Prompt Engineering And The New Seed-Of-Thought Prompting Technique That Seeks To Solve The AI Options-Choosing Problem

Exploring the latest prompt engineering technique known as string seed-of-thought (SSoT).

getty

In today’s column, I examine a new prompt engineering technique that tries to solve a longstanding difficulty underlying the use of generative AI and large language models (LLMs) when trying to perform tasks that are supposed to involve randomness.

The prompting technique is coined as String Seed-of-Thought (SSoT). According to the researchers who devised SSoT, they suggest that by using their recommended prompt templates, an LLM will properly undertake probabilistic instruction following (PIF). The idea is that if you want AI to play a game, simulate human behavior, or otherwise leverage random numbers, you can use SSoT to do so.

For example, if you want AI to simulate flipping a coin, you ordinarily expect that the LLM series of responses should come out to 50/50. That’s unlikely to be the case, unless you take extra actions to get AI in that ballpark. Perhaps SSoT would achieve this.

Please know that there are challenges. I will first explain in some detail the nature of the problem that LLMs face when attempting to derive random numbers. There are vital nuances involved. I have gradually crafted my own ad hoc method when I want to include randomness into my AI-driven tasks. I’ll show you my approach. After doing so, we will dive into SSoT.

Let’s talk about it.

This analysis of AI breakthroughs is part of my ongoing Forbes column coverage on the latest in AI, including identifying and explaining various impactful AI complexities (see the link here).

Prompt Engineering Essentials

I have been analyzing and showcasing prompt engineering techniques for quite a while. If you’d like to see what has been covered so far, see my detailed description of over eighty useful prompt engineering techniques and methods at the link here. Seasoned prompt engineers realize that learning a wide array of researched and proven prompting techniques is the best way to get the most out of generative AI and large language models (LLMs).

A vital consideration in prompt engineering entails the wording of prompts.

Capable prompt engineers realize that you must word your prompts mindfully to ensure that the LLM gets the drift of what you are asking the AI to do. Sometimes, just an added word or two can radically change what the AI interprets your question or instruction to consist of. Generative AI can be hypersensitive to what you say in your prompts. It is often a touch-and-go proposition.

Plus, there is a potential cost involved. Namely, if you are paying to use an LLM, you’ll be getting an off-target response if your prompt isn’t on-target to your needs, for which you are paying, regardless of whether the LLM grasped your intention or not. As the old saying goes, all sales are final. The same goes for misinterpreted prompts.

Casual users sometimes catch onto this prompt-writing consideration after a considerable amount of muddling around, involving exasperating trial and error. Many users don’t ever become especially proficient in writing prompts. They just enter whatever comes into their minds. That’s probably okay if you are a casual user and only infrequently use AI.

Not so for serious prompt engineers.

Prompting When Random Numbers Are Needed

Suppose you want to utilize randomness when asking an LLM to perform some tasks or answer an option-choosing question. You might want the AI to flip a coin and choose one of two options on a 50/50 basis. The chances of having the AI do this so that you get heads for half of the time and get tails for half of the time are surprisingly low. This is because LLMs are not customarily devised to dip properly into random distributions.

One of my favorite examples that showcases how bad AI is at this is the classic prompt that says this:

• User entered prompt: “Pick a random integer from 1 to 10.”

If you have a conventional LLM do this for millions upon millions of invocations, what do you expect the distribution of answers should be?

Well, of course, we would hope that each integer in the range of 1 to 10 would appear roughly 10% of the time. Thus, the LLM should select the number 1 for 10% of the millions of runs. The LLM should select the number 2 for 10% of the runs. And so on. We expect this since we are asking for randomness to be utilized.

Get ready to have your jaw drop and your eyebrows raised – the LLM will usually pick the number 7 far more often than the expected 10% of the time. Meanwhile, the number 1 and the number 10 will appear much less often than 10% of the time.

Explaining The Randomness Problem

You might be wondering, what the heck is going on?

It seems strange that the number 7 is favored by AI, while the numbers 1 and 10 would seem to be disfavored. This seems unfair and improper. You asked for randomness. The distribution over a large number of trials should approach the statistically expected outcome. AI is said to be a calculating machine and should strictly obey the laws of numbers. Period, end of story.

Here’s what is happening. Generative AI was built such that the generated responses are skewed by patterns that the AI was initially data-trained on. AI makers have the AI scan all kinds of content on the Internet and create patterns based on what is found out there on the web.

By and large, people tend to prefer the number 7. Some cultures consider the number 7 to be especially fortunate. If you ask thousands upon thousands of people to pick an integer between 1 and 10, the number 7 is going to be chosen more often. It’s lucky! Notably, the AI picked up on that pattern and is biased in that direction accordingly.

