XTC (Exclude Top Choices) sampling

XTC (Exclude Top Choices) is a sampler that occasionally removes the most predictable token candidates so the model picks a less obvious but still plausible one.

What XTC is

Most samplers remove unlikely tokens. XTC, also by p-e-w, does the opposite: it sometimes removes the most likely ones. The goal is to fight the cliched phrasing models fall back on (the "shivers down the spine" problem) without raising temperature to a level that also lets junk through.

How it works

XTC has two settings: a threshold and a probability.

On each token, with chance equal to probability, XTC activates. Otherwise nothing happens.
When active, it looks at every token whose probability is above the threshold.
If at least two tokens are above the threshold, it removes all of them except the least likely one.

Keeping the least likely of the high-probability group is the important part. There is always at least one good option left, so the model is never forced into a bad token. And if only one token clears the threshold (a name, a closing quote), XTC does nothing, which protects text that has one correct answer.

Effect on roleplay

At moderate settings prose becomes less formulaic and character voices sound less like the model's default narrator. At aggressive settings, coherence drops: the model keeps taking the second-best path, which adds up over long replies and can derail plot or instruction following. It also tends to make replies longer or shorter in odd ways, because the end-of-turn token can itself be excluded; some implementations protect newline and end tokens for that reason.

Sensible values

Threshold 0.1, probability 0.5: the commonly suggested starting point.
Lower the probability (0.2 to 0.3) for a light touch.
Raise the threshold to make XTC act less often; a threshold above 0.5 means it can never find two tokens and is effectively off.

Pair XTC with min P and a moderate temperature. It is not meant for structured output, code or tool calls.

XTC on Wild West API

XTC is a backend specific sampler, available in KoboldCpp, text-generation-webui, llama.cpp and similar local servers. It is not part of the OpenAI spec, so the API may ignore XTC fields unless the upstream model server supports them. Over a standard connection, the closest portable tools are temperature and top_p, plus writing guidance in your system prompt.

FAQ

What does XTC threshold mean?

It is the probability a token must exceed to count as a top choice. Only when two or more tokens exceed it does XTC remove all but the least likely of them.

Is XTC good for every model?

No. It helps models with strong stock phrasing, but on models that are already varied it mostly costs coherence. Test it with swipes on your own chats.

The DRY (Don't Repeat Yourself) samplerDRY (Don't Repeat Yourself) is a sampler that penalizes a token when choosing it would continue a sequence of text that already appeared earlier in the context. Min P sampling explainedMin P is a truncation sampler that removes every token whose probability is less than a set fraction of the most likely token's probability. Temperature: how it changes what the model writesTemperature is a sampling setting that makes a language model's word choices more predictable when lowered and more varied when raised. Mirostat sampling explainedMirostat is an adaptive sampler that adjusts itself token by token to keep the text's average surprise close to a target value.