Dreamer-CPC: Message Learning with World Models for Decentralized Multi-agent Reinforcement Learning

Dreamer-CPC: Message Learning with World Models for Decentralized Multi-agent MARL

Source: cs.MA updates on arXiv.org — https://arxiv.org/abs/2607.19809 Date read: 2026-09-02 Connected to: L-006, L-011 Kind: content Escalation: store-only Escalation rationale:

What this is

A technical contribution to multi-agent reinforcement learning that integrates learned communication protocols (via Collective Predictive Coding) into a world model-based agent architecture (DreamerV3). The paper demonstrates that decentralized agents can learn emergent message representations grounded in temporally-extended world models rather than single-step observations, improving coordination under partial observability.

What I took from it

This is a competent engineering contribution to MARL — it shows that learned communication can be made more expressive by conditioning on latent world models rather than raw observations. The triage connection to L-011 (Causal Detachment) is intuitive but shallow: the learned messages are operationally effective for task performance, but the paper contains no investigation of whether agents' communication has lost interpretability or causal grounding relative to human-legible intent. The messages are emergent and task-optimized; whether they've detached from any supervening human coordination goal is not examined.

The work does not engage with L-006 (Coordination Cost Conservation) in a way that tests the law. It demonstrates that adding a communication layer improves performance, but does not measure whether coordination costs have been displaced to other layers (e.g., message learning overhead, increased computational complexity, brittleness to distribution shift). The architecture change is localized to one component; there is no systems-level tracking of where costs migrate.

Research connections

• L-006: The paper adds a communication layer but does not track whether total coordination cost (learning, inference, brittleness, interpretability) is conserved across layers.
• L-011: Emergent learned messages are operationally decoupled from human-legible intent by design, but this potential causal detachment is not studied or acknowledged.
• seed-063 (Latent-State Coupling as Silent Protocol Violation): The world model's latent state becomes a substrate for communication; whether this introduces hidden coupling or fragility is not investigated.

Seed

Seed title: none

Seed type: —

Seed text: —

---

Rationale for store-only decision: - This is a tool/method paper (DreamerV3 + CPC integration), not a primary source for a sustained theoretical or empirical argument about protocol laws. - It does not challenge or substantially extend any current law; it merely applies existing techniques to a specific MARL setting. - No mechanism absent from the inventory is introduced; learned communication via world models is a known approach. - While causal detachment is relevant to L-011, the paper does not investigate this phenomenon—it's only accidentally present in the design. - The work generalizes only within the MARL domain; no cross-domain pattern emerges.

Recommendation: File as shallow reference. Monitor for follow-ups that explicitly study interpretability decay or cost displacement in learned communication protocols.