Large Language Model (LLM)
An AI model trained on vast text corpora to predict and generate language — the technology behind ChatGPT, Claude, and Gemini, increasingly used in research for writing, summarization, and literature synthesis.
What it means
A large language model (LLM) is an AI model trained on large amounts of text to predict the next token (word fragment) given the preceding context. Through this deceptively simple training objective on very large datasets, LLMs develop broad capabilities: they learn grammar, factual knowledge, reasoning patterns, and stylistic conventions present in their training data.
“Large” refers both to training data (typically hundreds of billions of words or more) and model size (billions of parameters). Scale is what distinguishes modern LLMs from earlier, smaller language models — capabilities emerged at scale that smaller models didn’t exhibit.
How researchers use LLMs
LLMs have become practical tools across several research tasks:
- Writing and revision: drafting, restructuring, improving clarity of scientific text — particularly useful for non-native English speakers
- Literature synthesis: summarizing a set of papers, identifying themes, comparing methodological approaches
- Code generation: writing analysis scripts, data cleaning pipelines, plotting code — tools like Claude and GitHub Copilot are widely used for this
- Structured data extraction: pulling specific information (sample sizes, effect sizes, study characteristics) from paper text
- Brainstorming: generating hypotheses, identifying analogies across fields, stress-testing experimental designs
What LLMs don’t do
- They don’t retrieve information — a standard LLM responds from training data only; it has no access to the web or new papers unless augmented with retrieval (see Retrieval-Augmented Generation)
- They don’t reason reliably from first principles — they pattern-match on training data; complex multi-step logical or mathematical reasoning is less reliable
- They hallucinate — see Hallucination: LLMs generate plausible-sounding text that may be factually wrong, including invented citations
Key models used in research contexts
| Model | Developer | Notes |
|---|---|---|
| GPT-4o / GPT-4 | OpenAI | Strong general performance; widely integrated into tools |
| Claude (3.5/3.7 Sonnet, Opus) | Anthropic | Strong for long documents and scientific writing; larger context window |
| Gemini | Google DeepMind | Strong multimodal capabilities; integrates with Google Workspace |
| Llama 3 | Meta | Open-source; can be run locally for sensitive data |