Gemini last models: temperature, top_p, and top_k are deprecated and ignored

https://ai.google.dev/gemini-api/docs/latest-model

by greatgib • 1 day ago

127 43 1 day ago

ai.google.dev

greatgib3• 1 day ago

1. Sampling parameter deprecation (temperature, top_p, top_k)

temperature, top_p, and top_k are deprecated and ignored. In future model generations, supplying these parameters returns an HTTP 400 error. Remove these parameters from all requests.

tolugenius• 23 hours ago

To improve determinism, define a system instruction with explicit rules for your specific use case.

Is this guaranteed to work any better than top_k or top_p? This just sounds like making a smaller version of a Agent.md doc.

janalsncm• 21 hours ago

It is guaranteed to work worse than top_k=1, that’s for sure.

furyofantares• 20 hours ago

Depends on what you mean by determinism.

tough• 23 hours ago

fwiw sonnet-5 also drops temperature (sonne-4 had it)

aesthesia• 22 hours ago

My guess is that RL training being done with particular generation parameters makes models much more brittle to changes in these parameters, and that's why we're seeing changes like this across model providers. But I don't really know.

pixelmelt• 22 hours ago

I'm inclined to agree given how unstable Gemma 4 is when not using the "official" sampler settings

impulser_• 22 hours ago

Good. These have been basically useless for the past few generations of models, and most of the time made the model perform worst.

kouteiheika2• 21 hours ago

Obligatory "The Conspiracy Against High Temperature Sampling":

https://gist.github.com/Hellisotherpeople/71ba712f9f899adcb0...

NooneAtAll32• 21 hours ago

where can one learn what top_k and top_p mean?

cratermoon• 20 hours ago

https://www.ibm.com/docs/en/watsonx/saas?topic=prompts-model...

krapht• 20 hours ago

ironically, any frontier LLM will easily generate a tutorial at any detail you like explaining what these are.

if don't have time for that, just know that these are technical parameters that affect how likely it is an llm will produce the same result after being asked the same question.

imperio59• 19 hours ago

https://www.longformexplainers.com/llm-inference/10-sampling

nodja2• 16 hours ago

The posted answers are either behind a paywall or very obtuse so I'll just explain. I'll assume you know what tokens are.

A models output is not a single token, but a list with the probability for all the tokens that it knows, so we need to use a sampler to select the token that it's going to be the next token in the sentence. For example a simple greedy sampler will choose the token with the highest probability, but samplers normally pick a random token weighted by probability. A model usually knows about ~250 thousand tokens and the probability of some of these tokens are gonna be high, but the vast majority is close to but not actually 0% so there's a chance the sampler might pick some random token that doesn't make much sense, so we filter tokens.

top_k filters the tokens so that only the k top tokens are selected. So top_k=50 will filter those 250k tokens to only 50. This is assuming the list of tokens is sorted by probability.

top_p filters the top tokens until a percentage is accumulated. So if for example if you set the top_p to 0.6 and the model gave the top token a 0.5 (50%) probability and the second top token a 0.2, those 2 token accumulated to 0.7 which is greater than what you set it to (0.6) so no more tokens are selected. If this ran after top_k=50 it'll turn the list of 50 tokens into one of 2.

After each filter parameter is processed, the probability of the tokens is adjusted to sum to 1 (100%), Also note that order of operation here matters, i.e. top_p could be applied before top_k, but most providers follow what's on huggingface, I think I've only seen different implementation in certain local model hosting frameworks.

est• 15 hours ago

wow I wish I knew this post before I installed various UIs with tons of temp/top-X sliders.

They are extremely confusing.

gdiamos• 20 hours ago

thank god, these parameters are so confusing

salamo• 20 hours ago

Possible reasons:

- They might be dynamically adjusting these at inference time [1]. For example, start with a low temperature and generate samples with increasingly high temperatures until one of them passes some quality gate.

- They don't want you to fine-tune on high temperature completions (rejection fine-tuning). You could call this "rejection fine-tuning rejection".

[1] https://rlhfbook.com/c/09-rejection-sampling#related-best-of...

regularfry• 10 hours ago

Couple of other more businessy reasons:

- SynthID hides the watermark in the sampling RNG. No randomness -> no watermark.

- If you want to distil on the model outputs, you want temp=0 outputs. No temp=0 -> worse distillation.

orbital-decay2• 10 hours ago

You probably don't want temp 0, especially with Gemini which often fails in the greedy sampling mode in practice, often in ridiculous ways (e.g. multiple thousand token loops). As another comment says it's pretty brittle and this is even reflected in their docs somewhere IIRC. You want the "normal" temperature (whatever it is) and oversampling if necessary.

