Large Language Model
Subject: Computer and Technology noun /ˌlɑːrdʒ ˈlæŋ.ɡwɪdʒ ˈmɑː.dəl/ 1 definition
- artificial-intelligence
- machine-learning
- deep-learning
- natural-language-processing
-
Definition (noun)
Entry #215 · UpdatedA Large Language Model (LLM) is a high-parameter neural network—most commonly built on the transformer architecture—that uses self-supervised pretraining over massive text corpora to model conditional probabilities across sequences of tokens. By predicting subsequent or masked tokens over wide context windows, it generalizes across diverse linguistic tasks such as text generation, summarization, translation, and code synthesis from natural-language prompts. Unlike rule-based parsers or narrow single-task classifiers, an LLM operates as a general-purpose probabilistic system rather than a deterministic database, meaning its outputs reflect statistical patterns learned during training and can exhibit hallucinations unless grounded by external retrieval or verification.
Common Types
Base Models: Models trained solely on self-supervised next-token prediction over raw corpora; they serve primarily as checkpoints for downstream fine-tuning rather than direct conversational use.
Instruction-Tuned Models: Base models refined through supervised fine-tuning (SFT) and preference alignment methods such as reinforcement learning from human feedback (RLHF) to follow user prompts reliably and safely.
Reasoning Models: Models trained to allocate extra inference-time compute to step-by-step deliberation before producing a final answer, improving accuracy on multi-step mathematics, logic, and coding tasks.
Multimodal Language Models: Extended architectures that align image, audio, or video encoders with a language backbone so a single model can interpret and reason across multiple media types alongside text.
Example
An engineering team integrates a 70-billion-parameter instruction-tuned LLM into its continuous integration pipeline to summarize pull request diffs and flag potential edge cases before human review. Using a 32,000-token context window, the model reads modified source files alongside repository style guidelines and emits structured review comments formatted as JSON.