When Does Latent Communication Pay? A Causal Audit of Relayed KV Caches in Multi-Agent LLMs

TL;DR - An arXiv preprint that causally audits whether multi-agent LLM systems relaying KV caches actually transmit example-specific "latent thoughts," by swapping in deranged (mismatched-example), zeroed, and moment-matched random caches. It matters because benchmark gains alone are shown to be insufficient evidence for latent-thought transfer.

• Two regimes are tested: when the receiver needs the sender's private information, the audit reads ceiling (100% vs 23–25% for answer-irrelevant relays on the primary backbone), replicated across three model families, five checkpoints, and a prose document-QA surface.
• When private information isn't needed, a pre-registered five-seed protocol with Holm-corrected TOST establishes equivalence within 2.8 points on GSM8K and ARC-Challenge across three Qwen3 scales and MedQA at 8B; one cell shows a small detected advantage inside the margin, and a second family shows none.
• Cache magnitude ≠ pairing: in one natural cell, zeroing the relay costs 14.7 points while a mismatched cache costs only 0.4 — so a large cache effect need not be example-specific.
• Delivered channels differ under the same test: LatentMAS's native relay hits ceiling, KVComm's layer subset is partial, and C2C's released projector shows no detected example-specific transfer; the authors release the mismatched-cache audit.