Hugging Face has introduced NeoHorse-1, a family of agent-native models designed to explore recursive self-improvement (RSI) through a novel agentic post-training approach. The system combines a heterogeneous model pool with intelligent routing, recording capability demands and service tiers for each user turn to generate training examples that preserve interleaved reasoning and tool calls.
The training pipeline employs structural validation, six-dimensional semantic evaluation, and subscene-level labeling to admit high-quality examples. Routing signals organize supervised fine-tuning into a three-stage curriculum and extend to routing-guided on-policy distillation, where a teacher supervises student responses under the same progression.
A key innovation is the capability-guided allocation loop, which converts evaluation feedback into the next training mixture, closing an evaluation-selection-update cycle. This mechanism allows the system to learn from what it has just learned, a core requirement for RSI.
Across eleven benchmarks covering harness-based agents, tool use, coding, and instruction following, post-training raised macro-average scores from 58.94 to 64.87 for the 4B model and from 65.60 to 69.04 for the 9B model, substantially narrowing the aggregate gap between the post-trained 4B and the 9B base model.
NeoHorse-1 serves as an initial prototype of this feedback-driven process, offering a path toward harness-mediated RSI across successive iterations. The project is available on Hugging Face and GitHub.