identifies representation mismatches between linear token sequences and molecular topology or 3D geometry, alongside data bias, brittleness, computational cost, hallucination, and limited interpretability.Source: §5.3.8 · pp. 141–142, §5.3.9 · p. 144
• LLMs may generalize poorly to novel or sparsely represented chemical spaces, limiting reliability in cutting-edge research and niche industrial applications [1083].Performance depends heavily on the breadth and quality of training data.Source: §5.3.9 · p. 144
• Chemical hallucinations can produce invalid or practically unsynthesizable molecules, while inadequate systematic validation undermines trust in real-world synthesis planning [1084].Rule-based filters partially mitigate the problem but do not resolve the validation gap.Source: §5.3.9 · p. 144
• Specialized applications are constrained by the cost or impracticality of producing high-quality annotated datasets for every niche subfield, absent robust few-shot or low-data learning [1085].Source: §5.3.9 · p. 144
• Hybrid LLM–mechanistic frameworks and multimodal representations could combine chemical theory with language understanding and capture 3D and electronic structure.The proposed inputs include SMILES, molecular graphs, and spectroscopic or crystallographic data.Source: §5.3.9 · p. 144
• Closed-loop experimental pipelines could enable rapid hypothesis testing and feedback-driven refinement, while data-efficient fine-tuning could improve performance in sparse domains.Suggested data-efficient methods include transfer learning, few-shot prompting, and synthetic augmentation of underrepresented reaction data.Source: §5.3.9 · p. 144
• Explainability, uncertainty quantification, and governance-first deployment could provide confidence metrics and mechanistic rationales while supporting validation, transparency, and responsible use.The proposed governance measures include model cards and ethical guidelines.Source: §5.3.9 · p. 144

5.4.1.1 Introduction to Life Sciences

• Life sciences study living organisms and processes from molecular to ecosystem scales, producing knowledge that underpins health, agriculture, and environmental conservation.Research spans tasks from genetic analysis and cellular biology to physiology, medicine, ecology, and evolution.Source: §5.4.1.1 · p. 145

5.4.1.2 Introduction to Bioengineering

• Bioengineering combines biological understanding with engineering design to create practical products, including therapies, devices, biomaterials, engineered tissues, and scaled biological processes.Its domains also use computational modeling and biological-data analysis alongside laboratory methods.Source: §5.4.1.2 · pp. 146–147

5.4.1.3 Current Challenges

• LLMs cannot replace human expertise in ethical decisions, physical experiments, or the specialized modeling and analysis needed for complex biological systems and heterogeneous data.They may assist with protocols or hypotheses, but novel data acquisition and the final validation of biological or engineered solutions still require laboratories, clinical trials, or real-world testing.Source: §5.4.1.3 · p. 148
• LLMs can support literature synthesis, biological sequence annotation and mutation-effect prediction, and generation of biologically plausible sequence candidates for engineering.ProGen [1244] generated protein sequences with predictable functions across protein families; some generated sequences have been experimentally shown to fold and function, but designs still require laboratory testing.Source: §5.4.1.3 · pp. 149–150
• Reported applications include oncology-note summarization, prediction of synergistic drug pairs in rare tissues, radiology-report generation, patient risk estimation, and expert-level biomedical question answering.The passages describe these capabilities but provide no quantitative comparisons.Source: §5.4.1.3 · p. 150

5.4.1.4 Taxonomy

• A data-centric taxonomy classifies life-science and bioengineering tasks by computational characteristics to clarify model compatibility, cross-domain transfer, and multimodal pipeline design.It makes modality differences explicit and supports composite LLM roles such as generation, orchestration, and interpretation.Source: §5.4.1.4 · p. 150
• The taxonomy distinguishes sequence-based, structured/numeric, textual-knowledge, and hybrid predictive tasks; LLMs are naturally suited to text, can interpret sequences, and may bridge numeric work by generating code, while complex numeric optimization favors specialized algorithms.Hybrid predictive tasks may use LLMs to interpret outputs or orchestrate specialized tools across modalities.Source: §5.4.1.4 · p. 151

