Causal Audit Finds Some Multi-Agent 'Latent Thought' KV Relays Transfer Nothing Example-Specific: Zeroing Costs 14.7 Points, Mismatching Costs 0.4

arXiv 2608.04893·high signalXBlueskyLinkedInCopy link

Multi-agent LLM systems that relay key-value caches instead of text credit their gains to exchanged latent thoughts, but that is a claim about which example's cache is relayed. Replacing the cache with deranged, zeroed, and moment-matched random counterparts shows the claim holds only when the receiver actually needs the sender's private information — there the contrast is 100% versus 23-25%, replicated across three families, five checkpoints, and a prose document-QA surface. When it does not, a pre-registered five-seed protocol establishes equivalence within 2.8 points under Holm-corrected TOST on GSM8K and ARC-Challenge across three Qwen3 scales; in one natural cell zeroing the relay costs 14.7 points while a mismatched cache costs 0.4, and among released systems LatentMAS's relay hits ceiling, KVComm's layer subset is partial, and C2C's projector shows no detected example-specific transfer.