Working with LLMs responsibly in research

Working with LLMs responsibly in Research - an introduction to non-computer scientists - Thilo Schmid, Leon Rohmann, and Jana Schor supported by Hannes Bohring and Riccardo Massei Additionally supported by: Julia Porath, Thomas Schnicke, Jan Bumberger Part 1 - Introduction to LLMs in Research - Agenda 1. Welcome, goals 2. From Chatbots to Agents 3. Agentic AI 4. What you can realistically expect from LLMs today 5. How to talk to LLMs: simple prompting patterns 6. Hands-on: “One task of my work, but with an LLM” 7. Responsible use & key take-aways 2 Goals Show how LLMs can realistically support your work. Clarify limits & avoid typical mistakes. Give practical prompt patterns & ideas for daily workflows. 3 It is not about making you AI engineers → It’s about using tools that already exist to save time, while staying scientifically rigorous. What are LLMs (and what are they not)? ” A Large Language Model (LLM) is a system that predicts the next word based on patterns in huge amounts of text. Input text Output text LLM Good at Language: ● Drafting ● Rewriting ● Summarizing ● Explaining Not designed as: ● Database of facts ● A statistician ● A domain expert → Can produce fluent, convincing text in many styles → Can also produce confident nonsense (“hallucinations”) → Quality depends strongly on how you ask and what data it has seen 4 Simple example Explain quantum dots Field Use Case Displays Quantum dot TVs (QLEDs) offer better color range and brightness. Biomedical Imaging Used as fluorescent probes that are brighter and longer-lasting than dyes. Solar Cells Enhance efficiency by tuning absorption across the spectrum. Quantum Computing Serve as qubits due to discrete energy levels and charge/spin properties. Quantum dots are tiny semiconductor particles, typically only a few nanometers in size (1–10 nanometers), that have unique optical and electronic properties due to quantum mechanics. Key Points: 🧠 What makes them special? ● Because they're so small, electrons inside them are confined in all three dimensions. ● This confinement leads to quantized energy levels, similar to what happens in atoms—hence they're sometimes called "artificial atoms." ● The result: their color (wavelength of emitted light) depends on their size. 🌈 Tunable Color: ●Smaller quantum dots emit blue light (shorter wavelength). ●Larger quantum dots emit red light (longer wavelength). ● This size-dependent light emission is what makes them great for displays and imaging. 🔬 What are they made of? ● Common materials: CdSe (cadmium selenide), PbS (lead sulfide), InP (indium phosphide), and others. ● They usually have a core-shell structure to improve stability and brightness (e.g., CdSe core with ZnS shell). Applications: Simple Analogy: Think of quantum dots like tunable light bulbs at the nanoscale. By making them slightly bigger or smaller, you can make them glow in different colors. Would you like a visual diagram or want me to explain one of the applications in more depth (e.g., in biology, electronics, or energy)? 5 🌀 Simple example You are a science communicator. Explain quantum dots in 3–4 sentences for a newspaper article, without equations. Quantum dots are incredibly small particles—just a few billionths of a meter wide—that can glow in bright, vivid colors. What makes them special is that their color changes with size: smaller dots shine blue, while larger ones shine red. This happens because their tiny size changes how they interact with light, thanks to the rules of quantum mechanics. They're already used in TV screens, medical imaging, and solar panels, offering brighter, more efficient, and more precise technologies. Same model, same knowledge - different output quality for the intended audience, just because we frame the task differently. This is the core of ‘prompting’. 6 🌀 From Chatbots to Agents Agentic AI builds on the same LLMs – but changes how we organize work. LLM as Chatbot 👤 Ask a question or give a short task 🌀 Returns a single answer ● Rewrite a paragraph ● Suggest bullet points for a section ● Explain a concept at a different level You probe and verify answers manually Agentic AI 👤 Define a goal 🌀 Plans multiple steps, may call tools, iterates until done ● Scan new documents, extract key points, update a shared report ● Pull data from a source, run predefined analysis code, draft a figure caption ● Monitor a mailbox, draft responses, and compile a weekly status summary You design the workflow, review intermediate outputs, approve or reject actions Interaction pattern Goal Workflow LLM Prompt Response LLM Typical capabilities Human role 7 What is an AI Agent? Agents use LLM’s to iterate on tasks and extend the model’s abilities with Tools. The role of the LLM is often specified using a system prompt. 