Learning from What You Can Do ── Toward recursive self-improvement in agent-native models

The Problem: From Observation to Learning

Recursive self-improvement has become a frequent reference in recent agent research. Yet the literature often describes what RSI is without explaining how it actually works. The NeoHorse-1 team confronts this gap directly. For RSI to function, they argue, three elements must be linked by a concrete mechanism. First, the system must record the capability demands it encounters and choose a corresponding model tier to respond. Second, those choices and their outcomes must be converted into training examples. Third, evaluation feedback from those examples must inform the mixture of data in the next training round. A system where these three loops rotate not once but repeatedly—this, the authors propose, is what recursive self-improvement looks like in practice.

Capability Observation via the Routing Harness

At the system's core sits a heterogeneous pool of models coupled with what the authors call intelligent routing. On each user turn, the system must select which model to deploy from its available options. Rather than optimizing for speed alone, the selection criterion is predicted capability demand. Simple queries route to lighter models; complex reasoning tasks route to heavier ones. This choice, the selected service tier, and the actual interaction outcome are all recorded as a single training example.

This record is no ordinary input-output pair. It includes the interleaved reasoning process, all tool invocations, and the harness context in which the model operated. The examples then pass through structural validation, six-dimensional semantic evaluation, and subscene-level labeling. By this multi-stage filtering, only training examples with genuine value for learning are promoted to the next phase.

Three-Stage Curriculum for Staged Improvement

Once training examples are collected, routing signals organize them into a three-stage curriculum. Each stage advances—from unsupervised pattern recognition, through routing-guided supervised fine-tuning, to teacher-supervised responses generated by the student model itself. The first stage focuses on basic pattern recognition. The second stage performs explicit fine-tuning based on routing signals. In the third stage, the teacher supervises responses the student generates, and both teacher and student progress through the same curriculum sequence—here, the students' current stage of learning is made common ground.

The design intention is transparent: rather than drowning models in data, implement the classical learning principle of presenting challenges calibrated to current ability and gradually increasing difficulty—but in a computational system.

Capability-Guided Allocation and Mixture Adaptation

The final piece of the post-training pipeline is capability-guided allocation. When evaluation feedback accumulates, it does not remain a static record of which model succeeded at which task. Instead, the feedback directly determines how data is mixed in the next training round. Tasks where capability was required, tasks where improvement remains, and information density across all tasks are quantified and factored into the mixture for the following iteration.

This is curriculum learning evolved. Traditional curriculum learning relies on either human designers to set difficulty levels or statistical heuristics. Here, the difference between what the system learned to do and what it could not learn becomes the input to the next curriculum plan itself.

Empirical Results: The 4B and 9B Numbers

The authors evaluated their approach on eleven benchmarks spanning agent tasks, tool use, coding, and instruction-following. The scope is significant: results matter only if they hold across diverse domains.

The headline numbers tell a particular story. For the 4B model, macro-average performance rose from 58.94 to 64.87. For the 9B model, it rose from 65.60 to 69.04. The most notable observation is that the post-trained 4B model substantially narrows the gap to the base 9B model. This suggests that post-training does not merely add points to a score, but potentially enables a smaller model to approach the performance of a larger one—a shift from scale-dependent gains to capability-specific gains.

A critical qualifier must be noted: these results come from a preprint, a format published before peer review. Whether independent verification will support the authors' claims remains an open question.

The Dream of Recursive Self-Improvement

What strikes an observer of NeoHorse-1 is the balance between ambition and care. The authors invoke the ambitious vision of recursive self-improvement while grounding it in concrete, implementable detail. Reasoning traces, tool calls, staged training curricula—everything is presented in executable form.

Yet the meaning of recursion itself warrants scrutiny. What the authors demonstrate is a system in which multiple improvement loops rotate within a single training round. Whether a deeper kind of recursion—in which the model redesigns its own learning process across multiple independent iterations—has been achieved remains unclear at present.

Agent-Native Post-Training in its Context

Why has agent-native post-training become a focus of research attention now? The backdrop is a shift in what models do. Models are transitioning from generation tools that respond to input toward autonomous systems that act, decide, and learn from failure. Earlier models had a simpler role: generate the most likely response to a given prompt. Agent-native models face a different task. They must orchestrate multiple tools, make decisions under uncertainty, and incorporate feedback into future action.

Such a complex action space cannot be handled by simple supervised fine-tuning alone. What becomes essential is for the model to know its own capability boundaries. This self-knowledge becomes the foundation for the next phase of growth. NeoHorse-1 offers one answer to this new challenge. Through the routing harness infrastructure, a model observes its own ability limits, and this observation influences what it learns next—a pattern that may well define the direction of agent research in the years ahead.