Cache Merging as a Convergent Replicated State for Multi-Agent Latent Reasoning
7 hours ago
- #cache merging
- #latent reasoning
- #multi-agent systems
- Prior work (Agent Primitives) concatenates KV-caches for multi-agent latent reasoning with BagMerge, which is non-commutative and depends on unpredictable input ordering.
- CanonicalMerge makes the exchange a convergent replicated state by ordering caches by mean K-norm at a middle layer, ensuring byte-identical merged cache under any input permutation.
- The replicated state is separated from decode-time layout, using content-addressed latent fragments with set union as a state-based CvRDT, where CanonicalMerge is its deterministic render.
- CanonicalMerge matches the best BagMerge ordering across regimes and budgets without prior knowledge, trading minimal accuracy loss for structural guarantees.
- On multi-document QA (HotpotQA), CanonicalMerge outperforms output-fusion baseline PackLLM by 45 points at matched budget, showing cache-level merging as distinct from output-level fusion.