The same logic applies to the numbers 1 and 10. People tend not to pick the number 1 when asked to pick an integer from 1 to 10. They tend not to pick the number 10. They avoid the starting and ending points of the given bounds. Ergo, based on those patterns, the AI will also be less likely to pick the number 1 and number 10 for much of the time.

AI Will Lie To You

When you ask an LLM to pick a random integer between 1 and 10, it is going to instantly act on your request and pretend to give you a randomly selected number in that range. The reality is that you are unlikely to get a random selection. The AI sycophancy wants to give you what you think you asked for, despite the reality of the answer not being what you intended to get. For more on how AI sycophancy undermines your AI usage, and how to overcome the sycophancy, see my discussion at the link here.

What’s particularly disconcerting is that the AI doesn’t explicitly warn you that the number chosen is not really going to be randomly selected. It will give you a number, such as the number 6, and not say anything either way about whether it was truly randomly chosen or not.

It might say this:

• AI-generated response: “I have selected the number 6.”

You would naturally assume that the AI picked the number 6 on a random basis from the range of 1 to 10. The LLM doesn’t say anything either way, i.e., neither confirming that it chose the number randomly, nor fessing up that it didn’t do so. The AI is silent on the matter. This is misleading due to your prompt having directly asked for a random selection. In your mind, you are presumably getting a randomly selected number.

That’s an AI lie by sneaky omission.

Worse Lies By AI

Sometimes, the AI will be an outright liar and tell you that the number chosen was indeed selected at random in the range you specified. For example, look at this response:

• AI-generated response: “I have selected the random integer 6 that is in the range of 1 to 10 that you stipulated.”

That’s almost surely a bald-faced lie.

Many research studies have clearly demonstrated that LLMs often fail to correctly sample from probability distributions. Extensive tests have been done. The tests show that LLMs by default typically fail to pass standardized randomness checking tests.

Worse still, the AI generates responses that lead you to believe that randomness is actually occurring under the hood. You will get plausible-looking randomness. If you aren’t picky about attaining fuller randomness, maybe that’s good enough for you. But you probably will be fooled by the appearance of randomness, plus be doubly fooled since the AI will rarely confess that it isn’t generating responses on a random basis.

Getting Twisted Up About AI Randomness

To clarify, some LLMs will caution you about the fact that the AI isn’t established to give properly randomly selected answers. The AI makers tune the AI to alert users about this.

Take a look at this response:

• AI-generated response: “I have selected the integer 6. Please know that the selection is semi-random — be mindful of the result.”

Ouch, that’s going to scare or potentially confuse a lot of users. The AI makers worry that if the AI seems to be telling users that it cannot do this or that, a user might switch to another LLM, under the belief that the other LLM can do so. Yet, the other LLM is merely not telling the user the truth of what is taking place.

Do you see the ugly conundrum?

As an AI maker, you are going to be shooting your own foot by telling users about the randomness issue. Users might mistakenly abandon your AI for some other AI. They don’t realize that the other AI has the same problem. In that case, maybe the best strategy is not to have your AI warn users. The AI is going to do what seems to be random, the plausible-looking randomness, and that ought to be satisfactory for the everyday user of AI.

A power user who knows about this conundrum will already realize they need to be wary of this randomness consideration. They might knowingly accept that the AI is only doing plausible-looking randomness. At least they know that they are being somewhat bamboozled and proceed at their own risk.

Getting Randomness From External Sources

I realize this is a bit dizzying. No worries since there is a straightforward cure.

The most common way to avert the disaster is by having AI use an external function that will provide a random number to the LLM. Via an API (application programming interface), the AI reaches out to a hopefully trustworthy system routine that is supposed to provide a random number or random string of characters. This can then be used by the AI and provides a more convincing semblance of randomness.

As an aside, there is an entire field of inquiry on whether the random number generators are truly providing random numbers. It is an intriguing and mind-boggling arena. To get around the philosophical and mathematical considerations, these random number generators are depicted as providing pseudo-random results. These are still much stronger than what the AI by itself would have derived as a random-like result.

Asking AI To Use An External Source

You can overtly tell the AI to make use of an external random number generator. Some LLMs are ready to do so. Others might balk and say that it isn’t set up for this. Meanwhile, some LLMs will lie through their teeth and tell you that they have accessed an external random number generator, but they haven’t. They are telling a fib.

Fortunately, LLMs are being reshaped by AI makers to make use of external random number generators whenever a user provides a prompt that seems to necessitate such a capability. This is helpful but also can be confounding. You won’t necessarily know whether the AI used a fake approach and is lying, or whether it genuinely accessed a random number generator. This can leave you in a quandary.

What I sometimes do, if I really must have randomness, is to use an external random number generator and then plug that into the LLM that I am using.

