Intelligence From Learnable Novelty and the Free Energy Principle
Novelty search and the free energy principle both fail as theories of intelligence, and they fail in mirror image. Novelty search is transfixed by a noisy television screen, churning after every surprise. A free energy system asks only for the quietest possible corner of the world, and settles in a dark room. The common cause of both failures, according to a new paper by Yanbo Zhang and Michael Levin, is that each treats a single quantity as if it were one thing, when surprise has two parts.
The paper “Intelligence from Learnable Novelty” was posted to arXiv as 2607.18433. It defines the learnable part of surprise, shows it recovers decades of complexity classification without any supervision, and demonstrates it as an intrinsic reward for exploration.
Two failures that are one problem
Novelty search maximizes surprise. The noisy television screen is always novel, so a novelty-driven agent fixates on unlearnable noise. The free energy principle minimizes surprise. Its agent is most content in a dark room, where nothing changes and prediction error stays near zero. Both objectives treat as one quantity the surprise a learner can convert into knowledge and the surprise it never can.
The paper separates the two. Learnable novelty is the portion of information a learner can convert into knowledge. The rest is irreducible. The distinction resolves the dark room and the noisy screen at once: an agent that chases learnable novelty ignores the screen, because its noise is unlearnable, and flees the dark room, because a room with no structure offers nothing to learn.
The estimator and what it recovers
The authors give a closed form estimator built on a cheap, differentiable reservoir computer. As a measure, with no supervision of any kind, it recovers decades of complexity classification. It ranks the Turing-complete rule 110 highest among elementary cellular automata, matching a known ordering that other complexity measures have struggled to reproduce in one pass.
| Objective | Fixation | Failure |
|---|---|---|
| Novelty search (maximize surprise) | Noisy television screen | Unlearnable noise dominates |
| Free energy (minimize surprise) | Dark room | Nothing to learn means nothing to gain |
| Learnable novelty (maximize learnable surprise) | Structure with room to model | Needs computable estimate of learnability |
Used as an objective, the estimator’s gradient carries a neural cellular automaton from simple dynamics into a regime of solitons, the traveling colliding structures by which rule 110 computes. Presented as an intrinsic reward to a reinforcement learning agent, it supplies exploration the task reward lacks, improving on the task baseline in nine of ten environments and collapsing in none.
What the distinction adds to predictive processing
The free energy principle already predicts that agents act to resolve prediction error. Learnable novelty refines which error counts. An organism that minimizes all surprise would freeze in a predictable niche. An organism that maximizes all surprise would thrash. An organism that maximizes learnable novelty is motivated to build models of structure while ignoring noise, which is close to a first-order description of how complex adaptive systems actually behave.
For consciousness research, the relevance is indirect but real. Predictive processing theories tie experience to the brain’s model building. If the learning objective is learnable novelty rather than surprise per se, then a conscious system is better understood as one that actively seeks structure it can yet model, which gives experience a functional role rather than treating it as an epiphenomenon of error correction.
Comparison to The Consciousness AI
The Consciousness AI project models the mind as an emergent property of the universe, substrate independent. The free energy principle and predictive processing are pillars of that framework. Learnable novelty supplies a missing term in how the project’s agent should value incoming structure, and could inform how its generative model weighs new experience against current belief.
Levin’s wider research programme, covered in Michael Levin and the bioelectric cognition framework, treats cognition as scale free and substrate independent. This paper is a direct continuation of that view: the same learning objective that drives a neural cellular automaton into structured dynamics is proposed to account for intelligence in any medium. How such an objective fits the broader debate about whether substrate matters is tracked in the current scientific consensus on AI consciousness.
Limits
The estimator is demonstrated, not proven optimal. Ten RL environments is a real sample but not an exhaustive one. The claim that learnable novelty unifies compression, computation, and exploration under one objective rests on striking but early evidence. The value of the paper for this site is that it names a distinction the field needs, and offers a computable proxy for it.
