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

A causal audit framework evaluates whether multi-agent LLM systems utilize latent representations for task-relevant information rather than mere message presence. Testing on Qwen3-4B and 8B across GSM8K and MATH-500 benchmarks, the method employs controlled message replacements to decouple encoded content from receiver sensitivity and agent-specific value.