NeoHorse-1：第一个认真搭递归自我改进原型的开源小模型

📌 论文：NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness arXiv：https://arxiv.org/abs/2609.08183 GitHub：https://github.com/TokenRhythm/NeoHorse HuggingFace：https://huggingface.co/collections/TokenRhythm/neohorse-1 ModelScope：https://modelscope.cn/collections/TokenRhythm/NeoHorse-1 License：Apache 2.0 · 发布：2026-09-07

---

BLUF：NeoHorse-1 是 TokenRhythm 发布的 4B/9B 开源 Agent 后训练模型，基于 Qwen3.5，Apache 2.0。它的特别之处不在于数字——尽管十项基准平均 +5.93 已经够扎实——而在于它的设计目标：用一个叫 Routing Harness 的机制，搭出递归自我改进（RSI）的第一个原型闭环。训练出来的模型可以重新进入 harness，形成评估-选择-更新的循环。这是一个起点，而不是终点。

---

递归自我改进：为什么这个方向重要

“递归自我改进”（Recursive Self-Improvement，RSI）是 AI 安全领域多年来反复讨论的概念——一个系统用自己的能力改进自己，改进后的版本再去改进，如此循环。

但大多数讨论停留在假设层面。NeoHorse-1 做了一件有意思的事：用一个工程化的 harness，把这个概念做成了可以实际运行的东西。

当然，它现在只是”原型”。但”一个真实运行的原型”和”一个思想实验”之间的距离，远比很多人以为的大。

---

Routing Harness：核心机制

NeoHorse-1 的核心不是模型本身，而是训练它的 Routing Harness（路由挽具）。

这个机制的运作逻辑是：

任务进入 → Routing Harness 分发给异构模型池
              ↓
模型执行任务 → 记录工具调用轨迹和结果
              ↓
Harness 估算能力需求缺口
              ↓
把能力级别的反馈注入下一轮训练混合
              ↓
更新后的模型重新进入 Harness → 循环

关键设计：反馈不是人工设计的——是模型自己在执行任务时暴露的能力边界，被 harness 捕获，转化为训练信号。

这和普通的 SFT 或 RLHF 有本质区别：

• 普通 SFT：人工收集数据 → 训练
• 普通 RLHF：人工或模型打分 → 优化
• Routing Harness：让模型在真实 agent 环境里跑任务 → 自动捕获执行轨迹和失败模式 → 直接转化成下一轮训练的信号

---

训练方法：两个核心组件

Routing-guided Curriculum SFT

不是把所有数据随机混在一起训练，而是根据模型当前的能力边界，动态决定用什么难度和类型的数据。Harness 知道模型在哪些任务上容易失败，就把相关数据多塞一点进下一轮训练。

Routing-guided On-policy Distillation

在训练时，用的是模型自己生成的轨迹，而不是纯粹的人工数据。这保证了训练分布和实际执行分布的一致性——训练时见过的东西，和推理时遇到的东西，尽量是同一种结构。

数据质量管道：精确/近似去重、评估集去污染（防止基准数据泄露）、结构验证、六维语义评估、Scene/Goal/Outcome 三层子场景标注。

---

测评结果

测评用 SGLang v0.5.17，thinking 模式开启，十项基准覆盖 Agent 能力、编程、指令遵循三个维度。

4B 档（对比 Qwen3.5-4B）

基准	Qwen3.5-4B	NeoHorse-1-4B	Δ
QwenClawBench	38.47	44.68	+6.21
WorkBuddy Bench	24.62	34.41	+9.79
PinchBench	71.19	77.33	+6.14
VitaBench	21.50	32.00	+10.50
BFCL v4	61.02	61.79	+0.77
tau2-Bench	84.29	88.46	+4.17
HumanEval	87.20	96.95	+9.75
LiveCodeBench v6	53.71	59.43	+5.72
IFBench	60.33	65.33	+5.00
IFEval	87.06	88.35	+1.29
十项平均	58.94	64.87	+5.93