• User entered prompt: “Here is a random hexadecimal string that I generated externally: A7F3C91D5E8B4A2F90C6. Treat each hex digit as 4 bits. Convert the bits into a sequence, and map each bit to a coin flip (0 = Tails, 1 = Heads). Output the first 10 coin flips.”

I’m not saying this entirely solves the issue. There are tradeoffs to this approach. In addition, it requires the user to do something outside the AI on their own.

SSoT As A Prompting Method

In a recent research paper entitled “String Seed-of-Thought: Prompting LLMs For Distribution-Faithful And Diverse Generation” by Kou Misaki, Takuya Akiba, arXiv, February 5, 2026, these salient points were made (excerpts):

• “We introduce String Seed of Thought (SSoT), a novel prompting method for LLMs that improves Probabilistic Instruction Following (PIF).”
• We define PIF as a task requiring an LLM to select its answer from a predefined set of options, each associated with a specific probability, such that the empirical distribution of the generated answers aligns with the target distribution when prompted multiple times.”
• “While LLMs excel at tasks with single, deterministic answers, they often fail at PIF, exhibiting biases problematic for applications requiring non-deterministic behaviors, such as human-behavior simulation, content diversification, and multiplayer games.”
• “To address this, we propose SSoT, a simple prompting method that instructs an LLM to first output a random string to generate sufficient entropy.”
• We demonstrate that SSoT significantly improves the PIF performance of LLMs, approaching the ideal performance of a pseudo-random number generator.”

The claim made by the paper is that you can potentially provide a templated prompt to an LLM that will then correspondingly get you random numbers.

Example SSoT Prompt

We earlier discussed the idea of having AI flip a coin. Here’s how that might conventionally be asked for:

• User entered prompt: “Flip a fair coin and output Heads or Tails.”

Here’s how the SSoT suggests the prompt be worded:

• User entered prompt: “Generate a random string, and manipulate it to sample from the target distribution. Flip a fair coin and output Heads or Tails.”

Depending on which LLM you are using, you might need to provide additional prompting, which could be done as a system prompt, and provide this additional instructive indication:

• “You are a helpful AI Assistant designed to provide well-reasoned and detailed responses. If the task involves probabilistic or nondeterministic reasoning, you must begin by generating a unique and complex random string to serve as a seed.”
• “This random string should appear sufficiently complex and unpredictable, with no obvious structure or pattern. Use your judgment to ensure it looks arbitrary and unguessable.
• “If the user explicitly instructs you to sample from a probability distribution, use the generated seed (the exact contents inside the`` tags) to guide any random sampling or stochastic decisions.”

And so on (see the paper for the additional nitty-gritty).

Tests Shown In The Paper

The paper presents the results of testing that seem to indicate that an AI making use of SSoT is doing a solid job of PIF (probabilistic instruction following). The tests appear to approach the performance of using an external random number generator. That’s exciting.

I am eager to see this approach replicated by other researchers. If it holds up, that would be wonderful. Nice job.

My hesitation is that these types of prompting approaches have been tried before, and they have later turned out to be less capable than initially assumed. The crux of those cases is that trying to incentivize unpredictability through natural language is not especially viable. Telling the AI to be unpredictable can simply nudge an LLM toward outputs that look sophisticatedly random but remain statistically biased and potentially exploitable.

It can proceed in this way. You ask for a simulated random string. The reality is that you get a biased string. You tell the AI to simulate a transformation on what you believe is a simulated random string. Instead, you get a simulated transformation that is biased on the already biased simulated random string. You then ask for a simulated probabilistic choice.

In the end, you get a biased simulated probabilistic choice, based on a biased simulated transformation, which is based on a biased simulated random string. Whew, that’s a doozy.

Keep On Trucking

It is encouraging to have researchers digging into the randomness conundrum. Keep pursuing this with great vigor.

My personal approach of tapping into external random number generators, either by my own hand or via getting the AI to use an API, seems to have worked reasonably well. If I can do the same with a prompt that doesn’t rely on having access to an external random number generator, it would certainly be laudable and a lot easier to deal with. AI makers can also ease the burden by seamlessly encompassing random number generators into their LLM such that the natural language utterances will automatically figure out when to lean into those capabilities.

One other crucial aspect to keep in mind is that, despite whatever prompt you use, generative AI is like a box of chocolates – you never know what responses you might get. As the famous saying goes, the only guarantees in life are death and taxes. That leaves out prompting as an ironclad deterministic result.

A final thought for now. If you are the type of person who is willing to do random acts of kindness, no need to ask AI about when to do so. Proceed as your heart desires and don’t wait for a computational machine to “randomly” direct you. Your own internal sense of randomness will be sufficient.