Do Latent Channels Actually Communicate? A Causal Audit of Latent Multi-Agent LLM

This paper introduces a causal audit for latent communication in LLM-based multi-agent systems. Instead of relying on end-task accuracy, it replaces sender-produced representations at the receiver boundary to separate message presence, message identity, example-specific content, and the incremental value of another agent. Experiments with Qwen3-4B and Qwen3-8B on GSM8K, ARC-C, and MATH-500 show that aggregate gains can conceal opposing causal components. For example, the Qwen3-4B GSM8K effect is -1.00 percentage point overall, combining a -6.17-point other-example component with a +5.17-point example-specific component.