5.4.2 Genomic Sequence Analysis

• Sequence models capture long-range dependencies and sparse functional patterns to annotate genomic regions and estimate CRISPR off-target mutations, reducing sequence-analysis time and cost.In the human genome, protein-coding regions make up about 1% of DNA, while promoters are also rare; models trained on massive sequence datasets can identify such regions.Source: §5.4.2 · p. 152
• DNA language models have progressed from short-window CNNs to long-context models that predict regulatory features, non-coding variant effects, and downstream tasks, but face limits in length, multi-omic integration, rare-variant generalization, and clinical interpretability.Future directions include more efficient ultra-long-sequence architectures, cross-modal biological data, and alignment with mechanistic knowledge.Source: §5.4.2 · p. 153
• RNA language models learn structural and functional features from sequences, improving predictions of splicing, RNA-protein interactions, and post-transcriptional regulation, but struggle with novel classes, evolutionary and tertiary information, and biologically meaningful explanations.The review calls for more diverse datasets, multi-resolution structural priors, and experimentally validated functional annotations.Source: §5.4.2 · pp. 154–155

5.4.3 Clinical Structured Data Integration

• Clinical language generation turns structured records into notes, radiology impressions, patient-friendly text, and conversation transcripts, reducing documentation burden while requiring factual accuracy and contextual coherence.The review identifies limited cross-institution and cross-language generalization, clinical safety, and nascent multimodal integration as challenges.Source: §5.4.3 · pp. 155–156
• 3.1 ROUGE-L improvement on discharge-summary generation was reported for ClinicalT5, while discharge-summary rewrites reached readability scores 18% higher than physician-authored versions.ClinicalT5 outperformed long-text baselines such as BART [1343], but its single-center English inpatient-note training data limits generalization; the rewrites targeted a 6–8th grade reading level.Source: §5.4.3 · p. 156
• Med-BERT, pretrained on 28,490,650 patients, surpassed baseline deep-learning models by over 20% in AUC with small fine-tuning datasets and matched models trained on datasets ten times larger.EHR prediction uses longitudinal records for tasks such as disease-onset prediction, risk stratification, and treatment planning; BEHRT models temporal dependencies in coded records, while GatorTron scales clinical representation learning.Source: §5.4.3 · pp. 156–157
• EHR-focused LLMs use medical codes, temporal data, and multimodal signals to forecast disease onset, treatment outcomes, and patient risk, including in low-resource settings.Long sparse timelines, interpretability, and institutional domain shifts remain obstacles; proposed directions include integrating imaging and genomics and improving temporal reasoning across fragmented records.Source: §5.4.3 · p. 157

5.4.4 Biomedical Reasoning and Understanding

• Med-PaLM 2 reached about 86.5% USMLE accuracy, near expert physicians’ 87%; GPT-4 remains competitive zero-shot, while Huatuo-2 reportedly surpassed GPT-4 on specialist and licensing exams.Biomedical question answering ranges from domain-pretrained models to instruction-tuned and reasoning-focused systems, but hallucination and clinical deployment risks remain.Source: §5.4.4 · pp. 158–159
• Biomedical language models support NLI, classification, and reading comprehension, with domain adaptation and instruction tuning improving contextual reasoning across specialized texts.Challenges include clinical negation, long fragmented records, and interpretability; future work should improve cross-domain generalization and ontology-grounded reasoning.Source: §5.4.4 · pp. 159–160

5.4.5 Hybrid Outcome Prediction

• BAITSAO’s LLM-derived drug and cell-line embeddings, multitask pretraining, and transfer learning outperformed classical, tabular, and transformer-based models on benchmark datasets.Drug synergy prediction can exploit contextual embeddings, attention, and prompt-based methods, but interpretability and cross-dataset consistency remain open challenges.Source: §5.4.5 · p. 161
• AlphaFold achieved an average GDT_TS of 92.4 at CASP14, and protein LLMs have since expanded from structural prediction toward multimodal generation and protein design.Generative and integrated tools support sequence creation from functional descriptions and automated design-build-test-learn engineering workflows.Source: §5.4.5 · pp. 163–165

