Which should I change, this or temperature?

Whichever your interface exposes, and only one of them. They act on the same choice by different means, so changing both leaves you unable to say which produced the difference. If both are available, most teams settle on temperature simply because more documentation and more shared examples are written against it.

Will a lower setting stop it inventing things?

No, and this is the most common reason people find it. Invention comes from the system not having the information rather than from adventurous word choice, so a narrow setting produces the same fabrication in plainer language. Grounding it in real material and checking the output are what address that.

Why can I not find it in my assistant?

Because consumer products deliberately do not expose it. Generation settings appear on developer interfaces, routing services and local runners, where the audience is expected to understand the trade. Its absence from a chat window is a design decision about who the product is for rather than a missing feature.

Glossary

Top-p

Top-p limits how wide a pool of candidate words a system may pick from at each step, narrowing output towards the likeliest choices or opening it up to less obvious ones.

In plain terms

A dial for how adventurous the wording is. Turned down, it keeps to the safest, most expected words and sounds flatter but steadier. Turned up, it reaches further into the unusual ones and gets livelier and less predictable. It is a knob about phrasing, and people reach for it hoping it is a knob about correctness.

Why it matters

Because it is one of the two settings people reach for when output feels wrong, and it addresses none of the reasons output is usually wrong. Turning it down makes phrasing steadier and does nothing for a missing condition, an invented fact or a misread document, so time spent on it is frequently time not spent on the actual cause.

How it works

It caps the candidate pool by cumulative likelihood rather than by count, which is what distinguishes it from simpler cut-offs. At each step the possible next words are ranked and only the most likely ones, taken together up to a share of the total probability, remain eligible. A low setting can leave very few candidates on an obvious continuation and many on an open-ended one.

It overlaps heavily with temperature, and that overlap is the source of most confusion. Both widen or narrow the same choice by different means, so adjusting them together produces compound effects nobody predicted and makes the result impossible to attribute. The standing advice from vendors is to move one and leave the other alone.

Its reach is phrasing rather than substance. Narrowing it produces steadier, flatter wording and does not make a system more accurate, better at following instructions or less likely to invent something, because those failures happen for reasons this setting does not touch. A confident wrong answer stated in safe words is still a confident wrong answer.

Most people never see it, which is worth knowing before hunting for it. Consumer assistants do not expose it, and it surfaces on developer interfaces, routing services and local runners where the settings are on show. If a product does not offer it, that is a deliberate decision rather than an omission.

What the dial actually moves

Laying it out as a single axis makes the useful point, which is that nothing on this line is labelled correct. Every position produces fluent, confident text; what changes is how surprising the word choices are. That is why the setting disappoints the people who go looking for it hardest. Somebody whose assistant keeps inventing details arrives at the generation settings expecting a reliability control, turns this one down, and gets the same invented details in more measured prose, which is arguably worse because it now reads as considered. The failures that send people here live somewhere else entirely: the system was not given the material, or was given too many conditions at once, or was asked something its sources do not answer. Those are addressed by grounding, by splitting the request, and by checking the result. This dial is for tone, and it is good at tone.

Seen in the wild

A desktop application for running open models locally, where the generation and hardware settings sit on visible controls rather than behind flags.
LM Studio
A local runner exposing an interface other software builds against, so the generation settings are part of the request rather than an interface option.
Ollama
A routing service in front of many providers, where the same request and its settings reach whichever model a policy selects.
OpenRouter

Common misconceptions

People assume

Turning it down makes the system more accurate.

In fact

It makes the wording more predictable, which is a different property entirely. Invented facts, dropped conditions and misread documents all survive a narrow setting intact, so a steadier voice can make a wrong answer read more authoritatively rather than less. Accuracy is addressed by grounding and by checks, not here.

People assume

It should be tuned alongside temperature.

In fact

That is the classic mistake, because both act on the same choice from different directions. Moving them together produces effects nobody can attribute afterwards, which is why vendors generally advise adjusting one and leaving the other at its default. Pick whichever your interface exposes and change only that.

Telling them apart

Top-p vs Temperature

Top-p

Limits which candidates are eligible at all.

Temperature

Reshapes how likely each candidate is.

They pull the same lever by different mechanisms, which is exactly why moving both at once makes the result impossible to attribute to either.

Questions

Which should I change, this or temperature?: Whichever your interface exposes, and only one of them. They act on the same choice by different means, so changing both leaves you unable to say which produced the difference. If both are available, most teams settle on temperature simply because more documentation and more shared examples are written against it.
Will a lower setting stop it inventing things?: No, and this is the most common reason people find it. Invention comes from the system not having the information rather than from adventurous word choice, so a narrow setting produces the same fabrication in plainer language. Grounding it in real material and checking the output are what address that.
Why can I not find it in my assistant?: Because consumer products deliberately do not expose it. Generation settings appear on developer interfaces, routing services and local runners, where the audience is expected to understand the trade. Its absence from a chat window is a design decision about who the product is for rather than a missing feature.