Transforming Science with Large Language Models: A Survey on AI-assisted Scientific Discovery, Experimentation, Content Generation, and Evaluation

This survey provides a comprehensive overview of the emerging ecosystem of large language models and tools that support researchers across the scientific lifecycle, covering key tasks from literature search and idea generation to content creation, experimentation, and evaluation, while addressing associated datasets, methods, limitations, and ethical concerns.

The Gist

Imagine the world of science as a massive, bustling library that is growing so fast it's about to burst its shelves. For centuries, scientists have been the librarians and explorers of this library, doing all the heavy lifting: finding books, reading them, figuring out what's missing, building experiments, and writing reports.

Now, a new kind of super-intelligent assistant has arrived: Large Language Models (LLMs). Think of these as a team of incredibly fast, well-read, but sometimes overly confident robots. This paper is a giant "user manual" and "safety guide" for how we can use these robots to help scientists, while being very careful not to let them run the show.

Here is a breakdown of how these AI assistants are changing the scientific game, using simple analogies:

1. The Super-Librarian (Literature Search)

The Old Way: A scientist spends weeks digging through piles of papers, trying to find the one sentence that answers their question. It's like looking for a needle in a haystack made of other needles.
The AI Way: The AI is a Super-Librarian. You can ask it, "What do we know about curing this specific type of cancer?" and it instantly scans millions of books, summarizes the key points, and hands you a neat report.

• The Catch: Sometimes the librarian gets a bit confused and invents a book that doesn't exist (called a "hallucination"). So, you still have to double-check the shelf.

2. The Creative Brainstormer (Ideas & Hypotheses)

The Old Way: A scientist sits in a quiet room, staring at a wall, trying to come up with a new idea. It's like trying to invent a new flavor of ice cream from scratch.
The AI Way: The AI is a Creative Brainstormer. It reads everything ever written about ice cream and suggests, "What if we mix mint with spicy chili?" It can generate hundreds of wild ideas in seconds.

• The Catch: While it's great at being creative, it's not always practical. It might suggest an idea that sounds cool but is impossible to build or test. It needs a human to say, "Okay, that's a fun idea, but let's try this one instead."

3. The Speed-Writer (Writing Papers)

The Old Way: Writing a scientific paper is like building a house brick by brick. You have to write the title, the abstract, the introduction, and the conclusion, making sure every fact is perfect.
The AI Way: The AI is a Speed-Writer. It can draft the title, summarize the abstract, and even write the "Related Work" section (which is like listing who built houses before you).

• The Catch: The AI is great at grammar and style, but it's terrible at facts. It might write a beautiful paragraph about a bridge that collapses because it forgot to mention the steel beams. Also, it can't be the "author" of the house; the human scientist must take responsibility for the blueprint.

4. The Artist (Figures, Tables, and Slides)

The Old Way: Turning data into a chart or a slide deck is like painting a picture by hand. It takes time and skill to make sure the colors match the data.
The AI Way: The AI is a Digital Artist. You give it a description like "Show me a graph of rising temperatures," and it instantly paints the picture, draws the chart, or even designs a presentation slide.

• The Catch: Sometimes the artist gets the colors wrong. The graph might show a line going up when the data says it's going down. You need a human to look at the painting and say, "Wait, that's not right."

5. The Critic (Peer Review)

The Old Way: Before a paper is published, other scientists (peers) read it to check for mistakes. It's like a food critic tasting a dish before it goes to the customers.
The AI Way: The AI is a Junior Critic. It can read the paper quickly, check for grammar, and even spot if a claim sounds suspicious. It helps the human critics by doing the boring first pass.

• The Catch: The AI doesn't have "common sense" or deep intuition. It might miss a subtle lie or a brilliant new idea that breaks the rules. We can't let the robot be the only judge; a human must have the final say.

The Big Warning: The "Fake News" Problem

The paper sounds a lot of alarms about Ethics.

• The "Hallucination" Risk: The AI is like a storyteller who loves to make things up to make the story sound better. In science, making things up is dangerous.
• The "Plagiarism" Risk: If the AI writes the paper, who owns it? The paper says the human must always be the boss. The AI is just the tool, like a hammer. You can't say the hammer built the house.
• The "Bias" Risk: If the AI was trained on books written mostly by men from one country, it might think that's the only way to do science. We need to make sure the AI sees the whole library, not just one corner.

The Bottom Line

This paper is essentially saying: "AI is a fantastic co-pilot, but it cannot be the pilot."

Science is moving from a time where humans did everything alone, to a time where humans and robots work together. The robot handles the heavy lifting, the speed, and the data crunching. The human provides the judgment, the ethics, and the "spark" of true discovery. If we use these tools wisely, we can solve big problems faster. If we let the robot drive without a human in the seat, we might crash into a wall of fake science.

Technical Summary

Based on the survey paper "Transforming Science with Large Language Models: A Survey on AI-assisted Scientific Discovery, Experimentation, Content Generation, and Evaluation" by Eger et al. (2026), here is a detailed technical summary.

1. Problem Statement

The scientific research lifecycle is facing a crisis of scale due to the exponential growth of scientific literature, making it increasingly difficult for researchers to efficiently search, synthesize, and generate new knowledge. While Large Language Models (LLMs) and Large Multimodal Models (LMMs) have shown promise in automating various tasks, the field lacks a comprehensive, workflow-centric overview. Existing surveys are often too narrow, focusing on specific domains (e.g., only physics or social sciences) or specific tasks (e.g., only literature search). Furthermore, there is a critical gap in understanding the limitations, ethical risks (such as hallucination and integrity erosion), and the practical feasibility of integrating AI into the full scientific cycle, from ideation to peer review.