Certain open models have/had the temperature locked on the official APIs, I assume they just have sampling incompatible with static temperature or do some fancy speculative decoding. It's clearly not to hide anything, as the weights are open and there are always alternative providers.

charcircuit• 19 hours ago

Along with everything else. These parameters can make speculative decoding less accurate increasing the inference cost.

nichohel• 18 hours ago

"Last" or "latest"? Those are rather different.

froh• 16 hours ago

hehe... fwiw, the un-editorialized title is "Using the *latest" Gemini models"

I speculate OP wanted to put focus on their chosen detail in the title...

greatgib3• 2 hours ago

Indeed, the focus was on this section but it is direct link to the relevant paragraph.

For the "last/latest", it is the result of trying to trim the title enough to enter in the size limited title box and then I didn't notice the typo.

At first I wanted to put the 3.5/3.6 model number because "latest" is not clear and time dependant but it was too long then.

david-gpu• 10 hours ago

The OP does not appear to be a native English speaker. And, for example, languages like Spanish don't clearly differentiate between "last" and "latest".

henriquecm8• 10 hours ago

I am not a Spanish speaker, but Portuguese is usually similar. In it we say "último" for last, and "mais recente" for latest, so while there isn't a single word translating, we still have its equivalent. I imagine it's the same for Spanish.

herpderperator• 16 hours ago

I'm curious: If someone wanted to serve models off hardware/silicon directly (like Cerebras or Taalas, and soon Google I think) rather than GPUs, would these parameters still be adjustable at request time? Or would they have to decide that before the model can even start serving and it would be locked in until they reload it (which would make it briefly unavailable)?

StrauXX• 9 hours ago

I'd expect that sampling would happen in software. Probably the hardware system would output a vector of probabilities over the tokenspace, just as the nets do when run in software.

I don't know though and am not aware of any docs going into detail here. That being said, sampling is really cheap. So implementing it in hardware wouldn't be worth it.

franky47• 16 hours ago

To improve determinism, define a system instruction with explicit rules for your specific use case.

"Please be deterministic".

rf15• 13 hours ago

please oh god please, I don't know how to

Zambyte• 10 hours ago

In all seriousness, this seems like it could be fine if done well. You can just have a model do a pass over the system prompt and set reasonable parameters based on that. That's probably not what they're doing, but it could be.

dev_dan_2• 10 hours ago

improve determinism

Oh well. I might be to picky here, but how I see things, determinism cannot be improved or worsened, but achieved or not achieved. Or Partially archieved, when analyzing a system that has both components that are deterministic or non deterministic.

There are times to think in absolutes, and when talking about deterministic behavior of technical systems, this is one of them. Join the sith side, we have cookies, and when we say we have cookies, we do.

gregdaniels421• 7 hours ago

Probably what is meant is consistency instead of determinism

bob1029• 15 hours ago

To improve determinism, define a system instruction with explicit rules for your specific use case.

What if I want to do the other thing? When performing research with many sub agents, having a lot of diversity in the hypotheses is a big deal. If my 5 parallel sub agents all produce the same conclusion I might as well have only ran one.

The latest OAI models have done the same thing. I'm currently adding random variation to prompts to compensate for the lack of higher temperature sampling.

orbital-decay2• 10 hours ago

Current post-trained models are stochastically and semantically collapsed to a large extent, you'll never get meaningful semantic diversity if your prompt stays the same. This makes the temperature brittle, the model becomes dumb much earlier than meaningfully diverse. There are fancy samplers that claim to overcome this (comments ITT mentioned them), but a) you need a local inference stack to use them and b) I've yet to see any definitive evidence they add any deep semantic diversity without dumbing the model down, most of what I've seen is pretty superficial, including distribution-aware samplers.

I'm currently adding random variation to prompts

This is better than naive temperature tweaks, if done right.

marcinjachymiak• 8 hours ago

It's not surprising since these are all reasoning models now. Token sampling/decoding isn't exposed because the reasoning models are trained around specific configurations. That's why it's been replaced with "reasoning effort" instead (low, medium, high "thinking" levels).

kittikitti• 8 hours ago

The API endpoints are less useful if you can't even fine-tune these parameters. These endpoints used to be developer oriented and this sort of change would be a huge cause of concern. If you primarily vibe-code, this is not a problem but understand that they're making their product worse on purpose.

greatgib3• 2 hours ago

Same thing as the new APIs like assistant API that rip us of the control of the history of messages for the context that is now hidden server side.

Gemini last models: temperature, top_p, and top_k are deprecated and ignored

Related Stories