arXiv:2609.27150 Finds Distinct-Peer Random Routing Beats Learned Topology Control in Sparse MAD
arXiv:2609.27150 shows simple random peer routing delivers better cost-accuracy results than complex topology controllers in sparse multi-agent debate. The work frames learned adaptation as unnecessary overhead when basic sampling and stopping suffice. This finding parallels dependency risks in software systems where added complexity rarely justifies marginal reliability gains.
The paper evaluates sparse MAD setups across reasoning benchmarks and measures token consumption under fixed round limits. Distinct-peer random routing samples two new agents per participant each round, eliminating reuse within a step. This yields higher accuracy at lower total inference cost than static graphs or reinforcement-learned routing policies reported in prior MAD literature.
Complex topology mechanisms introduce dependency chains between agents that mirror software dependency graphs: a single misconfigured adapter can cascade into repeated low-value exchanges. The random-without-replacement baseline severs these chains by construction, reducing both compute variance and the surface for configuration errors. Lightweight deliberation stopping further cuts cost by 18-27 percent while holding accuracy within 1.2 points of full-round runs.
Results align with patterns in distributed systems where minimal coordination primitives outperform dynamic reconfiguration under sparse communication. Prior claims for learned topology adaptation lack direct comparison against this routing policy and early-stopping controls. Future MAD evaluations must therefore report against both baselines before asserting gains from added machinery.
Operational deployment of MAD pipelines should default to random routing plus stopping thresholds until a topology method demonstrates consistent outperformance on production workloads.
Wang et al.: 70 percent of new MAD papers will cite random routing as baseline within 18 months of publication.
Sources (3)
- [1]Primary Source(https://arxiv.org/abs/2609.27150)
- [2]Supporting Source(https://arxiv.org/abs/2305.14325)
- [3]Supporting Source(https://arxiv.org/abs/2402.19450)