NeoHorse-1: An Open Model Trained on Its Own Production Router's Logs

NeoHorse-1, a pair of open-weight language models from TokenRhythm Technologies and Infinigence AI, is post-trained on data generated by the very system that serves it in production. A routing harness that assigns every incoming request to one of four capability tiers logs, turn by turn, what it predicted the request needed, which tier actually handled it, and how the interaction went. The paper behind it — arXiv:2609.08183, posted September 8 by 36 authors — turns those logs into the next round of training data, in what it calls an early prototype for harness-mediated recursive self-improvement (RSI): a system using evidence of its own performance to decide what to train on next.

From router logs to a training curriculum

Before a log entry becomes training data, it passes deduplication against the evaluation suite, a rule-based structural check, and a six-dimensional semantic judgment of the interaction. What survives is ordered into a three-stage curriculum by predicted difficulty, then reused for on-policy distillation, where a fixed teacher model scores the student’s own generated answers at each stage. In a separate, controlled comparison, the authors trained the same model on their routing-harness data and on Toucan, a public synthetic tool-agent dataset, under matched recipes and budgets: the routing-harness data won by 6.26 points on average across five benchmarks.

Ten benchmarks, and where the biggest gain comes from

Across ten benchmarks of agentic execution, tool use, coding and instruction-following, the 4B model’s average score rose from 58.94 to 64.87 after this post-training, and the 9B model’s from 65.60 to 69.04, both measured against their own Qwen3.5 base models. The single largest gain in either table belongs to VitaBench — the one benchmark the authors evaluated with a substituted user-simulator and judge model, DeepSeek-V4-Flash, because the benchmark’s own recommended setup was no longer available to them. The paper discloses that substitution itself, but the biggest jump in either table still rests on the one evaluation link that departs from everyone else’s protocol.

One loop, no safety section, and a vote count that isn’t a verdict

The authors call this an initial attempt at RSI rather than a demonstration: the reported results are one pass through the evaluate-select-retrain loop, and whether gains compound across further rounds is left as future work. Neither the paper nor the GitHub README carries a safety or risk section — no red-teaming, no refusal-behavior study, no discussion of what agentic post-training on a system’s own traffic might introduce. The paper reached this queue through Hugging Face Papers’ community voting, where it holds 414 votes, a figure that measures attention on the platform; no independent press coverage of the paper turned up in this research. Both checkpoints, plus GGUF and MLX quantized versions, are published on Hugging Face and ModelScope under Apache 2.0, though the training pipeline and the routing harness itself are not.

From router logs to a training curriculum

Ten benchmarks, and where the biggest gain comes from

One loop, no safety section, and a vote count that isn’t a verdict

Further Reading