5.4.6 Benchmarks

• DNA language models can approach expert-method performance on some BEND tasks, but struggle with long-range dependencies and vary in modeling gene structure and non-coding features.BEND evaluates DNA models on seven human-genome tasks covering functional elements at different sequence lengths.Source: §5.4.6 · pp. 165–166
• Across BEACON’s 13 RNA tasks, single-nucleotide tokenization and ALiBi positional encoding perform well across multiple tasks, motivating the lightweight BEACON-B baseline.BEACON evaluates RNA structural analysis, functional studies, and engineering applications using sequence-level and nucleotide-level classification and regression tasks.Source: §5.4.6 · p. 166
• Biomedical QA benchmarks assess factual recall, reasoning, comprehension, and contextual synthesis across exam-style and literature-grounded tasks.MedQA and MedMCQA use multiple-choice medical exam questions, whereas PubMedQA asks yes/no/maybe questions grounded in PubMed abstracts.Source: §5.4.6 · pp. 167–169
• Medical QA benchmarks are concentrated in English, though benchmarks also exist in other languages, including Chinese.Some benchmarks provide only question stems and answer choices, while others supply reference documents for contextual answering.Source: §5.4.6 · pp. 167–168

• CancerGPT evaluates drug-pair synergy prediction across seven rare cancer tissue types and compares results with larger models such as GPT-3 [7].Source: Drug Synergy Prediction · p. 169
• Even with fewer parameters, LLMs can effectively predict drug-pair synergies in scarce-data contexts, according to the evaluation framework.The evaluation includes few-shot and zero-shot scenarios.Source: Drug Synergy Prediction · p. 169
• BAITSAO combines regression and classification tasks using drug-pair, cell-line, and LLM-embedding inputs, then evaluates held-out data with multiple metrics.Its multi-task pretraining captures single-drug inhibition and pairwise synergy; ablations examine embeddings, training-data scale, and scaling laws.Source: Drug Synergy Prediction · p. 169
• For BAITSAO classification, drug embeddings are averaged and concatenated with cell-line embeddings, and Loewe-score thresholds define synergy labels.The setup tests generalization to unseen drug–cell-line combinations, including out-of-distribution samples.Source: Drug Synergy Prediction · pp. 169–170

Protein Modeling

• TAPE covers five protein-sequence tasks and finds pretraining improves performance on nearly all, although traditional non-neural methods still lead in some cases.Source: Protein Modeling · p. 170
• PEER benchmarks 17 protein-sequence tasks across five categories using curated splits to assess generalization across sequence-based methods.Source: Protein Modeling · p. 170
• Across life-science task families, domain-trained transformers outperform classical baselines; long-context models improve enhancer or eQTL correlations by 20–40% over CNN/RNN hybrids.The review also describes gains in clinical generation, EHR prediction, drug-synergy prediction, and protein-sequence activity.Source: Protein Modeling · p. 170
• Limitations include degrading accuracy on ultra-long genomic context, open distal-linkage and rare-variant problems, ad-hoc multimodal fusion, biased datasets, and interpretability concerns.Source: Protein Modeling · p. 170

