Allocating compute to intrinsic novelty search accelerates skill accumulation by creating foundational primitives that compound into complex capabilities. As explored in Play, prioritizing early exploration drives non-linear growth through structural feedback across the system. Under this setup, High Level Skills rises from 1351 to 3844 at step 15, with these illustrative values demonstrating how expanding the basic toolset multiplies the pace of higher-order mastery over time.
Key takeaways
- Doubling Intrinsic Exploration Compute from 2 to 4 increases High Level Skills to 3844 at step 15 instead of 1351.
- Doubling Composition Efficiency Rate from 1 to 2 raises High Level Skills to 1844, showing that discovery expands downstream output more than assembly speed alone.
- A manager planning agent budgets should prioritize expanding the primitive library because basic behavioral components act as a multiplier on subsequent synthesis.
- High Level Skills synthesis relies on the accumulated contents of Primitive Library, making early novelty search essential for compounding long-term capabilities.
Which compute allocation builds more complex skills?
Anyone who directs training budgets for autonomous agents faces a difficult trade-off between spending compute on pure novelty search or pouring those cycles into synthesizing discovered behaviors. In this model, doubling exploration compute produces a final stock of 3844 High Level Skills at step 15 compared to 1351 in the baseline. Doubling composition efficiency instead raises that stock to 1844. These figures are illustrative model values that map how different allocation choices alter capability over time.
The difference between these two levers emerges because basic building blocks feed future synthesis repeatedly. A system that discovers more behavioral primitives early creates a wider foundation for downstream assembly. Understanding how these pathways feed on each other helps planners decide whether to fund discovery engines or optimize task compilers.
How do exploration and composition levers change the outcome?
The first lever, Intrinsic Exploration Compute, sets how many units of compute go into seeking novelty each step. In the baseline setting of 2, the accumulation of High Level Skills reaches 1351 at step 15. Raising this lever to 4 pushes that outcome to 3844 at step 15, demonstrating how an expanded base accelerates downstream synthesis.
Tracking the steps shows this divergence building across time. At step 3, the baseline sits at 10 while the higher exploration path reaches 20. By step 8, the baseline reaches 147 while the boosted path stands at 369, showing that early additions compound as more building blocks become available.
The second lever, Composition Efficiency Rate, controls the speed multiplier used to assemble existing building blocks into structured capabilities. Leaving exploration compute at its baseline while raising this efficiency lever from 1 to 2 increases High Level Skills to 1844 at step 15.
Looking at intermediate points reveals an earlier pop that later levels off. At step 1, the baseline shows 1 skill while the boosted efficiency setting shows 2. By step 5, the efficiency path reaches 60 compared to 38 in the baseline, but it lacks the larger reservoir of primitives needed to match the 3844 reached under higher exploration.
How does the skill acquisition model work?
The playable model connects discovery directly to downstream capability through a set of reinforcing components. Novelty Search acts as a discovery pump, turning allocated Intrinsic Exploration Compute into items stored within Primitive Library. This pool begins with 2 primitives and collects candidate motor patterns created across training epochs.
Skill Synthesis then draws upon the contents of Primitive Library to construct complex behaviors. Synthesis Flow Rate calculates the creation of new abilities using the size of the library, the setting of Composition Efficiency Rate, and feedback from already acquired capabilities. The assembled behaviors accumulate in High Level Skills, which starts at 0 and grows through this ongoing assembly.
What does this result mean for planning compute budgets?
For anyone balancing compute allocations in autonomous agent design, this model suggests that expanding the primitive discovery budget matters more than solely refining assembly efficiency. When the raw repertoire of behaviors remains narrow, a higher composition multiplier quickly runs into diminishing additions. Investing in the underlying library provides the raw material that compounding synthesis requires.
In this model, the practical rule is to examine the input pool before attempting to optimize assembly pipelines. Improving synthesis efficiency produces faster gains in the opening epochs, but higher exploration delivers far greater total capability by step 15. Planners should watch whether their agents face a shortage of basic actions before committing resources to downstream compilation.
What does the model leave out?
This playable model adopts several deliberate simplifications. Each step of Play is treated as a uniform agent training epoch, holding timing and updates to a fixed grid. The baseline values and lever choices are constant settings rather than dynamic distributions that adjust based on internal performance.
The structure also assumes that compute allocated to novelty search directly adds usable items to Primitive Library without waste or duplication. Real systems are more complicated than any model, but these focused components make it easy to isolate how discovery inputs interact with hierarchical synthesis.
Where does the topic come from?
This model was inspired by the publication titled Intrinsic Goals for Autonomous Agents: Hierarchical Exploration Captures Future Volitional Action in Drosophila, published by openRxiv on 27 October 2026. The scenario presented here, along with its specific values and steps, represents the model's own illustrative scenario exploring hierarchical capability growth.
Can you test different compute trade-offs yourself?
You can explore these trade-offs by shifting the levers directly in the playable model. Try setting Intrinsic Exploration Compute to 3 while dialing Composition Efficiency Rate down to 0, or test balanced intermediate combinations. Observing how the accumulation curve bends across epochs will clarify how your budget choices shape agent mastery.
The parts of the model
The question it asks: How does allocating compute between intrinsic novelty search and task execution govern the accumulation of reusable agent skills?
- Novelty Search · Source
- Produces candidate motor primitives each epoch via intrinsic novelty exploration.
- Primitive Library · Pool
- Stores modular behavioral primitives available for assembly into complex tasks.
- Skill Synthesis · Source
- Constructs complex behaviors by hierarchically assembling available primitives.
- High Level Skills · Pool
- Measures the total repertoire of mastered high-level skills.
- Intrinsic Exploration Compute · Lever
- Adjusts compute units allocated per step to exploratory novelty search.
- Composition Efficiency Rate · Lever
- Adjusts the efficiency multiplier for combining primitives into functional skills.
- Synthesis Flow Rate · Register
- Computes new skills synthesized per step based on primitives, efficiency, and feedback.
Assumptions and limits
- Each step of Play represents one agent training epoch (illustrative).
- Compute allocated to intrinsic discovery directly populates the Primitive Library.
- Synthesis of High Level Skills draws upon the size of Primitive Library via a reinforcing multiplier.
A teaching model with illustrative values; each Play is a scenario, not a forecast.
Sources
- openRxiv (via Crossref) (doi.org)
About this model
Machinations makes these models with AI from a public source, and each one is a teaching scenario: its numbers are illustrative values chosen for the model, not measurements or forecasts. You can open this diagram in the Community to remix it and change it.