Can language help agents imagine goals they've never seen?

Goal generation is a bottleneck for open-ended learning agents. Fixed hand-defined goal spaces limit agents to predefined objectives. Learned generative models of states constrain goals to the distribution of known effects. Neither enables creative discovery — imagining outcomes the agent has never experienced.

IMAGINE models how children leverage language compositionality to overcome this: language descriptions can specify goals the agent has never seen because language is compositional — familiar words recombine to describe unfamiliar outcomes. An agent that has seen "red ball" and "blue box" can imagine "red box" without having encountered it.

Three mechanisms enable this:

1. Language as goal representation — descriptions of outcomes serve as target goals during play, enabling targeting of out-of-distribution states
2. Social peer guidance — a social partner provides language descriptions, bootstrapping the imagination process (paralleling how children's exploration is scaffolded by adult language)
3. Modularity — decomposition between learned goal-achievement reward function and policy (using deep sets, gated attention, and object-centered representations) enables generalization from imagined goals to actual exploration behavior

This challenges purely intrinsic motivation approaches to exploration. Since Can communication pressure drive agents to learn shared abstractions?, language is not just a communication tool but a cognitive scaffold: it structures the agent's goal space, enabling combinatorial expansion of what can be targeted.

The connection to Can LLMs reason creatively beyond conventional problem-solving? is direct: IMAGINE implements combinational creativity (recombining familiar concepts into novel goals) as a concrete mechanism. Exploratory and transformative creativity would require additional mechanisms, but the combinational foundation enables a rapid expansion of the search space.

The social guidance requirement echoes Can human-AI research teams improve faster than autonomous AI systems? — the most effective exploration is not fully autonomous but socially scaffolded.