5.4.7 Discussion

• The proposed research agenda calls for diverse multimodal data, biology-aware architectures, rigorous validation, and responsible integration with biological priors and experimental platforms.The conclusion frames hybrid systems as a way to bridge computational prediction and empirical validation.Source: Protein Modeling · pp. 170–171, §5.4.7 · p. 172
• Tokenizable biological data supports shared LLM representations across genomic, chemical, and medical substrates, unlike traditional size-specific architectures requiring manual features.The discussion gives k-mers, SMILES, and ICD codes as examples of tokenizable representations.Source: §5.4.7 · p. 171
• HyenaDNA processes 1 Mb genomic windows in one forward pass to learn enhancer–promoter logic, avoiding fragmented, manually curated pipelines.Source: §5.4.7 · p. 171
• Med-PaLM 2 handles note drafting, coding, and instruction rewriting in one prompt; the discussion also highlights weak experimental grounding and ethical risks.ProGen2-generated incretin peptides had a 7% activity success rate, and CRISPR-GPT missed over 30% of off-target sites in complex whole-genome data.Source: §5.4.7 · pp. 171–172

5.5.1.1 Earth Sciences Introduction

• Earth sciences study Earth’s structure, processes, history, and surrounding systems, spanning the solid Earth, water, atmosphere, and space.These domains support applications including disaster prediction, climate modeling, resource exploration, and environmental protection.Source: §5.5.1.1 · pp. 172–173

5.5.1.2 Introduction to Civil Engineering

• Civil engineering manages and builds infrastructure through tasks including structural optimization, health monitoring, material behavior prediction, and built-system modeling.Traditional approaches include mathematical and numerical optimization, smart-sensor anomaly detection, material-specific experimental or numerical models, and machine-learning methods such as SVMs and Graph Neural Networks.Source: §5.5.1.2 · pp. 173–174

5.5.1.3 Current Challenges

• LLMs cannot replace physics-based modeling, missing field measurements, physical validation, cross-scale integration, or human judgment in Earth science and civil engineering.Nonlinear, uncertain systems require high-resolution simulation and empirical calibration; sparse or noisy data, inaccessible environments, experiments, and regulatory or ethical trade-offs further limit text-based systems.Source: §5.5.1.3 · pp. 174–175
• LLMs can assist with literature and standards synthesis, field-report interpretation, sensor insights, and conceptual design, but not replace experts or physical experimentation.The section identifies documentation, communication, and early-stage design as useful roles, with future work relying on collaboration among people, physical models, and AI support.Source: §5.5.1.3 · p. 176

5.5.1.4 Taxonomy

• A computational taxonomy distinguishes LLM-compatible tasks from numerical simulation or high-dimensional optimization, revealing shared structures and opportunities for hybrid workflows.The framework groups tasks by computational characteristics to clarify model fit, support cross-domain workflow reuse, and let LLMs coordinate, explain, or augment scientific and engineering pipelines.Source: §5.5.1.4 · p. 176

5.5.2 Geospatial and Environmental Data Tasks

• LLMs are advancing from geospatial task operators toward agents that select data and tools, generate executable programs, and address user requirements, improving processing efficiency.Current limitations include limited domain knowledge, inferior data quality, and difficulty with complex sequential reasoning.Source: §5.5.2 · pp. 179–180

5.5.3 Engineering Simulation and Physical Modeling Tasks

• LLM-assisted simulation remains underexplored and is largely limited to simple modeling cases, with transparency and interdependent requirements posing further challenges.LLMs are becoming practical as assistants in early prototyping, but physical simulation remains a supporting rather than core role.Source: §5.5.3 · p. 180, §5.5.8 · p. 185

5.5.4 Textual and Document-Centric Tasks

• Domain-specific LLM adaptation supports geoscience question answering and document extraction, while compliance-checking and document workflows remain vulnerable to hallucination and inadequate training data.Methods include retraining on geoscientific records, specialized pretraining, and combining text classification, information extraction, and compliance checking.Source: §5.5.4 · pp. 180–181

5.5.5 Monitoring and Predictive Maintenance Tasks

• LLMs have been applied to energy optimization and digital-twin monitoring, but maintenance research remains preliminary and constrained by training-data quality and limited transferability.The review calls for exploiting few-shot learning and exploring additional application scenarios.Source: §5.5.5 · p. 181

