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

2026-08-05

AI Summary

This study addresses the lack of causal evidence regarding whether “latent communication” via key-value (KV) cache transfer in multi-agent large language models genuinely depends on alignment between sender-private information and specific task examples. The authors propose the first causal auditing framework, combining mismatched, zeroed, and moment-matched randomized caches with TOST equivalence testing, Holm-corrected multiple comparisons, and pre-registered multi-seed experiments across diverse models and benchmarks. Results show that in tasks requiring private information, correctly aligned caches achieve 100% accuracy—significantly outperforming irrelevant caches (23–25%)—whereas in tasks not requiring private information, most systems exhibit no example-specific gains, and some channels demonstrate no effective communication at all. The work clarifies that cache effects do not equate to example-pairing effects and releases a reproducible auditing protocol.

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 \emph{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.

Problem

• latent communication
• multi-agent LLMs
• KV cache
• causal audit
• example-specific transfer

Innovation

• causal audit
• latent communication
• KV cache relaying
• multi-agent LLMs
• mismatched-cache ablation