8 + LLM Tools ● Accuracy ● Consistency ● Reduced Hallucinations ● More specific to the task ● Access to external knowledge outside of training set Tools are accessed through text. ● Search a database ● Create/Read a file ● Analyze an image “You are an expert in environmental research, toxicology, and chemistry…” Let’s build a calendar manager 9 Scan my emails for meeting requests, check for conflicts, suggest alternatives when needed, and add all conflict-free events to the calendar. Agent Loop LLM Response … TOOL OUTPUT: Email sent to John Doe Let’s look under the agents hood 16 Prompt Agent Loop Add Date LLM Response Emails Check Calendar LLM: Hi Thilo, I checked your emails and you have 3 new appointment requests. 2 of the requests don’t conflict with your calendar, these are: 05.12.2025 at 13:30 and 25.01.2026 at 9:00. I added them to the Calendar. For the third meeting I drafted an email to suggest a reschedule on the 30.1.2026 and sent it to John Doe. Can I help you with anything else? Let’s look under the agents hood 17 .ics is a common calendar file format Prompt … TOOL OUTPUT: Email sent to John Doe LLM: Hi Thilo, I checked your emails 📧 and you have 3 new appointment requests 💯. 2 of the requests don’t conflict with your calendar, these are: 05.12.2025 at 13:30 and 25.01.2026 at 9:00 😃. 📅 I added them to the Calendar. 🪶 For the third meeting I drafted an email to suggest a reschedule on the 30.1.2026 and sent it to John Doe. 🚀 Can I help you with anything else? Agent Loop Add Date LLM Response Emails Check Calendar What can you realistically expect from LLMs today? 18 Think of 1–2 things you would like LLM agents to do for you in your work (papers, reports, emails, data, project management). Write a first draft of a related-work section. Give me all important papers on X with correct references. Analyze my dataset and choose the ‘best’ model. Draft a project status report from meeting notes and emails. Decide which hypothesis we should test next. Place them under “realistic” or “unrealistic”. What can you realistically expect from LLMs today? 19 Think of 1–2 things you would like LLM agents to do for you in your work (papers, reports, emails, data, project management). Place them under “realistic” or “unrealistic”. Write a first draft of a related-work section. Give me all important papers on X with correct references. Analyze my dataset and choose the ‘best’ model. Draft a project status report from meeting notes and emails. Decide which hypothesis we should test next. What can you realistically expect from LLMs? 20 High-quality language support ● Rewrite, shorten, or clarify paragraphs for different audiences (expert, interdisciplinary) ● Improve grammar, flow, and structure while keeping your technical content Template and boilerplate drafting ● First drafts of emails, project updates, lay summaries, cover letters, slide headlines ● Standard phrases (limitations paragraphs, reproducibility statements, reviewer responses), which you then adapt With agentic GPTs, assuming good prompts and human verification Idea generation & structuring ● Propose outlines for sections (e.g., Introduction, Methods, Impact) ● Brainstorm questions, hypotheses, or angles to consider (which you then filter critically) Summarization & reformatting ● Summarize abstracts, papers, reports, or meeting notes into key points ● Convert between formats: bullets → prose, notes → email, report → executive summary Assistance with known methods / code ● Suggest example code snippets in R/Python for standard analyses, to be reviewed and tested ● Explain methods or statistical concepts at different levels of detail Semi-automated workflows for repetitive tasks ● Orchestrate multi-step workflows like: ○ Collect text from documents or emails → extract key fields → populate a table or draft report ○ Monitor a source (e.g. list of abstracts) → summarize → send you a digest for review ● Always with you, reviewing important steps and final outputs What can you not expect from LLMs? 21 Perfect factual accuracy or “truth” ● can still “hallucinate” facts, numbers, and references. ● not authoritative databases; all important claims must be checked against primary sources (papers, databases, legal texts) Full automation of complex analyses without inspection ● “Upload my data and tell me