5.5.6 Design and Planning Tasks

• LLMs support structural and urban planning through natural-language design tools and planning agents, but these applications remain preliminary and complex sequential actions are a challenge.The review also notes that building layout planning and design rendering remain unexplored.Source: §5.5.6 · p. 182

5.5.7 Benchmarks

• Benchmarks show satisfactory ChatGPT and LLaMA performance on text tasks, LLaVA’s advantage on image tasks, and CodeGemma’s geoscience code-generation capability.Existing benchmarks do not adequately evaluate LLMs as interfaces for knowledge retrieval, data analysis, and numerical simulation; broader evaluation is needed.Source: §5.5.7 · p. 184

5.5.8 Discussion

• LLMs can lower barriers to technical knowledge, accelerate documentation, support early design and simulation, and connect domain experts with computational tools.These capabilities may enable faster prototyping and interdisciplinary collaboration, but numerical physical modeling requires precision and physical fidelity that current LLMs lack.Source: §5.5.8 · pp. 184–185
• LLMs cannot resolve sparse, uncertain field data or replace expert oversight for opaque, safety-critical, regulatory, and ethically contested decisions.Promising directions include domain-specific fine-tuning, hybridization with physical simulators, domain-constrained agents, digital-twin integration, and human-centered safeguards; realizing the potential requires empirical observation and expert validation.Source: §5.5.8 · p. 185

5.6.1 Overview

• Traditional programming and circuit-design methods face challenges with scalability, code complexity, debugging time, human error, complex-circuit verification, long design cycles, and specialized expertise; LLM applications span six coding and hardware-design task categories.The section focuses on software development and circuit design, including code generation, debugging, codebase analysis, HDL generation, functional verification, and high-level synthesis.Source: §5.6.1 · p. 189

5.6.2 Code Generation Tasks

• Copilot reduced average task-completion time by 55.8% among developers who completed the study task, from 160.89 minutes in the control group to 71.17 minutes in the treatment group.Peng et al. [1547] studied 95 developers; the improvement had p = 0.0017 and a 95% confidence interval of [21%, 89%], and remained robust when four control-group outliers were excluded.Source: §5.6.2 · p. 189

5.6.3 Code Assistant in Debugging

• LLMs assist debugging by interpreting code and error messages, localizing faults, explaining reasoning, and proposing fixes; UniDebugger [1551] uses multiple LLM agents collaboratively to diagnose and repair bugs.Combining static analysis with LLM reasoning can improve explainability and accuracy in crash-fault localization [1553], while COAST [1555] uses communicative-agent strategies to improve detection and correction of subtle logical errors.Source: §5.6.3 · pp. 189–190

5.6.4 Code Analysis in Codebases

• LLMs can explain code snippets, summarize code functionality, and suggest or generate refactorings to help developers navigate, understand, and maintain complex or unfamiliar codebases.Source: §5.6.4 · p. 190

5.6.5 Hardware Description Language Code Generation

• LLMs can translate natural-language hardware descriptions into Verilog or VHDL, while feedback-driven systems such as AutoChip [1559] iteratively correct generated HDL using compiler errors and simulation messages.Related work explores HDL assistance, debugging, and testing; AutoSVA [1573] uses formal verification to generate assertions, and VerilogReader generates tests aimed at code-coverage convergence.Source: §5.6.5 · p. 190

5.6.6 Functional Verification

• LLMs can generate testbenches, assertions, and test stimuli for circuit verification, with LLM4DV outperforming constrained-random testing on straightforward DUT scenarios but losing effectiveness as tasks grow more complex.AutoBench generates testbenches from a design-under-test description, AssertLLM generates assertions from specifications in three phases, and LLM4DV interactively elicits test stimuli.Source: §5.6.6 · pp. 190–191

5.6.7 High-Level Synthesis

• HLSPilot automates high-level application acceleration on hybrid CPU-FPGA architectures and achieves performance comparable to manually crafted counterparts, sometimes outperforming them.It applies C-to-HLS optimization strategies to code patterns through in-context learning with exemplary C/C++-to-HLS prompts; natural-language hardware specifications may also help make design more accessible and shorten design cycles.Source: §5.6.7 · p. 191

