19 Mitigating Output Variability

An important aspect of using generative AI models for educational tasks is managing output variability. As we have seen (and demonstrated) several times so far, when using a model with default parameter settings, you should expect slightly different responses from the model, even when using the exact same prompt. These models are not deterministic in the same way as regression models or many other statistical techniques. Instead, generative models produce a probability distribution over possible next tokens and then sample from that distribution.

Sometimes, the differences are minor. Word choices or sentence structure may vary, but the semantic meaning is generally the same. However, sometimes the variation can subtly shift the meaning. For example, if you are summarizing notes about a student’s performance that equally highlight both knowledge and engagement, one summary might emphasize knowledge, another might focus more on engagement, and a third might balance both equally.

There are several factors that influence response variability. The first is output length; longer responses usually¹ tend to show more variation. Second is prompting method—reasoning models, or prompting methods that encourage more extensive reasoning, may introduce more response variability. Another factor is task complexity—the more complex the task, the more room there is for variation in the model’s response. For example, “Write a paragraph on symptom overlap between cardiological and respiratory illnesses” will likely show more variation than “Identify the correct multiple-choice option.”

This variability is not inherently bad. In fact, it is beneficial for creative tasks like brainstorming or ideation. But it can be problematic for scoring tasks, where consistency is critical.

I developed a 3-part framework that helps me think about sources of variability in model responses:

Model-based strategies
Prompt-based strategies
Materials-based strategies

19.1 Model-based strategies

These strategies are specifically focused on changing the model’s generation parameters. As we discussed earlier and experimented with in an activity, lowering the temperature can encourage the model to select the same or similar tokens when generating a response.

19.2 Prompt-based strategies

These strategies focus on how you structure your prompts and workflow. Certain prompting techniques, such as using examples, using punctuation or formatting to separate sections, or breaking tasks into steps, can all lead to more consistent outputs.

19.3 Materials-based strategies

These strategies relate to the content within your prompt. While closely related to prompt-based strategies, I find it useful to treat them as a separate category. This includes things like providing more context (filling in the Mad Lib blanks) and clarifying or editing the materials included in the prompt.

I have found that this also highlights another good use of generative models: rubric refinement. The model can serve as a proxy for how a reasonably intelligent reader might interpret the rubric, so passing an assessment artifact and rubric through the model several times can help identify potential ambiguity or places in need of clarification.

I observed this in some of my own research. In one study, I had an analytic rubric element that read: “Tenderness to deep palpation on the right medial heel: 1 point.” In one of the notes created to see how well the model could detect this concept, I wrote: “A musculoskeletal exam revealed some pain during a deep palpation on the right heel.” I intended this to fully represent the concept. I found that the model applied the rubric in 3 different ways (paraphrasing the model’s rationales):

Full credit; 1 point.
Learner did not say “medial”; 0.5 points (which was not part of the rubric, although other elements had partial credit).
Learner did not say “medial”; 0 points.

This can easily be corrected by changing the rubric to include guidance for this edge case: “Tenderness to deep palpation on the right medial heel: 1 point. It is not necessary for the learner to include the word ‘medial.’”

I liberally use the “usually” qualifier because, as models improve, these factors influencing variability may become less of an issue.↩︎

# Mitigating Output Variability An important aspect of using generative AI models for educational tasks is managing output variability. As we have seen (and demonstrated) several times so far, when using a model with default parameter settings, you should expect slightly different responses from the model, even when using the _exact same prompt_. These models are not deterministic in the same way as regression models or many other statistical techniques. Instead, generative models produce a probability distribution over possible next tokens and then sample from that distribution. Sometimes, the differences are minor. Word choices or sentence structure may vary, but the semantic meaning is generally the same. However, sometimes the variation can subtly shift the meaning. For example, if you are summarizing notes about a student’s performance that equally highlight both knowledge and engagement, one summary might emphasize knowledge, another might focus more on engagement, and a third might balance both equally. There are several factors that influence response variability. The first is output length; longer responses usually[^1] tend to show more variation. Second is prompting method—reasoning models, or prompting methods that encourage more extensive reasoning, may introduce more response variability. Another factor is task complexity—the more complex the task, the more room there is for variation in the model’s response. For example, “Write a paragraph on symptom overlap between cardiological and respiratory illnesses” will likely show more variation than “Identify the correct multiple-choice option.” This variability is not inherently bad. In fact, it is beneficial for creative tasks like brainstorming or ideation. But it can be problematic for scoring tasks, where consistency is critical. I developed a 3-part framework that helps me think about sources of variability in model responses: - Model-based strategies - Prompt-based strategies - Materials-based strategies ## Model-based strategies These strategies are specifically focused on changing the model's generation parameters. As we [discussed earlier](08-gen-ai-parameters.qmd#sec-gen-params) and [experimented with in an activity](activity-parameter-testing.qmd#sec-act-parameter), lowering the `temperature` can encourage the model to select the same or similar tokens when generating a response. ## Prompt-based strategies These strategies focus on how you structure your prompts and workflow. Certain prompting techniques, such as using examples, using punctuation or formatting to separate sections, or breaking tasks into steps, can all lead to more consistent outputs. ## Materials-based strategies These strategies relate to the content within your prompt. While closely related to prompt-based strategies, I find it useful to treat them as a separate category. This includes things like providing more context (filling in the Mad Lib blanks) and clarifying or editing the materials included in the prompt. I have found that this also highlights another good use of generative models: **rubric refinement**. The model can serve as a proxy for how a reasonably intelligent reader might interpret the rubric, so passing an assessment artifact and rubric through the model several times can help identify potential ambiguity or places in need of clarification. I observed this in some of my own research. In one study, I had an analytic rubric element that read: "Tenderness to deep palpation on the right medial heel: 1 point." In one of the notes created to see how well the model could detect this concept, I wrote: "A musculoskeletal exam revealed some pain during a deep palpation on the right heel." I intended this to fully represent the concept. I found that the model applied the rubric in 3 different ways (paraphrasing the model's rationales): - Full credit; 1 point. - Learner did not say "medial"; 0.5 points (which was not part of the rubric, although other elements had partial credit). - Learner did not say "medial"; 0 points. This can easily be corrected by changing the rubric to include guidance for this edge case: "Tenderness to deep palpation on the right medial heel: 1 point. It is not necessary for the learner to include the word 'medial.'" *** [^1]: I liberally use the "usually" qualifier because, as models improve, these factors influencing variability may become less of an issue.