post-training

advancedwidespread

Post-training is the second phase of language model training, in which a raw, text-completing model becomes a helpful, instruction-following, and safely behaving assistant.

In plain terms

Imagine someone who has read half the world's libraries but has never worked with people. They know a lot, yet instead of answering your question, they just continue with whatever comes to mind. This is what a language model is like after pre-training: it understands how to continue text, but not how to converse.

Post-training is like onboarding at a first job. Here the model learns to answer questions, follow instructions, what tone to use, when to say no, and how to use tools, such as a search engine or program code.

Most of a model's knowledge comes from pre-training, while its behavior, style, and 'personality' are largely shaped by post-training. This is why two models with similar knowledge can be very different conversational partners.

How it works

Starting point: there is a pre-trained base model that has learned to predict the next word (more precisely, a text segment, i.e., a token) on a huge amount of text, but does not yet behave as an assistant.
Learning from examples: developers show it carefully compiled question–answer pairs and sample conversations so it learns instruction following and the form of dialogue.
Learning from feedback: humans, other models, or automated checks evaluate its answers, and the model tunes itself toward the answers deemed better (reinforcement learning).
Safety and principles: the model learns what not to do, how to handle sensitive requests, and what values to weigh.
Special capabilities and oversight: the system practices tool use, multi-step task solving, and extended reasoning, then undergoes test series before release to measure what it can do and where it fails.

An everyday example

You ask the chatbot how to write a resignation letter, and it politely answers in bullet points, then asks at the end if you need a sample. It learned this helpful behavior during post-training. If you entered the same into a raw base model, it might simply continue your sentence.

Who is affected

everyonePost-training determines how helpful your AI assistant is, what tone it uses, and what it refuses to do.

developerThis determines how reliably the model follows instructions and how usable it is as an agent, for example, as a coding agent. → Code

entrepreneur, executiveMuch of the difference between models is decided here, so when choosing a model, it is worth testing its behavior, not just its knowledge.

lawyer, citizenDuring post-training, the developer company's value choices are embedded into the model, raising questions about transparency and accountability. → Law

A common misconception

MythDuring post-training, the model receives fresh knowledge, enabling it to know about the latest events.

In realityPost-training mainly shapes behavior rather than updating knowledge. Fresh information typically comes from search or from documents loaded into the conversation.

Limits and risks

Post-training can reduce but not eliminate errors: the model can still confidently state falsehoods.
Learned constraints are not unbreakable; with clever phrasing, they can sometimes be bypassed.
Over-tuning can have a price: the model may become sycophantic or overly cautious and reject requests unjustifiably.
From the outside, you rarely see exactly what data and criteria were used for post-training, as companies only partially disclose this.

Similar, but not the same

pre-training – Pre-training provides general linguistic and world knowledge from massive text volumes, while post-training shapes this into a usable, rule-following assistant.
fine-tuning – Fine-tuning is a technique in which an already trained model is trained further on targeted data; the model's users (for example, companies) can also do this for a narrower task. Post-training, by contrast, is the model developer's entire second stage, of which fine-tuning is only one tool.
post-training quantization – Its English name (post-training quantization) is similar, but this compresses the finished model to run on smaller machines, not behavior training.

Related concepts

Broader

machine learning, model training, large language model

Narrower

fine-tuning, reinforcement learning from human feedback, instruction tuning, safety training

On the edge

pre-training, foundation model, AI alignment, Distillation, evaluation

A brief history

The combination of pre-training and subsequent fine-tuning became common with the GPT and BERT models released in 2018. OpenAI's 2022 InstructGPT study showed that with human feedback, a raw model can be transformed into an instruction-following assistant. The term 'post-training' then spread as the collective name for the entire second phase.

Where things stand

At the end of 2026, every major chat model undergoes post-training, and the large AI labs maintain dedicated teams for it. The emphasis is increasingly on teaching multi-step work done as an agent, and tool use.

Status: 2026 November

In the news

Appears in 1 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Gemini 4 could arrive "much earlier" than planned, says new Google DeepMind chiefSep 23, 2026 · Agent Age

Also known as

post-training · post-training · post-training · post-training phase · post-training · post training · posttraining

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.