The proposal suggests a hierarchical decoding architecture for language models, where smaller models handle most token generation, while larger models intervene only when necessary. This approach aims to reduce latency, energy consumption, and costs associated with using large models for every token, by having them act as supervisors that monitor for errors or critical reasoning steps. The system could involve a Mixture-of-Experts (MoE) architecture, where a gating mechanism determines when the large model should step in. This method promises lower inference latency, reduced energy consumption, and a better cost-quality tradeoff while maintaining reasoning quality. It raises questions about the best signals for intervention and how to prevent over-reliance on the larger model. This matters because it offers a more efficient way to scale language models without compromising performance on reasoning tasks.
