Cache Merging as a Convergent Replicated State for Multi-Agent Latent Reasoning

- This paper introduces "cache merging," a technique that merges the internal memory (KV cache) of multiple large language model inference runs to enable multi-agent reasoning without a central orchestrator. - KV cache is the hidden-state data an LLM stores during text generation; normally each agent's cache is isolated. By merging caches, agents can share unfinished thoughts and converge on a shared latent state. - The method targets "multi-agent latent reasoning" — letting several LLM instances work on the same problem by exchanging intermediate representations rather than natural language. - A key motivation: it avoids the overhead of a centralized controller or a separate aggregation model, since the merge itself acts as the coordination primitive.