Large language models (LLMs) are neural networks trained on vast text corpora to learn statistical patterns of language and then generate or analyze text. They rely on transformer architectures with self attention, which allow the model to weigh relationships among all tokens in a sequence and capture long range dependencies. Training uses next token prediction or masked language modeling on internet scale datasets, producing systems that can generalize across many tasks without task specific supervision.

A key property is in context learning. Without changing their weights, LLMs can adapt to new tasks given a few examples in the prompt, effectively performing meta learning. Scaling laws show that performance improves predictably with model size, data volume and compute, motivating ever larger models. At sufficient scale, models display emergent abilities such as chain of thought reasoning, code synthesis and multi step question answering that are weak or absent in smaller systems.

Despite strong capabilities, LLMs hallucinate, producing plausible but incorrect statements due to their reliance on learned statistical patterns rather than grounded world models. Safety research targets harmful content, biases, privacy risks and prompt injection attacks, using alignment techniques such as instruction tuning and reinforcement learning from human feedback. Evaluation is challenging because benchmarks can become saturated or contaminated by training data, so newer work emphasizes robustness, fairness and real world usefulness.

Recent efforts explore retrieval augmentation, tool use and modular architectures to integrate external knowledge, improve factuality and enable planning. Ongoing research also examines data efficiency, multilingual performance and the environmental and social impacts of large scale training.