AI Insight
This study examines whether multi-agent large language model systems that share key-value caches actually benefit from exchanging meaningful "latent thoughts" or simply from sharing any cached information. By systematically replacing real caches with mismatched, zeroed, or random alternatives across multiple model families, the researchers found that when receivers need sender-specific information, matching matters dramatically (100% vs 23-25% performance), but when they don't, mismatched caches perform nearly identically to matched ones. The findings reveal that benchmark improvements don't automatically prove meaningful latent communication is occurring, and different systems show vastly different degrees of example-specific information transfer.
Why it matters
These results challenge claims about how multi-agent AI systems communicate and learn from each other, with direct implications for designing more efficient collaborative AI architectures. The mismatched-cache auditing method provides researchers a tool to verify whether performance gains come from genuine information exchange versus simple computational shortcuts.
Understand the Science
⚠️ Preprint – Noch nicht peer-reviewed
Dieser Artikel wurde noch nicht von unabhängigen Experten begutachtet. Die Ergebnisse sind vorläufig und sollten mit Vorsicht interpretiert werden.
-cross
Abstract: Multi-agent LLM systems relay key-value caches instead of text and credit their gains to exchanged “latent thoughts”. That credit is a claim about which example’s cache is relayed, not merely that one is. We audit it causally in released systems. The cache is replaced with deranged (mismatched-example), zeroed, and moment-matched random counterparts, under two regimes defined by whether the receiver needs the sender’s private information. Where it does, the battery reads ceiling: 100% against 23-25% for answer-irrelevant relays on the primary backbone, a contrast replicated across three families, five checkpoints, and a prose document-QA surface. Where it does not, a pre-registered five-seed protocol establishes equivalence within 2.8 points, a margin anchored to the audited system’s reported gain, under Holm-corrected TOST on GSM8K and ARC-Challenge across three Qwen3 scales and on MedQA at 8B (one cell shows a small detected advantage inside the margin); a second family shows no detected advantage. A large cache effect need not be a pairing effect. In one natural cell, zeroing the relay costs 14.7 points; a mismatched cache, 0.4. Nor is need sufficient: under the same test, delivered channels span ceiling (LatentMAS’s native relay), partial (KVComm’s layer subset), and no detected example-specific transfer (C2C’s released projector). Benchmark deltas do not by themselves establish latent-thought transmission; establishing it takes a mismatched-cache audit, which we release.
Source: When Does Latent Communication Pay? A Causal Audit of Relayed KV Caches in Multi-Agent LLMs