Natural language processing (NLP) is the field of computational methods for language. Large language models (LLMs) are one family of models used within that field. NLP also includes rules, dictionaries, statistical models, retrieval, classifiers, speech-linked systems, and smaller specialized neural models.

What an LLM learns

An autoregressive language model estimates probabilities for tokens conditioned on context. Pretraining on large corpora can yield representations and generation abilities useful across tasks, but fluent output does not prove understanding, factuality, intent, or access to current information.

Tokenization affects inputs, sequence length, languages, and cost; see tokenization in NLP.

How LLMs support NLP tasks

Prompting, retrieval, tool use, or adaptation can support classification, extraction, translation, summarization, question answering, and generation. The best architecture depends on data, latency, cost, privacy, explainability, and error tolerance. A smaller model, rules, or search may outperform an LLM for a bounded task.

Review NLP applications and compare every approach on the same representative task set.

Adaptation is not automatic knowledge

Fine-tuning changes model behavior using examples; it does not guarantee current facts, safety, or domain expertise. Retrieval can provide updateable and citable information but may fail through poor retrieval, permissions, stale sources, or prompt injection. See fine-tuning guidance.

Evaluate the complete system

  • Define users, languages, task, unsupported cases, and unacceptable failures.
  • Measure correctness, source support, completeness, calibration or abstention, robustness, safety, privacy, security, latency, cost, and subgroup performance.
  • Validate prompts, retrieval, tools, permissions, interface, human review, fallback, and monitoring.
  • Protect training, evaluation, prompt, and log data; tokenization does not anonymize it.

Use LLM evaluation metrics as a design guide, not a checklist that proves quality.

Reviewed and substantially updated September 4, 2026. Original publication date preserved.