New Paper Unifies Intelligence Concepts via Learnable Novelty

The new preprint "Intelligence from Learnable Novelty" proposes a unified framework for intelligence. It attempts to bridge the differing definitions of intelligence across statistics, machine learning, complex systems, and agent behavior into one core principle: the pursuit of learnable novelty. The authors derived a closed-form approximation for Epiplexity, arguing its deep connection to "intelligence."

Confirmed

The paper posits that intelligence is not merely about experiencing "surprise," but rather splitting novel experiences into "learnable" and "unlearnable" components, optimizing only for the former. The authors propose the epiplexity rule to provide an unsupervised exploration goal for AI, ensuring it pursues only "truly learnable novelty" or "learnable surprise." According to @MacrinePhD, this rule directly addresses two common failure modes in AI exploration: getting distracted by meaningless noise (the "Noisy TV" problem) and becoming lazy in sparse environments (the "Dora the Explorer" problem). By introducing this rule, the system balances these extremes. In practical applications, @menhguin highlighted the paper's findings in Reinforcement Learning (RL): incorporating learnable novelty into RL tasks improves final returns across various tasks and effectively accelerates training convergence.

Why it matters

This research provides a clear mathematical and theoretical framework for agent exploration in unsupervised environments. By distinguishing between "learnable" and "unlearnable" novel experiences, it not only addresses long-standing exploration efficiency bottlenecks in RL but also offers a highly inspiring, unified perspective for cross-disciplinary discussions on the essence of "intelligence."

2026-07-22 ~ 2026-07-23 · 7 related posts

Primary sources

2 near-duplicate retellings: MacrinePhD · MacrinePhD