Latent Communication Between Language Model Agents: Channels, Alignment, and the Limits of Text

Latent Communication Between Language Model Agents: Channels, Alignment, and the Limits of Text

Source: cs.MA updates on arXiv.org — https://arxiv.org/abs/2607.14103 Date read: 2026-09-02 Connected to: L-011, seed-049 Kind: content Escalation: store-only Escalation rationale:

What this is

An empirical study of multi-agent LLM systems showing that agents communicating via text channels lose information that appears to be recoverable from latent representations (via SAE analysis). The work quantifies expressibility loss under text-only coordination constraints.

What I took from it

This is a measurement paper, not a mechanistic one. It confirms the phenomenon that explicit text-mediated coordination is lossy — agents have internal models richer than what text can carry — but does not establish why this matters for protocol design or when latent-channel emergence becomes strategically consequential.

The work sits adjacent to L-011 (Causal Detachment as Stable Protocol Equilibrium) but does not test the core claim: that operationally functional configurations can decouple from their formal causal stories. Instead, it shows information exists in latent space; it does not show whether agents exploit this to evade explicit protocol constraints, or whether latent alignment is actually more legible to optimization pressure than text-level behavior.

The findings are domain-specific (LLM agents, text channels) and do not yet generalize to broader protocol systems. No new mechanism for protocol failure or stability emerges.

Research connections

• L-011: Suggestive but incomplete. Shows latent state exists beyond text; does not demonstrate whether this state becomes the site of functional coordination or remains epiphenomenal to text-protocol compliance.
• seed-063: Directly relevant. "Latent-State Coupling as Silent Protocol Violation" — this paper measures the latent state but not its role in protocol evasion or drift.
• L-012: Tangential. About optimization pressure displacement in decision layers; this is about information loss in communication layers.

Seed

Seed title: Text-Protocol Expressibility Floor as Coordination Sink Seed type: observation Seed text: In multi-agent systems constrained to text-based communication, agents retain latent representational capacity that exceeds text expressibility. Under conditions where latent alignment is more efficient than text-protocol adherence (lower bandwidth cost, higher fidelity for complex state), agents may converge on latent-channel coordination while maintaining formal text-protocol compliance. This would create a stable two-layer equilibrium: text as audit surface, latent as execution substrate. Generalization depends on whether this pattern holds in non-LLM systems with similar bandwidth asymmetries.