NeoHorse-1: The Dawn of Self-Improving AI Agents

In the rapidly evolving world of Artificial Intelligence, the ultimate goal has always been "Recursive Self-Improvement" (RSI)—the ability for an AI to learn from its own experiences and get better over time without constant human intervention. A new research report on the NeoHorse-1 family of models provides a concrete blueprint for how to turn this vision into a reality through a process called agentic post-training.

Moving Beyond Static Data

Traditional AI training relies on static datasets—massive collections of text that models memorize and mimic. However, for AI "agents" that need to use tools, write code, and solve multi-step problems, static data isn't enough. NeoHorse-1 introduces a "routing harness" that acts like a coach. As the AI interacts with software or solves tasks, the system records everything: what the AI tried to do, which tools it called, and whether it succeeded. These real-world interaction records are then fed back into the model, creating a continuous loop where the system learns from its own successes and failures.

The Power of Intelligent Routing

A standout feature of the NeoHorse-1 architecture is its use of a heterogeneous model pool and intelligent routing. Instead of using one giant model for every task, the system predicts the "capability demand" of a user's request. It then routes the task to the most appropriate service tier. This routing signal does more than just save costs; it serves as a curriculum for the AI. By organizing training into a three-stage curriculum based on these signals, the researchers found that the models could learn complex reasoning and tool-use in a structured, efficient way.

Small Models, Big Results

The practical implications for businesses are significant. The research tested NeoHorse-1 at two scales: a 4-billion (4B) parameter model and a 9-billion (9B) parameter model. After the agentic post-training, the 4B model saw its performance jump from 58.94 to 64.87 on average across eleven benchmarks. Remarkably, the post-trained 4B model narrowed the gap so significantly that it began to rival the performance of much larger, base 9B models. For enterprises, this means more capable AI agents can be deployed using less computing power, reducing latency and operational costs.

Real-World Problem Solving

What does this look like in practice? The researchers analyzed how the models handled complex tasks like project scheduling and code repair. While standard models often failed when they encountered an unexpected error or a missing software dependency, NeoHorse-1 demonstrated "robustness." The 9B model, in particular, showed an ability to recognize when an approach was blocked, abandon the failing strategy, and switch to alternative tools to complete the task. In one instance, this adaptive behavior reduced token usage and execution time by over 70% compared to less sophisticated models.

The Future of Recursive Improvement

NeoHorse-1 represents a shift from building "static" models to building "learning systems." By closing the loop between evaluation, selection, and updates, the system ensures that what the AI learns today shapes how it solves problems tomorrow. As these models continue to iterate, we are moving closer to a future where AI agents don't just follow instructions—they refine their own expertise through experience.