Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

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

Source

arXiv

arxiv.org/abs/2608.04893v1

Read the source

Source

Plain-Language Brief

The study investigates how effective relaying key-value caches between multi-agent language models is, particularly when the receiver does not need the sender's private information. The research shows that when the receiver needs the sender's information, there is a significant improvement in performance, with up to 100% gain compared to answer-irrelevant relays on the primary backbone. However, when the receiver does not require the sender's private information, the difference is smaller and negligible. The study also finds that a large cache effect is not necessarily due to pairing effects, as zeroing the relay can cost 14.7 points in one scenario, while a mismatched cache only costs 0.4 points.

What the Document Says

  • The study looks into the effectiveness of multi-agent language models (LLMs) that exchange key-value caches instead of text and attributes their performance to 'latent thoughts'.
  • The research audits this claim causally in released systems by replacing the cache with deranged, zeroed, and moment-matched random counterparts.
  • When the receiver needs the sender's private information, there is a significant improvement in performance, with up to 100% gain compared to answer-irrelevant relays on the primary backbone.
  • When the receiver does not require the sender's private information, the difference in performance is smaller and negligible.
  • The study finds that a large cache effect is not necessarily due to pairing effects.
  • In one scenario, zeroing the relay costs 14.7 points, while a mismatched cache only costs 0.4 points.

Key Concepts Mentioned

  • multi-agent language models (LLMs)
  • key-value caches
  • latent thoughts
  • causal audit
  • private information
  • performance gain

Claims Made by the Document

  • Multi-agent LLM systems show significant performance improvement when exchanging key-value caches.
  • The improvement is only significant when the receiver needs the sender's private information.

Evidence Used

  • 100% gain against 23--25% for answer-irrelevant relays on the primary backbone when the receiver needs the sender's private information.
  • The margin between the performance of the delivered channels is anchored to the audited system's reported gain.

Methods / Approach

  • causal audit
  • benchmark deltas

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

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.