Hermes

1 · Contextual Reasoning

Two Complementary Dimensions of Reasoning

“The capacity of the human mind for formulating and solving complex problems is very small compared with the size of the problems …”

Test-time scaling spends more compute at inference, increasingly across many separate contexts coordinated by a harness. That brings out two complementary dimensions of reasoning.

The Best Use of Additional Contexts Changes as Reasoning Unfolds

Existing harnesses fix these decisions in advance: Reasoning Cache compresses and continues, Recursive Agent Optimization delegates subproblems. Yet a hard proof may first need several approaches in parallel, then a check of a promising step, then a clean continuation.

This motivates a flexible harness that exposes multiple ways to use additional contexts and lets the model choose among them based on the current state of the solution.

2 · Hermes

Hermes: A Family of Simple, Flexible Harnesses

Hermes progressively shifts contextual reasoning control from the harness to the model.

Design Desiderata

Contextual Reasoning via Delegation and Digestion

Two ways to reason beyond one context window: in parallel, by delegating tasks to fresh contexts, and in sequence, by digesting progress into a fresh context.

Configurable Compute: Breadth, Depth and Length

The root agent takes T steps per window and may launch up to B subagents at each step but the last. Change B, the depth D or the windows W to see the compute grow; every box is one 8K-token context.

Steerable Control over Delegation

The agent must also decide what each fresh context is for. Hermes groups delegation strategies into two categories:

• Search makes progress. Solve the whole problem again with no set approach (Same), with different named approaches (Different), or split it into subproblems (Split).
• Verify checks the current answer: directly (Answer), through the steps behind it (Steps), or by finding why candidates disagree (Disagreement).

Three steering levels set who chooses: at L3 the harness names the strategy, at L2 only the category, and at L1 the model picks from every strategy as reasoning unfolds.

Test-Time Scaling with Hermes Relies on Contextual Reasoning

Given up to seven 8K-token contexts, capable models gain, but Qwen3-4B, the small open model we train, loses more than half its accuracy: the contextual reasoning gap.

3 · Hermes-Learn

Hermes-Learn: Learning Contextual Reasoning

Hermes-Learn teaches delegation, deciding what to send to fresh contexts, in two stages. Digestion is not trained; the model relies on its existing ability to summarize.

A Curriculum That Progressively Reduces Harness Steering

RL trains on a mixture of harnesses that shifts over time, from guided L2 episodes to L1, so the model is increasingly given control over its contextual reasoning decisions.

Data Augmentation for Balanced Strategy Coverage

Competition problems do not exercise delegation strategies evenly, so RL adds 500 synthetic problems whose structures favor different strategies; the same pipeline builds the 88-problem synthetic benchmark.

4 · Main Results

Experimental Setup and Main Results

We answer four research questions:

1. Does Hermes-Learn Improve Contextual Reasoning?
2. What Contextual Reasoning Behaviors Emerge Through Training?
3. Which Components of the Training Recipe Drive These Improvements? Section 5
4. Do These Gains Generalize to New Settings? Section 5

5 · Analysis

Analysis and Discussion

Citation

arXiv: arxiv.org/abs/2609.38332

@article{li2026hermes,
  title   = {Hermes: Learning Contextual Reasoning Unlocks Test-Time Scaling},
  author  = {Li, Xinyu and Goswami, Mononito and Liu, Hao and Kanakaris, Nikos and
             Huang, Langlin and Jana, Prithwish and Bl{\"o}baum, Patrick and Jain, Purak},
  journal = {arXiv preprint arXiv:2609.38332},
  year    = {2026},
  url     = {https://arxiv.org/abs/2609.38332}
}