Large language model (LLM)

A large language model (LLM) is a neural network trained on very large amounts of text to predict the next token, a word or piece of a word, from the text before it. The developers of GPT-3 described it as an autoregressive language model with 175 billion parameters: each new token is predicted from the ones already written 2.

Why do LLMs matter in healthcare?

Clinical work runs on language: notes, messages, referral letters, guidelines. LLMs "can respond to free-text queries without being specifically trained in the task in question" 4, so one model can be pointed at many of those jobs. The early signal was striking. ChatGPT performed at or near the passing threshold on all three USMLE steps without specialized training 5. Versions adapted for healthcare are covered under clinical LLM, and benchmark results sit in our LLM benchmark results tracker.

How does an LLM work?

Modern LLMs are built on the transformer, an architecture "based solely on attention mechanisms, dispensing with recurrence and convolutions entirely" 1. They are made in two broad stages.

Pretraining. The model reads broad text and learns to predict the next token. At the scale of GPT-3, that alone was enough to perform new tasks from a few examples placed in the prompt 2. Broad pretraining of this kind is what places LLMs among foundation models.

Fine-tuning for instructions. "Making language models bigger does not inherently make them better at following a user's intent" 3. InstructGPT was built by fine-tuning GPT-3 on human-written demonstrations and then with reinforcement learning from human rankings of its outputs 3. ChatGPT, similarly, was produced through fine-tuning of an underlying LLM 4.

In use, your prompt and any attached text go into the model's context window, and the answer comes out one token at a time.

Common misunderstandings

Fluent means correct. The InstructGPT authors wrote that large language models "can generate outputs that are untruthful, toxic, or simply not helpful to the user" 3. In healthcare the confident fabrication is the failure to plan for; see AI hallucination in clinical contexts.

An LLM looks facts up. A plain LLM generates from patterns learned in training. Checking against current sources requires a system built around it, such as retrieval-augmented generation.

Exam scores predict bedside performance. In a simulated clinical setting built from 2,400 real patient cases, the open LLMs tested diagnosed less accurately than physicians, followed neither diagnostic nor treatment guidelines, and could not reliably interpret laboratory results 6.

What to ask a vendor

• Which underlying model and version does the product use, and can it change without notice?
• Are answers grounded in sources we can inspect?
• What is the context window, and what happens with long records?
• How was it evaluated on our task, beyond exam benchmarks? See LLM evals vs clinical evaluation.
• Is our data used to train or fine-tune the model? See HIPAA and LLMs.

Related terms

• Clinical LLM: LLMs adapted and evaluated for healthcare.
• Foundation model: the wider class LLMs belong to.
• Context window: how much an LLM can take in at once.