NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness: plain AI paper summary
AI-generated summary. Read the original paper before relying on it.
Hugging Face paper page · Original paper on arXiv
Plain summary
This paper describes a way to train AI agents by learning from their own work. The system watches which model it picks for each task and how the task turns out, then turns those records into new training data. It uses that data to teach smaller models to do more of the work, so the system gets better over time.
What changed
The authors built a training loop that uses routing choices and task results as the source of new training examples. They tested it on two model sizes and eleven tasks covering agents, tools, coding, and following instructions. Scores went up on both sizes, and the smaller trained model closed much of the gap to the larger untrained one.
Why it matters
It shows a practical path to improving agent models without hand-labeling every example. For an agency, that points toward cheaper, smaller models that can still handle routine work, which matters for cost and speed.
Who should care
Teams that run AI agents in production and care about model cost, routing quality, and steady improvement from real usage logs.
Business lead ideas
A routing and logging audit that captures which requests need which model tier and where cheap models already work.
- Target buyer: Small SaaS or WordPress agency owners running AI features
- Problem: They pay for big models on every request, even simple ones, and have no way to improve from real usage.
- First action: Add turn-level logging of model choice, task type, and outcome to one existing AI feature.
- Measure: Share of requests served by the cheaper tier without a drop in task success.
A small pilot that converts logged agent turns into a validated fine-tuning set and re-tests on their own tasks.
- Target buyer: Product teams with agent workflows
- Problem: Their agent quality is static; nothing turns usage data into better behavior.
- First action: Sample 200 logged turns, validate them against a simple rubric, and run one fine-tune on a small open model.
- Measure: Before-and-after success rate on a held-out set of their real tasks.
Limits
The paper is a research prototype, not a product. It reports benchmark averages, not real customer outcomes. The loop depends on good routing signals and careful data validation, which are hard to get right. No evidence yet that the loop keeps improving over many rounds.
Verdict
Watch