Beyond Benchmarks: A New Intelligence Metric

The relentless pursuit of more capable AI models often gets bogged down in the limitations of static benchmarks. These benchmarks, while useful, can lead to models that overfit to specific tasks or datasets, failing to capture a true measure of general intelligence. This paper introduces a novel approach to bridge this gap.

Formalizing Indistinguishability as Intelligence

The core innovation presented is the Generalized Turing Test (GTT), a formal framework designed to compare arbitrary agents based on their indistinguishability. The GTT defines a comparator where agent B can reliably distinguish between interactions with agent A (instructed to imitate B) and another instance of B. This establishes a dataset- and task-agnostic measure of relative intelligence. The researchers explore the structural properties of this comparator, including conditions for transitivity, which allows for the induction of an ordering over equivalence classes of intelligence. Variants with modified interaction protocols, such as querying or bounded interactions, are also analyzed, offering flexibility in evaluation.

Empirical Validation of Stratified Intelligence

To ground the theoretical framework, the authors instantiate the GTT on a suite of modern AI models. Through thousands of pairwise indistinguishability trials, they empirically evaluate the proposed comparisons. The resulting data exhibits a discernible stratified structure, aligning with existing intuitions and rankings of model capabilities. This empirical evidence suggests that the GTT framework yields meaningful relative orderings of intelligence, moving beyond the limitations of traditional benchmarks.