9B 档（对比 Qwen3.5-9B）

基准	Qwen3.5-9B	NeoHorse-1-9B	Δ
QwenClawBench	44.04	48.73	+4.69
WorkBuddy Bench	39.60	40.15	+0.55
PinchBench	74.55	82.25	+7.70
VitaBench	31.25	42.25	+11.00
BFCL v4	64.88	67.43	+2.55
tau2-Bench	88.04	90.82	+2.78
HumanEval	92.68	98.17	+5.49
LiveCodeBench v6	65.14	65.14	+0.00
IFBench	66.33	66.33	+0.00
IFEval	89.46	89.09	-0.37
十项平均	65.60	69.04	+3.44

提升最明显的几项都在 Agent 类基准——这和 Routing Harness 的设计目标一致。VitaBench（+10.50/+11.00）的提升尤其大，这个基准测的是复杂 agent 任务的完整流程。

---

规格与部署

模型规格：

• 上下文长度：原生 262,144 tokens，可扩展到 1,010,000
• 权重格式：Safetensors / BF16
• GGUF 版本：BF16、8-bit、5-bit、4-bit 全提供（本地部署友好）

SGLang 部署：

pip install "sglang==0.5.17"
python3 -m sglang.launch_server \
  --model-path "/path/to/NeoHorse-1-4B" \
  --served-model-name neohorse-1-4B \
  --host 0.0.0.0 --port 30000 \
  --context-length 262144 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder

vLLM 部署：

vllm serve "/path/to/NeoHorse-1-4B" \
  --served-model-name neohorse-1-4B \
  --max-model-len 262144 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder

---

为什么值得关注

NeoHorse-1 有两个地方值得认真看：

一、RSI 原型是一个新的研究方向起点

到目前为止，大多数开源模型的后训练都是”一次性的”：收数据、训模型、发布。NeoHorse-1 的 Routing Harness 第一次把”模型进入 harness → 执行 → 能力反馈 → 重新训练”做成了一个可以迭代的循环。现在 NeoHorse-1 是这个循环的第一轮，下一轮会是什么？

二、Agent 类基准的大幅提升是可解释的

NeoHorse-1 不是在所有任务上都显著提升——LiveCodeBench 和 IFEval 的提升很小甚至持平。但在 Agent 类任务（WorkBuddy、VitaBench、PinchBench）上的提升格外大。这和 Routing Harness 的设计方向高度一致：它就是为了让模型更好地完成 agent 任务而设计的。结果和设计目标吻合，说明这套方法是在解决真实问题。

---

开源代码与模型仅供学习、勿直接用于工作。

© 2026 Author: Mycelium Protocol. 本文采用 CC BY 4.0 授权——欢迎转载和引用，须注明作者姓名及原文链接，不得去除署名后以原创发布。

📌 Paper: NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness arXiv: https://arxiv.org/abs/2609.08183 GitHub: https://github.com/TokenRhythm/NeoHorse HuggingFace: https://huggingface.co/collections/TokenRhythm/neohorse-1 ModelScope: https://modelscope.cn/collections/TokenRhythm/NeoHorse-1 License: Apache 2.0 · Released: 2026-09-07

---

BLUF: NeoHorse-1 is TokenRhythm’s 4B/9B open-source agent post-training model based on Qwen3.5, Apache 2.0. Its significance isn’t just the numbers — though +5.93 average across ten benchmarks is solid — but the design goal: using a mechanism called the Routing Harness to build the first prototype closed loop for recursive self-improvement (RSI). The trained model can re-enter the harness, forming an evaluate–select–update cycle. This is a starting point, not an endpoint.

---

Why Recursive Self-Improvement Matters

“Recursive Self-Improvement” (RSI) has been discussed in AI safety circles for years — a system using its own capabilities to improve itself, with each improved version doing the same, cycling forward.

Most of that discussion stays hypothetical. NeoHorse-1 does something interesting: it engineers a harness that turns this concept into something that actually runs.

It’s a “prototype” right now. But the distance between “a working prototype” and “a thought experiment” is much larger than most people assume.

---

The Routing Harness: Core Mechanism

NeoHorse-1’s core isn’t the model itself — it’s the Routing Harness used to train it.

