Mixture of Complementary Agents for Robust LLM Ensemble
The paper argues that picking LLMs for an ensemble should favor complementarity, not just strength or diversity, and proposes greedy selection methods.
Intelligence analysis by GPT-5.4 Mini
The paper reframes proposer selection in multi-LLM pipelines as a combinatorial problem: the best mix is not necessarily the strongest or most diverse, but the set that complements one another and the summarizer. It tests greedy methods that approximate that idea with a small labeled set.
A team of robot helpers is trying to answer a question. One robot writes one idea, another writes a different idea, and a bigger robot combines them into one final answer.
The paper says the best helpers are not always the smartest ones by themselves. Sometimes the best team is like a soccer squad: players matter most when they work well together, not just when they are stars on their own.
So the paper tries to pick helpers that fit together nicely. It tests simple choosing tricks that are fast enough to use in real life, and says this teamwork idea helps the final answer work better.
Analysis
What the paper tackles
Multi-AI collaboration, including ensembling and debating large language models, depends on a first step that is easy to overlook: deciding which proposer models should feed a summarizer. The paper says that choice is not trivial. If the system only picks the strongest models, it may miss useful interactions. If it only chases variety, it may still ignore how the proposers work together.
Core idea
The authors recast proposer selection as a combinatorial problem similar to feature selection. In that framing, the useful property of an LLM is not only its standalone accuracy, but how much it adds when paired with the others and with the summarizer. That is the paper’s idea of complementarity. Rather than applying standard feature-selection methods directly, the paper argues that those methods are too expensive for this setting.
What they do instead
To make the idea usable, the paper explores a broad set of greedy-style algorithms that can estimate complementarity from a small labeled set. These methods are designed to be computationally feasible, so they can work in real LLM ensemble pipelines without the heavy cost of exhaustive search.
Main takeaway
According to the abstract, the experiments support complementarity as a useful principle for choosing proposers. The paper also identifies methods that offer the best trade-off between performance and cost in practice. The result is a selection strategy aimed at making LLM ensembles more robust without turning the setup into an impractical optimization problem.
Key points
- The paper studies how to choose proposer LLMs for ensembling and debating setups.
- It argues that complementarity matters more than accuracy or diversity alone.
- It treats proposer selection like a feature-selection problem.
- It uses greedy-style methods that can work with a small labeled set.
- The experiments reportedly find good performance-cost trade-offs.
