AIAny
Icon for item

NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness

Explores a practical mechanism for recursive self-improvement by post-training LLMs: uses a routing harness to record agent executions and convert traces into curriculum-guided supervised fine-tuning and on-policy distillation data, closing an evaluation-selection-update loop and improving benchmark performance.

Introduction

Most model improvement pipelines treat user interactions as isolated examples. NeoHorse-1 reframes agent executions as structured curriculum signal: by recording routing decisions, tool calls, and outcomes, it converts real harnessed agent trajectories into validated training examples and then reintroduces updated models into the same harness to close a feedback loop.

Key Findings
  • Routing-guided curriculum and on-policy distillation: routing signals organize supervised fine-tuning into a three-stage curriculum and supervise student responses under the same progression, preserving interleaved reasoning and tool context.
  • Evaluation-selection-update loop: capability-guided allocation turns evaluation feedback into the next training mixture, creating a prototype closed loop for iterative improvement (a step toward recursive self-improvement).
  • Empirical gains: post-training raises the ten-benchmark macro-average from 58.94 to 64.87 for the 4B model and from 65.60 to 69.04 for the 9B model, narrowing the gap between model sizes across agent, coding, and instruction-following tasks.
  • Data & quality controls: heterogeneous model pool routing, structural validation, six-dimensional semantic evaluation, subscene-level labeling, and decontamination to preserve execution context while reducing noise.
Who it's for and trade-offs

Great fit if you run an agent harness or multi-model routing infrastructure and want to turn real executions and tool use into training signal without losing execution context. Look elsewhere if you need plug-and-play performance gains without building routing/evaluation pipelines: the approach requires investment in routing instrumentation, evaluation labeling, and curriculum design, and the paper presents an initial prototype rather than a turnkey solution.

Methodology overview

The system pairs a heterogeneous model pool with intelligent routing that records predicted capability demand and chosen service tier per turn. These traces are validated, semantically evaluated, labeled at subscene granularity (Scene/Goal/Outcome), and mixed into routing-guided SFT and on-policy distillation schedules; updated models can re-enter the harness to form repeated evaluation-selection-update iterations.

Information

  • Websitearxiv.org
  • AuthorsGuoliang Cao, Guohao Dai, Tianyu Guo, Kai Han, Hailin Hu, Zihan Jiang, Xiang Kuang, Boxun Li, Yulong Li, Zehua Pei
  • Published date2026/09/08

More Items

Evaluates a weak teacher's RL-induced policy shift on the student's own rollouts and amplifies verifier-supported updates so stronger models can learn from weaker supervisors and surpass them. It rescales only verifier-supported policy-gradient components to preserve optimization fixed points while accelerating learning, reducing student updates versus standard RL or distillation.

Unifies continuous streaming perception, full‑duplex real‑time interaction, and agentic tool use across video, audio, and text. Proposes a Cerebellum–Brain split—streaming Thinker‑Talker cerebellum for low‑latency interaction plus a Brain for complex reasoning—and is released with models, code, and data.

Combines sparse verifier outcomes with dense privileged‑hindsight token scoring to learn an outcome‑calibrated, normalized distribution over complete responses for on‑policy self‑improvement. Key features: sign‑gated guidance (retain/reverse/disable per verifier advantage), profiled trajectory balance with one log‑partition per rollout group, and explicit correction against false‑positive self‑guidance.