Priors That Travel: Proving Learned Content Transfers Across Worlds
Keywords: Cross-World Transfer, Generalization, Inductive Bias, Active Inference, World Model, Artificial General Intelligence, AGI, Cognitive Science, Machine Learning, Project Halo
If you teach an agent to navigate a specific virtual maze, and then change the colors of the walls, the positions of the corridors, and the coordinates of the goal, the agent will typically fail. It has overfitted to the specific sensory footprint of its training environment. It has learned the "coordinates" of the solution, not the "rules" of the world.
For a mind to scale, it must be capable of transfer. What it learns in World A must measurably accelerate its learning in World B, even when World B shares only deep structural laws while differing in surface details.
In Volume 3, Episode 7, we set out to prove this keystone capability. We built a new world family: ChemWorld.
In ChemWorld, raw resources are inert and useless on their own. Survival requires discovering a causal recipe: combining Resource X and Resource Y to synthesize fuel, which yields massive energy. The specific identities, positions, and gradients of the resources are randomized in every new run.
We wanted to know: Can an agent learn this abstract causal recipe in World A, carry that prior across, and use it to instantly survive in an unseen World B?
The Minimal Cross-World Engine
We built a minimal cross-world agent (ChemAgent). The agent's physical navigation was handled by body-affordances, but its mind learned one specific, abstract causal belief: combine_value—"does combining two complementary resources pay?"
This belief was designed to be identity-invariant. It wasn't about "Resource X plus Resource Y"; it was about the abstract act of combining resources.
We set up our experimental loop:
1. Training (World A): The agent lives for 40 lives in a world where combining resources is necessary to survive. It learns that combine_value is highly positive.
2. Transfer (World B): We freeze the agent's learned belief and drop it into a completely unseen random layout with different resource positions, concentrations, and gradients.
3. Control (World A'): We run a control agent in a world where combining resources yields zero energy, then transfer it.
4. Scratch: We run a from-scratch agent with no prior beliefs.
The Gated Results: Proving Transfer Works
We measured the first-contact survival lifespan (capped at 600 steps) in the unseen World B across 3 independent seeds:
| Agent Configuration | Life-1 Survival in World B | First-5 Lives Average | First-20 Lives Average |
|---|---|---|---|
| From-Scratch | $324$ steps | $261$ steps | $408$ steps |
| Transferred Agent (A $\rightarrow$ B) | $600$ steps (Cap) | $473$ steps | $522$ steps |
| Wrong-Prior Control (A' $\rightarrow$ B) | $184$ steps | $264$ steps | $409$ steps |
The data provided a clean validation of transfer: * Instant Competence: The transferred agent survived to the maximum cap ($600$ steps) on its very first life in an unseen world. The from-scratch agent struggled, dying at $324$ steps. * The Cost of a Wrong Prior: The control agent carrying the wrong belief (that combining is useless) performed the worst, surviving only $184$ steps. This proved the gain was specifically due to the transferred causal belief, not generic pretraining or feature design. * Durability: The transferred agent maintained its survival advantage across its first 20 lives, proving the prior was durable.
Decomposing the Transfer
We wanted to ensure this wasn't a black-box trick. We designed a wrong-prior control to settle a crucial question: was the transfer due to pre-built feature invariants (a design gimmick) or the actual learned content?
We ran the sweep on a scaled task (--gate12), comparing a from-scratch agent, a correct-prior agent, and a wrong-prior agent (which carried the wrong belief about the combine value but correct beliefs about navigation, acid-avoidance, and foraging).
Total Transfer Gain = Correct Prior - From-Scratch = +359 steps
Shared Structure (Wrong Prior - Scratch) = +195 steps (Foraging, navigation, acid-fleeing)
Specific Learned Fact (Correct - Wrong) = +164 steps (Value of the combine recipe)
The transfer decomposed additively. * The shared structural laws of the physics (how to navigate, how to avoid hazards, how to forage) carried $+195$ steps of survival value unconditionally. * The specific learned causal recipe ("combining pays") added another $+164$ steps on top. * The wrong-prior agent paid a penalty of exactly the value of the fact it got wrong.
This was not a black-box shortcut. It was an auditable, mathematical account of what knowledge carried and what it was worth.
The Next Wall: The Compounding Challenge
Cross-world transfer was now real. But we quickly hit the next cognitive wall: compounding.
When we presented the agent with a curriculum of levels where the active recipe changed per level, a single fixed policy could not survive. The agent could not determine which recipe was active in its current level.
To solve this context-dependent transfer, the agent needed to do more than hold a point-estimate belief. It needed to reason under uncertainty.
In the next chapter, we look at Episode 8: how version-space Bayes-adaptive planning solved the compounding wall and let the agent probe-then-exploit optimally.