the best model and conclusions” ● Agentic systems that run code or access tools can still mis-specify models, misinterpret outputs, or overfit – they require the same kind of review you would apply to a junior collaborator’s work Reliable literature search & citation without verification ● can suggest plausible-sounding but non-existent papers or wrong DOIs ● cannot replace proper searches in bibliographic databases, only complement them for brainstorming Autonomous scientific judgment ● Choosing the “right” study design, statistical test, or interpretation of results without your domain expertise ● Deciding which hypothesis to accept or reject, or what is ethically acceptable in a study Unrealistic / risky expectations → be extremely cautious! Fully autonomous “research agents” without human-in-the-loop ● Letting agents read mail, change documents, or make decisions without clear approval steps and logging is operationally and ethically risky ● Current best practice is human-in-the-loop: agents can automate steps, but humans approve critical actions What can you not expect from LLMs? .. ctd 22 Confidential or sensitive data handling in public systems ● Uploading protected personal data, unpublished results, or confidential project documents into public ChatGPT/Gemini is usually not acceptable under data protection or organization’s rules ● Even advanced models do not change organizational and legal constraints Replacing authorship, accountability, or originality ● LLMs cannot be authors of scientific papers; journals and funders require human accountability ● You cannot outsource responsibility for correctness, plagiarism, or ethical compliance to an AI system – you remain responsible Unrealistic / risky expectations → be extremely cautious! Fully autonomous “research agents” without human-in-the-loop ● Letting agents read mail, change documents, or make decisions without clear approval steps and logging is operationally and ethically risky ● Current best practice is human-in-the-loop: agents can automate steps, but humans approve critical actions How to talk to LLMs: What is a “prompt” and why it matters 23 Improve my text. A prompt = everything you type or provide (text, files, links, images) that guides the model. Good prompts answer: Who should you act as, what to do, for whom, and with what material. The model is “smart but literal” It does not know your hidden intentions or institutional context unless you tell it. 🌀 How to talk to LLMs: What is a “prompt” and why it matters 24 You are a scientific editor. Improve the clarity and flow of the paragraph below for an interdisciplinary audience. Keep technical terms, shorten by about 25%. Here is the paragraph: … A prompt = everything you type or provide (text, files, links, images) that guides the model. Good prompts answer: Who should you act as, what to do, for whom, and with what material. 🌀 The model is “smart but literal” It does not know your hidden intentions or institutional context unless you tell it. How to talk to LLMs: Role + Task + Context + Constraints 25 👤 Role how the model should “behave” e.g., “You are a scientific editor / project manager / reviewer.” 📋 Task what you want as output e.g., “Rewrite…,” “Summarize…,” “Draft an email…,” “List potential risks…” 📄 Context background and input material e.g., what the document is, who the audience is, why you need it 🎚 Constraints / Format style, length, structure e.g., “max 150 words,” “bullet points,” “neutral tone,” “include 3 strengths and 3 weaknesses” You are a [ROLE]. Your task is to [TASK]. Context: [who is the audience / what is this for]. Use the following material: [paste text / describe file]. Please answer in [FORMAT / LENGTH / STYLE]. A template might be: 🌀Use line breaks and blank lines. Use tags and “:”. Key take-aways 32 LLMs are powerful language tools, not automatic truth machines. Good prompts = clear role, task, context, constraints – and iterate. Use them to draft, structure, and summarize - but always verify and respect data/ethics rules. Homework: ● Pick one of your tasks from this week. ● Try doing it with ChatGPT/Gemini using today’s patterns. ● Evaluate the results and decide if it saves you time. There is no free lunch. Everything comes with a price. → you decide what you gain with how much effort AI or LLM? Beware of the differences and the scope 33 https://www.linkedin.com/posts/eva-sula-02b9a73_ai-in-defence-mini-series-part-1-llms-activity-7397544072925810688-TnF_?utm_source=share&utm_mediu m=member_desktop&rcm=ACoAABXDVAoBkaqPkK78HJYjAUVMzCkg1JGvC9w