MOC Fixes the Telephone Game Killing Multi-Agent AI Accuracy
MOC rewires how LLM agents share information across multi-hop paths, cutting token bloat while boosting accuracy up to 6.77% on reasoning benchmarks
Takeaways
- MOC is a new communication scheme for LLM multi-agent systems that exposes raw upstream responses from multiple hop distances, not just direct neighbors. [Paper]
- Current MAS suffer from semantic attenuation: key information gets distorted as it passes through intermediate agents who paraphrase upstream outputs.
- A Semantic-Topological Merging algorithm compresses redundant multi-hop messages using lightweight embeddings and a small distillation model, keeping token costs in check.
- Results show up to +6.77% accuracy on AQuA and +3.84% on MMLU; at 20 agents, MOC uses fewer tokens than a vanilla MAS baseline.
- K=2 hop depth is the optimal default; K=3 introduces redundancy in denser graphs and can hurt performance.
- MOC is complementary to topology optimization: plugging it into G-Designer adds +4.8% on HumanEval on top of an already-optimized graph. [Code]
Multi-agent systems built on LLMs have become one of the most active research fronts in AI, with frameworks like AutoGen, MetaGPT, and GPTSwarm showing that a team of specialized agents can outperform any single model. But almost all of the innovation has focused on one question: how should agents be connected? A new paper argues that the far more neglected question is what happens on those connections, and proposes a scheme called Multi-Order Communication (MOC) to fix it.
The telephone game problem in agent networks
In a typical LLM-based multi-agent system (MAS), agents are arranged as nodes in a directed graph. When agent C needs to reason, it receives the outputs of its direct upstream neighbors, B and A, concatenated into a prompt. That sounds reasonable until you realize what it means for deeper graphs: agent C only ever sees what B chose to say about A's output. Any nuance, correction, or key detail that A produced but B paraphrased away is simply gone.
The paper formalizes this as a restricted evidence receptive field: each agent is blind to everything more than one hop upstream. The authors draw an analogy to the telephone game, where messages get progressively distorted as they travel through intermediaries. They call this semantic attenuation, the gradual loss of meaning as information passes through multiple agents who summarize, reinterpret, or simply miss details.

The naive fix, just give every agent access to all upstream outputs, runs into a second wall: context inefficiency. LLMs process ordered text, not fixed-size vectors. Expanding the evidence scope causes prompt length to grow linearly with the number of agents, and research has shown that LLMs struggle to use information buried in long contexts effectively.