Adaptive assessment systems typically rely on fixed difficulty labels or black-box models, which can make rating updates unstable and hard to interpret. The authors of a new arXiv preprint propose MARS, a Momentum-Aware Adaptive Rating System that represents the question bank as a directed weighted knowledge graph. Learner proficiency is updated using dynamic rating sensitivity, response quality, and a momentum-aware velocity term, giving the system a smoother and more interpretable rating trajectory than standard Fixed Elo.

The system also addresses a practical constraint: time-limited tests with negative marking. MARS frames question selection as a time-budgeted mark-maximization problem, choosing questions that fit within a fixed test duration while maximizing expected score. In experiments on ASSISTments 2009 and a simulated exam environment, MARS achieved the highest total score under this routing scheme, improving by 8.09 over the strongest baseline. It also reduced overtime violations from 50% to 5.6%, suggesting the policy is better at respecting time limits.

The authors argue that MARS is suitable for real-time, interpretable, and deployment-friendly adaptive assessment. Because the model is not a black box, instructors can trace why a particular question was selected, and the smooth rating updates may reduce erratic jumps in perceived ability. The paper does not compare MARS against every existing adaptive system, but its results on routing and stability indicate a promising direction for personalized testing under realistic constraints.