← Back to context

Comment by Der_Einzige

18 hours ago

Yet another example of "comments that are only sort of true because high temperature sampling isn't allowed".

If you use LLMs at very high temperature with samplers which correctly keep your writing coherent (i.e. Min_p, or better like top-h, P-less decoding, etc), than "regression to the mean" literally DOES NOT HAPPEN!!!!

11 comments

Der_Einzige

Reply
hnlmorg  16 hours ago

Have you actually tried high temperature values for coding? Because I don’t think it’s going to do what you claim it will.

LLMs don’t “reason” the same way humans do. They follow text predictions based on statistical relevance. So raising the temperature will more likely increase the likelihood of unexecutable pseudocode than it would create a valid but more esoteric implementation of a problem.

  • Terr_  15 hours ago

    To put it another way, a high-temperature mad-libs machine will write a very unusual story, but that isn't necessarily the same as a clever story.

    • balamatom  14 hours ago

      So why is this "temperature" not on, like, a rotary encoder?

      So you can just, like, tweak it when it's working against your intent in either direction?

      1 reply →

  • bob1029  15 hours ago

    High temperature seems fine for my coding uses on GPT5.2.

    Code that fails to execute or compile is the default expectation for me. That's why we feed compile and runtime errors back into the model after it proposes something each time.

    I'd much rather the code sometimes not work than to get stuck in infinite tool calling loops.

adevilinyc  18 hours ago

How do you configure LLM température in coding agents, e.g. opencode?

  • Der_Einzige  18 hours ago

    You can't without hacking it! That's my point! The only places you can easily are via the API directly, or "coomer" frontends like SillyTavern, Oobabooga, etc.

    Same problem with image generation (lack of support for different SDE solvers, the image version of LLM sampling) but they have different "coomer" tools, i.e. ComfyUI or Automatic1111