Determine whether modular adaptation reuses prior experts

Determine whether adaptation by the pretrained, frozen-encoder PWM mixture-of-experts world model is based on reusing previously learned experts or on learning newly introduced experts for the composition task.

Background

The continual PWM variant introduces new dynamics experts for each incoming task while freezing experts learned for earlier tasks. Its router can combine all active experts, making it possible for composition-task performance to arise either from functional reuse of prior experts or from learning the newly added experts.

The authors directly identify this attribution as unresolved and investigate it with routing visualisations and expert-ablation experiments. The question is included because the paper does not assume that routing weights alone establish functional reuse; it tests whether prior experts materially support composition-task performance.

References

However, it remains less clear whether adaptation is based on reusing previous experts or on learning the newly introduced ones. Our benchmark's explicit compositional structure allows us to examine this distinction directly.

— Benchmarking World Models for Continual Learning on Compositional Tasks  (2609.22055 - Zhou et al., 18 Sep 2026) in Section 4.1, paragraph “Router weights recover task-level reuse structure in several curricula” / Section 4.1 opening

Future work could test whether task embeddings can guide layer-specific reuse or whether similarity-based selection could be combined with learned composition, including an adaptive balance between prior and current knowledge.

— How to Find and Reuse Policies for Continuous Adaptation in Lifelong Reinforcement Learning  (2610.03119 - Nath et al., 2 Oct 2026) in Discussion and Future Work, Section 6