Identify effective training objectives for the pattern-completion network

Identify which prediction-based training objectives over an individual actor’s experience stream most effectively update the pattern‑completion network p to consolidate implicit behavioral patterns in the proposed Linguistic Multi-Actor Environment framework.

Background

The framework allows many possible self-supervised prediction objectives to update the pattern-completion network p based on an individual’s experience. While one example objective is illustrated, the authors do not commit to a specific formulation.

They explicitly defer determining the best objective(s) to empirical investigation, marking an unresolved methodological choice crucial for implementing learning dynamics in the proposed architecture.

References

Here we do not prescribe or advocate for any specific formulation. We illustrate one possibility but leave it to future (empirical) work to discover what might work best.

— A Theory of Appropriateness That Accounts for Norms of Rationality  (2603.14050 - Leibo et al., 14 Mar 2026) in In-context learning versus long-term memory consolidation, Section 4 (Individual decision making by predictive pattern completion)