2. Methodology

The authors adopt a narrative survey methodology rather than a systematic review. This approach was chosen to synthesize heterogeneous and rapidly evolving bodies of work across different domains and methodological traditions.

• Selection Criteria: Papers and tools were selected by a diverse team of co-authors with domain expertise. Selection was based on technical depth, influence, publication venue reputation, and citation patterns.
• Structure: The survey is organized around the scientific research lifecycle, categorizing AI support into five core tasks:
  1. Literature Search, Summarization, and Comparison.
  2. Scientific Discovery (Ideation, Hypothesis Generation, and Experimentation).
  3. Text-based Content Generation.
  4. Multimodal Content Generation and Understanding.
  5. AI-assisted Peer Review.
• Analysis Framework: For each task, the authors analyze datasets, methods (e.g., RAG, fine-tuning, multi-agent systems), evaluation strategies, limitations, and future directions.

3. Key Contributions and Technical Findings

A. Literature Search and Summarization

• Tools: The survey categorizes tools into AI-enhanced search (e.g., Elicit, Consensus), graph-based systems (e.g., ORKG ASK, Connected Papers), paper chat/QA (e.g., NotebookLM, ChatPDF), and recommender systems.
• Techniques: Reliance on Retrieval-Augmented Generation (RAG) to ground answers in specific documents, reducing hallucinations. Graph-based systems utilize Knowledge Graphs (KGs) to structure unstructured text for complex querying.
• Limitations: Issues with data quality, coverage gaps in non-English or niche domains, and algorithmic bias in ranking.

B. Scientific Discovery (Ideation & Experimentation)

• Ideation & Hypothesis Generation: LLMs are used to generate novel research ideas and testable hypotheses. Methods include iterative refinement, human-in-the-loop alignment, multi-agent systems (e.g., VirSci, ResearchAgent), and Knowledge Graphs (e.g., SciAgents).
• Automated Experimentation: Systems like The AI Scientist and ScienceAgentBench aim to automate the full pipeline: task formulation, code generation (Python), execution, and evaluation.
• Evaluation: Approaches range from intrinsic (automatic metrics like ROUGE, LLM-as-a-judge) to downstream (simulation-based evaluation like LabBench and real-world experimentation in biofoundries).
• Findings: LLMs often generate ideas that are judged as more novel but less feasible than human ideas. They struggle with precise reasoning in complex domains (biology/chemistry) and lack access to real-time scientific data, leading to outdated or hallucinated hypotheses.

C. Text-based Content Generation

• Tasks: Covers title/abstract generation, related work synthesis, citation generation, and long-form manuscript drafting.
• Techniques: Shift from extractive methods to abstractive generation using transformers and LLMs. Advanced methods like LongWriter and LongWriter-Zero utilize reinforcement learning (RL) and hierarchical attention to manage ultra-long contexts (10,000+ words).
• Critical Issues: Citation hallucination is a major problem; studies show significant rates of fabricated references in LLM outputs. While proofreading and paraphrasing are strong suits, factual consistency remains a weak point requiring human oversight.

D. Multimodal Content Generation

• Scope: Generation and understanding of figures, tables, slides, and posters.
• Techniques:
  • Figures: Tools like AutomaTikZ and DeTikZify convert text/sketches into TikZ code. Recent approaches use Reinforcement Learning and Tree Search for better structural accuracy.
  • Tables: Focus on table-to-text (summarization) and text-to-table (extraction) using structure-aware pretraining.
  • Slides/Posters: Early extractive methods are evolving into generative models (e.g., DOC2PPT, SciDuet) that create persona-aware presentations.
• Limitations: Datasets are small compared to general image generation; models struggle with complex reasoning and often fail to generate accurate scientific visualizations without human correction.

E. AI-assisted Peer Review

• Applications: Automated review generation, meta-review synthesis, claim verification, and sentiment/politeness analysis.
• Status: LLMs (e.g., GPT-4) show promise in providing feedback and summarizing reviews but are unreliable as standalone evaluators. They struggle with assessing scientific rigor and novelty.
• Ethical Risks: High potential for bias (e.g., affiliation bias) and the risk of diminishing human critical engagement.

4. Results and Evaluation

• Performance Gaps: While proprietary models (GPT-4o, Claude) generally outperform open-source models in scientific tasks, all models still lag significantly behind human experts in complex reasoning, factual accuracy, and novelty assessment.
• Evaluation Challenges: Traditional metrics (BLEU, ROUGE) are insufficient for scientific text. There is a lack of standardized, domain-specific benchmarks. Evaluation often relies on small-scale human studies or simulation environments which may not reflect real-world scientific validity.
• Feasibility: Current AI systems are best viewed as complementary tools (copilots) rather than autonomous replacements. They require "human-in-the-loop" verification to ensure integrity.

5. Significance and Ethical Considerations

• Ethical Landscape: The survey highlights critical ethical concerns:
  • Integrity: Risks of "fake science," plagiarism, and authorship ambiguity.
  • Bias: Reinforcement of existing research paradigms and demographic biases.
  • Safety: Potential for generating harmful scientific ideas (e.g., toxic molecular designs) if not properly constrained.
• Policy Implications: The authors advocate for transparent disclosure of AI use, strict adherence to authorship guidelines (AI cannot be an author), and the development of robust safeguards.
• Future Outlook: The paper calls for the development of domain-specific benchmarks, trustworthy AI frameworks, and interdisciplinary collaboration to integrate AI into "AI4Science" systems responsibly.

Conclusion

This survey serves as a foundational resource for the "AI4Science" community, mapping the current state of AI across the scientific lifecycle. It concludes that while LLMs offer transformative potential to accelerate discovery and democratize research, their current limitations in reasoning, factuality, and safety necessitate a cautious, human-centric approach where AI acts as a powerful assistant rather than an autonomous agent.