5.6.8 Benchmarks

• Claude 3.7 Sonnet tops CodeMMLU (67 %), Aider-Polyglot (64.9 %), and SWE-Bench (62.3 %); Claude 3.5 Sonnet leads HumanEval (93.7 %), while GPT-o1 leads EvalPlus (89 %).GPT-o1 is second on HumanEval (92.4 %), and Claude 3.5 Sonnet is second on EvalPlus (81.7 %); Llama 3.3 scores 88.4 % on HumanEval and 43.9 % on CodeMMLU. HumanEval evaluates functional correctness on 164 programming problems via PASS@K, while LiveCodeBench includes self-repair and code-execution prediction and addresses data contamination.Source: §5.6.8 · pp. 192–194
• Claude 3.5 Sonnet and GPT-o1 lead general coding pass rates; Claude 3.7 Sonnet leads analysis and multilingual engineering and large-scale debugging and circuit-design tasks, while GPT-4o tops PICBench.GPT-4o, Qwen 2.5, and Llama-3 70 B are cost-efficient coding runners-up; Gemini 2.0 Flash and DeepSeek R1 are mid-tier analysis and multilingual options, and GPT-4o is the most balanced all-rounder.Source: Table 60 · p. 194, §5.6.8 · p. 194

5.6.9 Discussion

• In software, LLMs can automate repetitive coding, assist static and dynamic debugging, and explain complex codebases, potentially accelerating development and improving code quality and developer satisfaction.These capabilities may save developers time for more complex work and facilitate maintenance and collaboration [1546] [1549] [1550] [1551] [1556] [1557] [1558].Source: §5.6.9 · p. 194
• In digital circuit design, LLMs can generate HDL from natural-language specifications, help create testbenches and assertions for functional verification, and support high-level synthesis [1560] [1562] [1568] [1569].These applications could reduce verification bottlenecks, shorten design cycles, and make hardware design more accessible to software engineers.Source: §5.6.9 · p. 194
• Software code correctness and reliability remain challenging, with complex logic requiring human oversight and LLMs struggling with long-term dependencies and hallucinated code [1615].These limitations affect trustworthiness in software development.Source: §5.6.9 · p. 195
• Current LLMs lack the expertise for fundamental hardware architecture and complex-IC physical layout; novel or safety-critical verification may exceed their reasoning, analog design is nascent, and hardware data is scarce.Further work should improve code and verification reliability for critical applications, integrate LLMs with existing workflows and EDA tools, explore advanced design and optimization, and address hardware data scarcity.Source: §5.6.9 · p. 195

6 Conclusion: Navigating the Present, Shaping the Future

• The conclusion proposes a practical, auditable blueprint for cross-disciplinary LLM adoption, spanning multimodality, verifiable grounding, constrained tools and agents, calibrated decision support, human oversight, governance, and safety education.The blueprint also calls for temporal and causal adaptation and reproducible, rule-governed agent-based simulation.Source: §6 · p. 196

6.1 New Frontiers in LLMs

• Five research frontiers span architecture, training, knowledge and tools, domain customization, and responsibility, with models shifting toward efficiency, structured reasoning, grounded action, and domain-specific trustworthiness.The review frames these developments as advancing from core design through deployment and governance.Source: §6.1 · pp. 196–199

6.2 Beyond the Brief: LLMs for Arts, Letters, and Law

• Across history, philosophy, politics, arts, architecture, and law, LLMs can support interpretation, analysis, and creative production, but face recurring limits in reasoning, grounding, interpretability, and originality.The review organizes cross-disciplinary use around role play, understanding, generation, education, and human-in-the-loop collaboration.Source: §6.2 · p. 199

