Generate Variants with AI

Writing a second version of an exercise by hand is the expensive part of exercise variants. For a programming exercise it means re-theming the problem statement, template, solution and tests consistently, keeping every task-to-test link intact, and making sure the solution still builds green while the template still fails.

Artemis can do that work for you. Create Variant with AI takes an existing exercise, asks what should change, and produces a complete, verified Artemis exercise in the background.

ℹ️

Information:

Variant generation is part of Hyperion and is available only when your Artemis instance has the Hyperion module enabled. It requires at least the Editor role in the course.

Supported Exercise Types

Exercise type	Supported
Programming	Yes
Quiz	Yes, unless the quiz contains drag-and-drop questions
Modeling, Text, File Upload	No

The action is hidden for exercises it cannot handle, and for anyone below the Editor role.

Starting a Generation

On the course's Exercises page in course management, the overflow menu (…) of an exercise row offers Create Variant with AI. In exams, the same action sits on the rows of an exercise group.

The AI action in the overflow menu of an exercise row

A wizard opens and walks through the five stages below in order.

Select What Should Change

Pick one or more aspects. You can combine them freely.

Choose the aspects the AI should adapt

Aspect	Effect
Difficulty	Re-tunes how hard the exercise is. The variant keeps the core concept and the clear majority of its tests — it is never gutted, and a task never becomes easier by handing students working code
Application Domain	Moves the exercise to a different real-world theme. Algorithm, data-structure and design-pattern names stay as they are, because they are the learning content
Storytelling	Sets the narrative style, from a plain technical description to a fully told story
Custom	Free-text instructions of your own

Configure the Aspects

Fill in the details for the aspects you selected: the target difficulty, the new domain, the narrative style, and your custom instructions.

The narrative styles are:

Style	Description
Technical	No story — a plain, concise focus on the technical concepts
Realistic	A short real-world scenario introduces the task, the rest stays technical
Creative	A story carries the task: themed setting, named actors, story-driven examples

Choose the Placement

Decide where the finished variant goes.

Placement: create a group with the original, join an existing group, or stay standalone

Option	Result
Create new group with original	A new variant group holding the source exercise and the variant. You set the group title, points cap and timeline right here
Add to group "…"	The variant joins the group the source exercise already belongs to
Add as standalone exercise	No group — the variant appears on its own in the exercise list

Follow the Generation

The wizard shows the phases as they happen: the source exercise is analyzed, the change is planned, a real copy is provisioned, the changes are applied, and the result is verified and — if needed — repaired.

The live timeline of a running generation

A programming variant runs real template and solution builds, so a generation typically takes a few minutes. Quizzes are faster.

Click Run in Background to close the window and let the job continue. It then appears in the tray in the navigation bar, where you can follow it, reopen the full view, or cancel it.

Background generations in the navbar tray

Canceling ends the job and removes the half-finished exercise copy, including its repositories and their builds. Jobs stay in the tray for 24 hours so you can review a result later.

Review the Result

A finished variant with every check green

The result step summarizes what the AI did per phase, how many tokens the run used, and offers Open in Editor to review the variant.

What Artemis Checks

Before a variant is accepted, it has to pass objective gates. They run cheapest first, and every finding goes back into the next repair round. The first four are hard gates: a variant is never accepted while one of them reports a problem. The AI-backed checks that follow are best-effort — when the model is unavailable, fails or exceeds its time budget, the check contributes no finding for that round and the run continues without it:

• The solution builds and passes all tests.
• The template still fails, so students actually have work to do.
• Every task-to-test reference in the problem statement resolves to an active test case, so grading stays wired up.
• Quiz validity: questions, answer options and points are consistent.
• A semantic consistency check (best-effort) compares problem statement, template, solution and tests against each other (the same check described under Consistency Check).
• A quiz critique (best-effort) reviews the generated questions for wording and answer quality.

Outcomes

Outcome	Meaning
Completed	Every hard gate passed, and every best-effort check that ran reported nothing. The variant is placed as configured and ready for review. A best-effort check that could not run is skipped silently, so review the variant yourself rather than reading Completed as a full semantic sign-off
Draft with warnings	Some check did not pass within the attempt budget, or the placement failed. The exercise is created all the same, with a short note on what is wrong — open it in the editor and fix that. A warning about the placement means the variant is not in the group you chose, so move it there yourself
Failed	The run could not produce a variant. The result step explains the phase it failed in and how to continue, and the half-provisioned exercise is deleted
Canceled	You stopped the run, and the half-provisioned exercise is deleted

⚠️

Warning:

A generated variant is not held back from students. A course variant is a real exercise and copies the dates of the exercise it was generated from, whatever the outcome — Completed as much as Draft with warnings. Generate a variant of an exercise that is already released and the variant is released too, before you have reviewed it. Generate variants before releasing the original, or postpone the variant's Release Date as soon as the run finishes; for a variant placed in a group, that date belongs to the group.

ℹ️

Information:

Deleting the half-finished copy of a failed or canceled run can itself fail — if the repository or the build cannot be removed, the exercise survives. Artemis says so on the result step and links the copy with Open in Editor, so that a run reporting a cleanup problem is one to follow up on: delete that exercise by hand.

⚠️

Warning:

AI-generated content may contain errors. Review a variant carefully before releasing it to students — the gates verify that the exercise builds and hangs together, not that it teaches what you intended.

Exam Exercises

Generation works the same way for exercises inside an exam exercise group, with two deliberate differences:

• There is no placement step. The variant always joins the exercise group of its source, which is what makes it usable as a per-student variant, and it inherits the exam's timing.
• There is no difficulty option. Exam difficulty is reviewed in advance, and handing students variants of differing difficulty would be unfair.

For exam exercises the wizard has no placement step and no difficulty card

Troubleshooting

The Action Is Not Visible

Check that the Hyperion module is enabled on your instance, that you have at least the Editor role in the course, and that the exercise is a programming exercise or a quiz without drag-and-drop questions.

The Run Ended as a Draft with Warnings

Open Show technical details on the result step. It names the gate that did not pass — most often a build that stayed red or a task reference that no longer resolves to a test. Fixing that in the editor is usually a small edit.

The summary above it is written by the model, so read it as a description of what went wrong, not as instructions: the exercise already exists by then, and there is no later publish step to reach. If one of the warnings is about the placement, the variant is not in the group you chose and you have to move it there yourself.

Generation Is Slow

Programming variants trigger real template and solution builds in every verification round. If your build agents are busy, the run waits for them.