Tag
The paper introduces Hermes, a family of configurable harnesses that let models control context allocation across inference-time compute, and Hermes-Learn, a two-stage framework that trains models to make these contextual reasoning decisions adaptively, closing the gap for smaller open-source models.