6.2.1 Shared Opportunities

• Humanistic and social-science applications include document analysis, logical interpretation, social-behavior simulation, complex decision support, and content ideation.Examples range from historical analogical and legal reasoning to voter-behavior modeling, art creation, and legal drafting.Source: §6.2 · p. 199, §6.2.1 · pp. 199–200

6.2.2 Common Limitations

• Common limitations include shallow causal reasoning, weak historical and cultural grounding, derivative creativity, opaque justifications, and risks of bias and misuse.These shortcomings expose a gap between fluent output and the epistemic, cultural, and ethical demands of humanistic and social inquiry.Source: §6.2.2 · pp. 200–201

6.2.3 LLM Paradigms from History to Law

• Five adoption paradigms are role-play, knowledge-augmented expertise, multimodal learning, education and training, and ethically constrained generation; role-play requires attention to fidelity, flexibility, and responsibility.Role-playing can simulate historical, political, and artistic figures for inquiry and pedagogy, but must address persona consistency, knowledge gaps, and coherence.Source: §6.2.3 · pp. 201–202
• LLMs can augment humanities and social-science expertise through cross-disciplinary reasoning, multimodal evidence synthesis, and interactive learning, but learners need critical framing.Examples include historical analogies, philosophical and legal reasoning, and tutoring or exercises; proposed safeguards include reflection prompts, curated retrieval, and transparent assessment.Source: §6.2.3 · pp. 202–203
• Ethical constraints are central because biased or opaque outputs can marginalize perspectives, distort cultural meaning, and undermine trust and accountability in socially consequential domains.Proposed directions include bias audits, stakeholder participation, context-sensitive alignment, and interdisciplinary standards.Source: §6.2.3 · p. 203

6.3 Signals to Strategy: LLMs for Economics and Business

• LLMs introduce language-based reasoning into economics and business, where textual information can be transformed into structured decisions and interventions.The section characterizes this as a shift in domains traditionally led by numerical and symbolic methods.Source: §6.3.1 · p. 203, §6.3 · p. 203

6.3.1 Shared Opportunities

• 89% of financial institutions were using or piloting GenAI/LLMs, while over 70% of BFSI leaders remained in the proof-of-concept phase.The 2024 IIF-EY survey also reported that 100% increased AI/ML investments in 2024 and 80% prioritized internal process optimization over customer-facing applications.Source: §6.3.1 · p. 203
• LLMs connect unstructured text with structured data, automate expert language tasks, simulate economic behavior, and provide natural-language access to complex systems.Applications include combining financial narratives with risk models, simulating agents, and conversationally querying databases and regulatory documents.Source: §6.3.1 · p. 204

6.3.2 Common Limitations

• In economics and finance, LLM limitations include weak numerical grounding, inconsistent structured reasoning, temporal fragility, opacity, and inherited bias.These shortcomings affect precision-sensitive and high-stakes tasks such as finance, forecasting, compliance, credit scoring, and hiring.Source: §6.3.2 · p. 205

6.3.3 LLM Paradigms from Finance to Marketing

• Economic and business paradigms combine document grounding, controlled generation, agent simulation, causal adaptation, and high-stakes decision support, while emphasizing validation and oversight.Proposed safeguards include verifiable citations, human review, validation against real-world data, online learning, causal methods, confidence estimates, and fairness audits.Source: §6.3.3 · pp. 205–206

6.4 Models as Instruments: LLMs for Science and Engineering

• In science and engineering, LLMs are instruments that translate intent into operational artifacts and support workflows, but must remain embedded in auditable, constraint-aware pipelines.Their role is to complement established apparatus and physically grounded models through interfaces, knowledge translation, generative design, education, and human-in-the-loop verification.Source: §6.4 · p. 206

