NeoHorse-1 Pitches Routing Loop as Recursive Self-Improvement

Post-training with a routing-harness loop lifted a 4B model's macro-average across eleven benchmarks from 58.94 to 64.87, and a 9B model from 65.60 to 69.04, according to a NeoHorse-1 paper on arXiv.

The authors describe NeoHorse-1 as "a family of agent-native models" whose training system "combines a heterogeneous model pool with intelligent routing," recording predicted capability demand, selected service tier, and the interaction that followed for each user turn. Those records become training examples that preserve "interleaved reasoning, tool calls, and harness context," and are admitted through structural validation, six-dimensional semantic evaluation, and subscene-level labeling. Routing signals then organize supervised fine-tuning into a three-stage curriculum and extend into routing-guided on-policy distillation, where a teacher supervises student-generated responses under the same progression.

"What the system learns to do shapes what it learns from next," the paper writes of the resulting evaluation-selection-update loop. The team calls NeoHorse-1 "an initial prototype of this feedback-driven process and a path toward harness-mediated RSI across successive iterations." One round of post-training is reported. The abstract publishes no per-benchmark numbers and nothing for a second iteration.

The eleven benchmarks span harness-based agents, tool use, coding, and instruction following, and the paper says post-training "substantially" narrowed the aggregate gap between the post-trained 4B and the 9B base model. It arrives the same week as our coverage of a Dwarkesh study reporting that data mixtures beat model recipes by 3.24x in pretraining.