AI Cosmos explores how AI-assisted practice can leave evidence of what a learner did

AI Cosmos proposes a way to retain evidence of AI-assisted practice for assessment.

Researchers introduced AI Cosmos in an October 29 preprint on situated learning. The design places learners and assistants with specific roles in task scenarios, with relevant material and executable workspaces where needed. The system keeps parts of the work and its outcomes as reusable records, including information about where those materials came from and what they depend on. [AI Cosmos preprint](https://arxiv.org/abs/2610.40031) · [Research team repositories](https://github.com/THU-AICosmos)

Finishing a task and learning from it require different evidence

A learner may complete an assignment with help from an assistant and still struggle to apply the idea alone later. The paper starts with that distinction. Its scenario approach makes the activity concrete. An assistant can have a specific role and the learner can work with materials that resemble the task, instead of relying on an unrestricted chat.

Saving a record of the work gives a teacher or learner a way to revisit decisions. For example, a later exercise could ask the student to explain why a previous approach worked and adapt it to new inputs. That makes the saved record a starting point for another activity rather than a substitute for understanding. [AI Cosmos preprint](https://arxiv.org/abs/2610.40031)

An evaluation needs an independent follow-up task

A pilot can compare how learners perform with the assistants, then ask them to solve a related problem after the assistance is removed. Looking at both results can show whether the activity improved immediate performance, independent skill or both. It also matters whether the saved work is understandable to the learner and whether they can identify which parts they produced.

The preprint proposes an infrastructure and teaching approach. Its long-term value in a particular setting depends on learning outcomes measured over time and how the tools fit the course. [AI Cosmos preprint](https://arxiv.org/abs/2610.40031)