The logic:

Task arrives → Routing Harness assigns to heterogeneous model pool
                    ↓
Model executes → Records tool interaction traces and outcomes
                    ↓
Harness estimates capability demand gaps
                    ↓
Injects capability-level feedback into next training mixture
                    ↓
Updated model re-enters Harness → Loop

Critical design: the feedback isn’t manually curated — it’s the model’s own capability boundaries exposed during task execution, captured by the harness, and converted into training signal.

This is fundamentally different from standard SFT or RLHF:

• Standard SFT: humans collect data → train
• Standard RLHF: humans or model score outputs → optimize
• Routing Harness: run model on real agent tasks → automatically capture execution traces and failure patterns → convert directly into next training round’s signal

---

Training Methods: Two Core Components

Routing-guided Curriculum SFT

Not mixing all data randomly. Instead, based on the model’s current capability boundaries, dynamically decide what difficulty and type of data to use. The harness knows where the model tends to fail and feeds more relevant data into the next round.

Routing-guided On-policy Distillation

Training uses trajectories the model itself generated, not purely human-curated data. This keeps training distribution aligned with actual execution distribution — what the model sees during training stays structurally similar to what it encounters at inference time.

Data quality pipeline: exact/near-duplicate removal, evaluation decontamination (preventing benchmark leakage), structural validation, six-dimensional semantic evaluation, Scene/Goal/Outcome three-layer subscene labeling.

---

Benchmark Results

Evaluated with SGLang v0.5.17, thinking mode enabled, ten benchmarks across agent capability, coding, and instruction following.

4B track (vs Qwen3.5-4B)

Benchmark	Qwen3.5-4B	NeoHorse-1-4B	Δ
QwenClawBench	38.47	44.68	+6.21
WorkBuddy Bench	24.62	34.41	+9.79
PinchBench	71.19	77.33	+6.14
VitaBench	21.50	32.00	+10.50
BFCL v4	61.02	61.79	+0.77
tau2-Bench	84.29	88.46	+4.17
HumanEval	87.20	96.95	+9.75
LiveCodeBench v6	53.71	59.43	+5.72
IFBench	60.33	65.33	+5.00
IFEval	87.06	88.35	+1.29
Average	58.94	64.87	+5.93

9B track (vs Qwen3.5-9B)

Benchmark	Qwen3.5-9B	NeoHorse-1-9B	Δ
QwenClawBench	44.04	48.73	+4.69
PinchBench	74.55	82.25	+7.70
VitaBench	31.25	42.25	+11.00
tau2-Bench	88.04	90.82	+2.78
HumanEval	92.68	98.17	+5.49
Average	65.60	69.04	+3.44

The biggest gains are in agent-class benchmarks — exactly what the Routing Harness was designed for. VitaBench’s +10.50/+11.00 gains are notable; this benchmark tests complete-pipeline complex agent tasks.

---

Specs and Deployment

Model specs:

• Context length: 262,144 natively, extensible to 1,010,000 tokens
• Format: Safetensors / BF16
• GGUF: BF16, 8-bit, 5-bit, 4-bit all available (local-friendly)

SGLang deployment:

pip install "sglang==0.5.17"
python3 -m sglang.launch_server \
  --model-path "/path/to/NeoHorse-1-4B" \
  --served-model-name neohorse-1-4B \
  --context-length 262144 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder

---

Why This Is Worth Watching

Two things stand out:

One: RSI prototype as a new research direction starting point

Most open-source model post-training is a one-shot operation: collect data, train model, release. NeoHorse-1’s Routing Harness is the first to make “model enters harness → executes → capability feedback → retrain” an iterable cycle. NeoHorse-1 is round one. What does round two look like?

Two: Agent benchmark gains are mechanistically explained

NeoHorse-1 doesn’t improve uniformly across all tasks — LiveCodeBench and IFEval show minimal gains. But agent-class tasks (WorkBuddy, VitaBench, PinchBench) improve dramatically. This aligns precisely with the Routing Harness’s design purpose. When results match design intent, it’s evidence the method is solving a real problem.

---

Open-source code and models are for learning purposes only — do not use directly in production work.

© 2026 Author: Mycelium Protocol. Licensed under CC BY 4.0 — free to share and adapt with attribution. You must credit the author and link to the original; removing attribution and republishing as original is not permitted.