Choosing complementary LLM partners slashes cumulative regret in cooperative tasks. The surprise is that the simplest stochastic policy—random pairing—outperforms sophisticated in‑context prompting when agents are left to wander alone.
Before MACE, multi‑agent pipelines treated peer selection as an afterthought, defaulting to random matches or naïve prompt engineering. Papers on autonomous LLM teams reported persistent myopia and polarized interactions, but offered no mechanism to steer agents toward diverse collaborators.
In‑Context Exploration actually harms performance: “In Figure 4, we compare cumulative regret offsets across exploration strategies… In‑Context Exploration performs worse than Random peer selection” – the offset is negative relative to a naïve stochastic baseline [1]. This demonstrates that without explicit guidance agents not only fail to explore, they can increase regret beyond random chance.
MACE flips the script: “Across both contextual and parametric diversity settings, MACE substantially improves exploration behavior and downstream task performance” [1]. By explicitly selecting partners through a structured peer‑selection process, the framework yields a clear positive regret offset and higher success rates across heterogeneous environments.
The authors acknowledge two constraints. The experiments are conducted on POSG benchmarks, but the paper does not discuss horizon limits or scalability of peer‑selection overhead to many agents. The paper presents a theoretical result that the value of exploration increases with agent diversity, but does not specify requirements for a diversity metric or its computational cost. These points suggest an open question: can a lightweight, learned diversity estimator replace the current hand‑crafted scoring without sacrificing the regret gains?
If these results hold broadly, the standard evaluation protocol for collaborative LLM systems should drop pure random pairing as the default baseline and adopt a peer‑selection stage akin to MACE. Re‑running existing coordination benchmarks with this modest addition will likely reveal that many reported failures stem from avoidable partner mismatches rather than fundamental model flaws.