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

Abstract

Multi-agent latent reasoning composes the KV-cache contributions of several agents into one context for a final agent. Prior work (Agent Primitives) does this by concatenating per-agent caches with RoPE re-encoding, a construction we name BagMerge. BagMerge is non-commutative, and the best input ordering is not predictable a priori: it shifts with deployment regime, latent-step budget, model scale, and model family. We make this cache exchange a convergent replicated state. CanonicalMerge fixes the layout by a content-determined ordering (mean K-norm at a middle layer), making the merged cache byte-identical under any input permutation, verified on synthetic tensors (N <= 5) and bit-for-bit on real Qwen3-1.7B and Qwen3-4B KV state. We then separate state from layout: the durable object is a set of content-addressed latent fragments merged by set union, a state-based CvRDT, and CanonicalMerge is its deterministic render, so every accuracy number is inherited and re-delivered duplicates are absorbed. On a partitioned-reasoning benchmark CanonicalMerge matches the best BagMerge ordering without knowing which it is (within 4 points in all 12 cells at 1.7B; same picture at 4B), and on Llama-3.1-8B, where the ordering gap grows to 19 points, it again tracks the best ordering. On HotpotQA (n = 200) and MuSiQue it is the best cache-level method, at a 7-point F1 cost against shipping the text on HotpotQA and at parity on MuSiQue, while the output-fusion baseline PackLLM trails by 45 points. At k > 2 we delimit the approach: cache merge transports latent traces but does not by itself compose them.