Researchers Propose New C-test to Evaluate Supraintelligent AI

It is a test based on Levin Kt complexity that uses large language models as representational machines, generating inverse-inference questions with difficulty measurable in bits that can exceed current models and humans.

Existing benchmarks are often calibrated against human or existing model capabilities, making them circular or prone to saturation. The new test relies on the complexity of the underlying mechanism rather than human comparison, allowing generation of items harder than any frontier model.

The paper presents a theoretical framework and an initial generation pipeline, not a production-ready tool. Research teams and regulators can track it as a potential methodology for evaluating future models, but operational adoption requires independent verification and open tooling.