Committee of Weak AI Models Can Match Stronger Ones

A new arXiv paper (2605.14163) proposes that a committee of weak reasoning language models can achieve performance comparable to much stronger models through verifier-backed committee search. The mechanism involves repeated sampling to amplify coverage, but critics and comparators require a local soundness signal like execution or proof checking. Rank-based bounds show when local selection errors compose into reliable trajectories, and the proposer-side ceiling is characterized by oracle best-of-k converging only to the mass of task solutions.