E.10.INT:11 - SoTA-Echoing

Learning progress. Schmidhuber’s historical formal theory of creativity and intrinsic motivation distinguishes improvement of a predictor or compressor from the unpredictability of its input. Adopt that comparison for progress claims in §4.2. The competing shortcut, fixed-model surprise, can reward an unlearnable signal; §5.1 makes the related information-seeking error actionable.

Information-seeking policies. Paprika describes strategic information gathering for task completion without an intrinsic-motivation reward. Its training curriculum uses a coefficient-of-variation heuristic for sampling tasks. Adapt the policy distinction in §4.1; keep that sampling heuristic separate from a measurement of acquired capability.

Active inference. Friston et al. (2017), §2, supplies the historical coupling of epistemic and pragmatic value. Their 2025 technical note extends information-seeking to distinctions among model structures. Adapt those contributions in §4.2: recover the model, expected information and preferences behind the action. This supplies a richer continuation than explaining all exploration as aversion to an unexpected observation; the selected formal account still determines its assumptions.

Open-ended search. Stanley and Lehman’s Why Greatness Cannot Be Planned, Chapters 5 and 9, gives the historical argument for judging available stepping stones without requiring their final destination. Enhanced POET, §§2-3, combines a population-relative difficulty screen with novelty and transfer. Adopt the distinction between an attainable learning opportunity and retained possibilities in §§4.1-4.3. A progress-only selector can lose the latter; the particular search method must supply its generation, retention and resource rules.

Agent-relative challenge. Oudeyer’s 2026 preprint review compares curiosity processes across timescales and examines conditions for an intermediate-knowledge preference. Adapt this conditional treatment of the Goldilocks region. It improves on an agent-independent midpoint rule by retaining the learner, environment and changing knowledge.

Rhythmic interest. Toussaint’s The Geometry of Musical Rhythm, second edition, Chapter 41 and Epilogue, develops an account using several rhythmic properties and their combinations. Spiech et al. (2025) compare groove ratings across complexity and meter in three behavioral experiments. Adapt their listener-and-meter distinction in §5.3. It provides a more useful selection basis than a universal complexity optimum, while leaving rhythm construction and teaching to their field methods.

Reconsider these comparisons when a source provides a better account of the particular contribution, changes the conditions under which a signal works, or enables the same action with less inference or measurement.