AI 評測與代理系統
KV Cache Is Not “Latent Thought”: Mismatched Caches Reveal That Multi-Agent Gains May Be Merely Interface Effects
A new study re-audits latent communication in multi-agent systems using mismatched, zeroed, and random KV caches, finding that the large performance drop caused by removing the cache does not necessarily prove that the receiver used specific information from the sender. On some standard benchmarks, zeroing the cache reduced accuracy by 14.7 points, while replacing it with a cache from another problem cost only 0.4 points.
Multi-agent systems have begun directly transmitting Transformer KV caches to reduce the tokens and latency introduced by text-based mediation. However, [new research](https://arxiv.org/abs/2608.04893) argues that “a cache performs better than no cache” is insufficient evidence that agents exchanged useful latent thoughts. The researchers replaced authentic caches with mismatched caches from other problems, all-zero caches, and random caches with matched statistics, allowing them to separately measure the effects of interface presence, generic computational state, and problem-specific content.
When the receiving model genuinely lacked answer-relevant information held by the sender, LatentMAS’s native channel achieved 100% accuracy in the primary Qwen3-8B setup, while caches containing no private information achieved only 23.4% to 25.2%. Similar gaps appeared across three model families, five checkpoints, and text-based document question answering. But in GSM8K, ARC-Challenge, and MedQA settings where the receiver could solve the problem independently, the difference between correctly matched and mismatched caches fell within the prespecified ±2.8-point equivalence margin in all seven Qwen3 experimental cells. MedQA was especially revealing: zeroing the cache caused a 14.68-point loss, whereas a mismatched cache caused a loss of only 0.40 points, suggesting that most of the effect may come from a well-formed computational state rather than content specific to the problem.
The same audit also distinguished among implementations. [LatentMAS](https://github.com/Gen-Verse/LatentMAS), which transfers the full cache, demonstrated problem-specific effects close to the upper bound. KVComm’s partial-layer transfer produced gains of only about 7.2 to 10.8 points, while C2C’s publicly available projector showed no detectable, consistent transfer of specific content. From an engineering perspective, latent-communication evaluations should include mismatched samples instead of relying solely on cache-versus-no-cache ablations. One limitation is that each external system was tested using only a single publicly available configuration, and some Mistral ARC results were affected by parsing failures. The conclusions could differ after retraining the projector or changing the number of cache layers transferred.