Conjecture—Systematic generalization via rule learning as an inductive bias

Establish whether encoding systematic generalization through learned compositional rules is an effective inductive bias for embodied agents to achieve robustness to variations in tasks and environments.

Background

Embodied agents must generalize beyond limited, non-IID training data in changing environments. Systematic generalization—learning rules for composing concepts—may provide a powerful inductive bias.

The authors conjecture that encouraging learning of general rules can enable robust performance across task and environment variations.

References

We conjecture that an inductive bias that encodes systematic generalization by encouraging a set of general rules to be learned may be the best way to enable agents that are robust to variation in task and environment.

— From Machine Learning to Robotics: Challenges and Opportunities for Embodied Intelligence  (2110.15245 - Roy et al., 2021) in Section 2.1 (Inductive Biases for Embodied Intelligence: Limitations and Challenges)

Our transfer evaluation shows that changing surface details, such as layouts or characters, can reduce agents’ learning slopes even when the underlying rules remain the same (Table~\ref{tab:transfer}). In real-world tasks, agents also regularly encounter changes in their surroundings, available resources, and the people they interact with. This raises a broader question: how can agents adapt to these changes in their environments? Future methods could help agents compare experiences across different situations, distinguish shared rules from instance-specific details, and check when those rules still apply.

— Learn2Play Bench: How Well Do LLM Agents Learn from Experience in Unfamiliar Environments?  (2610.08215 - Li et al., 6 Oct 2026) in Section Discussions and Future Directions, subsection “How can agents learn what stays constant as environments change?”