Rice researcher proposes Pareto-based grading to test AI-resilient student skills

In a new paper on arXiv, Rice University computer scientist Anshumali Shrivastava tackles the challenge of assessing student work in an AI-native world. The core idea: instead of banning AI, grade students on their ability to achieve outcomes that surpass a strong AI baseline. This requires declaring an "AI-native Pareto frontier"—the best tradeoffs an advanced AI model can achieve on a given task—and then measuring each student's "Pareto surplus," i.e., the degree to which their submission improves upon that baseline. The grading certificate is executable and behavioral, not just subjective, though the author notes that surrounding protocols (design reports, prompt traces, oral checks) still provide context.

The paper provides a concrete example: an approximate-membership assignment centered on Bloom filters, used in Rice's COMP 480/580 course. Students must improve beyond AI-generated implementations to earn credit. The framework also addresses practical complications such as self-improving AI loops, budget neutrality, and server-mediated feedback. Shrivastava argues that this approach is fair—reducing the advantage of larger AI budgets—while still allowing students to use AI freely. The work is part of a growing effort to rethink academic assessment in an era of powerful, easily accessible AI tools.