Hermes: Learning Contextual Reasoning Unlocks Test-Time Scaling

Summary

Hermes reimagines test-time scaling by allowing the model itself to decide how context windows are allocated and reused during inference, rather than having these decisions imposed by a fixed skeleton. The two-stage training recipe, called Hermes-Learn, teaches smaller open-source models to develop adaptive contextual reasoning strategies that help them close the gap with larger models.

Why it matters

Test-time scaling is becoming the primary lever for extracting more capability from fixed-size models. However, most existing systems define context management as hard-coded logic within the inference skeleton. This work internalizes this decision-making process within the model and demonstrates that learned strategies can generalize beyond the computational power used during training and be extrapolated. This points to a more flexible, model-centric paradigm for distributing inference-time compute, an approach that could benefit a wide range of downstream tasks.

Highlights

• Shifts context allocation and reuse decisions from the inference skeleton to the model's own reasoning loop
• Hermes-Learn's two-stage training enables smaller open-source models to match the test-time scaling behavior of larger models
• Reported gains extrapolate beyond the inference compute observed during training and transfer to complementary scaling methods

Critical take

The summary contains strong claims regarding generalization and extrapolation. However, this work is still a 45-page preprint that has not undergone peer review. The scope of the comparisons, the magnitude of the 'gap' closed, and the robustness of extrapolation against unseen compute budgets have not yet been verified. Furthermore, the phrase 'simple, configurable skeletons' risks underestimating the engineering complexity required for a model to reliably manage its own context windows in practice.