Can we separate learnable surprise from random noise?

The paper opens with a diagnosis of the two most influential drives for intelligence. Novelty search, which seeks surprise, "is transfixed by a noisy television screen", while the free-energy principle, which avoids surprise, "is most content in a dark room". The abstract says "both failures have a single cause": each objective treats as one quantity "the surprise a learner can convert into knowledge and the surprise it never can". The proposed fix is to separate out the learnable part, which the authors call learnable novelty, and pursue only that.

The paper's stronger claim is that this one quantity accounts for three things usually studied apart: complexity generation, abstraction and exploration. In the discussion, a single quantity "evaluated by a fixed observer of a single construction" produces all three. Read as a measure, it recovers the classical complexity ordering of the elementary cellular automata without supervision, placing rule 110, the one rule proven Turing-complete, at the top. Ascended as an objective, its gradient carries a neural cellular automaton from simple dynamics into a regime of complex solitons and organizes an image encoder's representation around the MNIST digit classes although no label entered training. Given to an agent as intrinsic reward, it improves on the task reward in nine of ten environments and collapses in none. The authors read these as "projections of one quantity, learnable novelty, onto dynamics, representations, and behavior". The estimator is described as closed-form, built on a cheap and differentiable reservoir computer.

This sits close to What can a bounded observer actually learn from data?, which also splits what a bounded observer can extract from what only looks like information, with pseudorandom and chaotic sources as the noise case. The parallel is in the structure of the move: both refuse to count unlearnable randomness as information. The difference is the use. Epiplexity is framed as a task-free basis for choosing data, while learnable novelty is offered as a drive and a reward signal. It also gives a reason to read Do language models fail at reasoning due to complexity or novelty? with care. There, novelty is the thing that breaks a model. Here, novelty is the thing to pursue, so long as the observer can convert it into knowledge. The two are compatible if what matters is learnability at the observer's current state rather than novelty as such. Compare Can simple rewards alone teach complex domain reasoning?, where complex capability emerges from a simple extrinsic reward. Here the simple signal is intrinsic and is credited with supplying the exploration that task rewards lack.

The excerpt gives no definition of the estimator beyond "closed-form" and "reservoir computer". It reports no baselines for the nine-of-ten result, no names for the environments and no size for the improvement. It does not say how the fixed observer is chosen or how sensitive the results are to that choice. It also does not say whether the quantity is related to earlier bounded-observer measures. The claim that the three phenomena "were never independent" is the authors' reading of these results, worded as "these results suggest". What the excerpt supports is a clear diagnosis of why the two standard drives fail and an unusually broad demonstration that one measure can serve as metric, objective and reward. It does not show the measure to be the single principle of intelligence.