CMU Researchers Introduce Epiplexity: Rethinking Information for Computationally Bound AI

joecole · x · 2026-08-08

Researchers from CMU and other institutions published a new paper, From Entropy to Epiplexity, addressing the limitations of Shannon information theory and Kolmogorov complexity in evaluating data value for machine learning.

The authors identify three paradoxes in traditional information theory: deterministic transformations cannot increase information; information is independent of data order; and likelihood modeling is merely distribution matching. To resolve these tensions, they introduce Epiplexity, a formalization of information that captures what computationally bounded observers can actually learn from data. The metric isolates structural content while excluding time-bounded entropy, such as unpredictable outputs from pseudorandom number generators.

Original post →

More from Research

Research channel →