Researchers demonstrate recursive self-improvement prototype via post-training agents with smart routing

A new pre-print (arXiv:2609.08183) describes NeoHorse-1, a model family built to test whether a system can observe its own capabilities and feed those observations into the next round of learning. The core idea is an evaluation-selection-update loop in which what the model learns to do determines what it learns from next.

The system pairs a heterogeneous pool of models with a smart routing mechanism. For every user turn it logs the predicted capability demand, the service tier selected, and the interaction that follows. Those records become training examples that preserve interleaved reasoning, tool calls, and the context of the harness — the execution environment that mediates between the model and external tools.

Before entering training, examples pass through three filters: structural validation, semantic evaluation across six dimensions, and sub-scene-level labeling. Routing signals then organize supervised fine-tuning (SFT) into a three-stage curriculum and extend into on-policy distillation, where a teacher model supervises student responses generated under the same progression.

The step that closes the loop is capability-guided allocation: evaluation feedback is translated into the next training mix. That creates a concrete mechanism for recursive self-improvement, or RSI, in which observed capability dictates future training data rather than the reverse.

Across eleven benchmarks covering harness agents, tool use, code, and instruction following, post-training lifts the macro average from 58.94 to 64.87 on the 4B model and from 65.60 to 69.04 on the 9B model. In other words, the smaller model after post-training closes a substantial portion of the gap to the larger base model, a strong signal of the method's efficiency.

The paper is readable via `hf papers read 2609.08183`; as of now it is cited by 10 models and zero datasets. It remains a pre-print that has not undergone peer review, and the authors characterize NeoHorse-1 as an "initial prototype," not a finished product. Even so, the demonstration of a closed loop between evaluation and training data is a methodological step worth tracking in the state of the art of agent research.