Objectives, activities and evaluations are linked in a "skills program" that the learner and agents adjust together: 12.19 out of 50 points earned, compared to 5.61 for comparison.
The major language models already know how to tutor, generate content or adapt a path. But to develop real skills, these resources must form a coherent process that evolves with the learner, which means aligning objectives, activities and assessments and relying on learning evidence to decide what to do next. A team that includes mathematician Weinan E proposes a "multi-agent organization with human participation", implemented in Agentic University.
The learner is treated as a member of the organisation: he or she works with specialised agents to maintain a reviewable skills programme, which links his or her personal objectives, activities on different media and a multi-level assessment. His feedback and results then guide the support offered.
In the authors' evaluation, the system does better on the organization of the program, the design of the evaluations and the quality of the revisions. Users improved by an average of 12.19 points on a scale of 50, compared to 5.61 in the comparison condition, and report better cognitive support. The code is published in open access on GitHub.
Read the full article on the original source
Go to the source →