6.4.1 Probe into Individual Disciplines

• Chemistry benchmarks cover property prediction, reaction tasks, and molecule generation, but emphasize drug-like molecules and patent reactions while underrepresenting 3D structure and multi-step synthesis.The review also notes that molecular textualization, prediction, and synthesis planning struggle with novel structures, nuanced conformations, and underrepresented reaction conditions.Source: §6.4.1 · p. 208
• Genomics models improve enhancer and eQTL prediction by 20–40% over CNN/RNN baselines, while clinical models support risk scoring and generative models propose hypotheses for laboratory validation.The review also reports that BEHRT and GatorTron outperform traditional models with limited fine-tuning data.Source: §6.4.1 · p. 209
• Earth-science tools support remote-sensing analysis, geospatial workflows, and document summarization, while civil and computing applications assist compliance, coding, debugging, and circuit design.Their reliability remains limited by hallucinated tool calls or regulatory interpretations, weak physical reasoning, and the need for validation and human oversight.Source: §6.4.1 · pp. 209–210

6.4.2 Shared Opportunities

• Shared opportunities include domain-specific grounding, hybridization with simulators, natural-language interfaces to modeling pipelines, design exploration, and more accessible education and communication.In these roles, LLMs augment established numerical methods rather than replace their rigor.Source: §6.4.2 · p. 211

6.4.3 Common Limitations

• LLMs remain limited by weak mechanistic reasoning, costly experimental validation, hallucinations, sparse biased data, and brittle reasoning or innovation in complex tasks.These barriers constrain their use in safety-critical workflows and frontier scientific problems.Source: §6.4.3 · p. 212

6.4.4 LLM Paradigms for Science and Engineering

• The proposed adoption paradigms position LLMs as mediators across knowledge formats, constraint-grounded hypothesis generators, guides through verifiable design search, adaptive educational assistants, and safeguards against dual-use risks.A key challenge for mediation is translating between natural language and formal representations such as equations, CAD sketches, solver decks, or molecular graphs.Source: §6.4.4 · pp. 212–213
• Constraint-grounded filtering and simulation can narrow design spaces and check candidate solutions, while adaptive tutoring can scaffold learning and domain-specific safeguards can screen unsafe outputs.The review highlights the need for validation pipelines, oversight, and human guidance so that these systems complement rather than replace established practice.Source: §6.4.4 · pp. 213–214

6.5 Pilot the Present, Plot the Future

• The review synthesizes progress across arts, letters and law, economics and business, and science and engineering to assess how LLMs are changing knowledge creation, disciplinary methods, and cross-field integration.It frames the section as taking stock of current developments and charting future directions.Source: §6.5 · p. 214

6.5.1 Current State

• Across disciplines, LLMs scale information processing, support knowledge-augmented decisions, enable pre-experimental simulations, lower technical-workflow barriers, and foster human–AI co-creation.Source: §6.5.1 · pp. 214–215
• Together, these applications move LLMs from language generators toward cross-disciplinary intelligent infrastructures that democratize access to knowledge and practice.Source: §6.5.1 · p. 215
• Shared limitations include weak verifiable and causal reasoning, contextual and numerical precision, attribution, robustness, reproducibility, bias, accountability, and costly validation, yielding bottlenecks in reasoning, adaptation, and auditability.Source: §6.5.1 · p. 215, §6.5.2 · p. 215

6.5.2 Future Path

• Future systems should align multimodal schemas, orchestrate tools with physics validators, ground claims in traceable evidence, and adapt through online learning, drift detection, and causal methods.Source: §6.5.2 · p. 216
• Reproducible, rule-governed agents should log seeds and prompts and validate outcomes; high-stakes recommendations should include calibrated confidence, risk thresholds, domain limits, and audit trails.Source: §6.5.2 · p. 216
• The roadmap also calls for efficient domain-specialized models, logged human oversight and fairness audits, discipline-aligned education, and red-teaming, sandboxing, and incident response for safety.Source: §6.5.2 · pp. 215–216
• The review presents selected domains as a first cross-disciplinary step, aiming for dependable, auditable infrastructures that connect evidence to reasoning and support disciplinary-scale analysis, simulation, and decisions.It is not exhaustive, and its claims may be contested as AI evolves.Source: §6.5.2 · p. 216