TokenRhythm Open-Sources NeoHorse-1: Feeding Router Trajectories to Small Models for a Verified RSI Loop

机器之心 · wechat · 2026-09-09

TokenRhythm has open-sourced NeoHorse-1 (4B and 9B, built on Qwen3.5). The core idea: let the routing system (OpenSquilla), originally built to dispatch work across models, also record every execution trajectory, then use those trajectories to post-train a small model so it learns to complete what a whole model pool once did. The team calls this a reproducible single-round validation of an RSI (Recursive Self-Improvement) loop, with Tsinghua, PKU and Infinigence involved.

Training proceeds in four steps: preserve full reasoning-action-feedback trajectories (100K–1M anonymized traces); filter quality via deterministic rules plus six-dimension semantic evaluation; perform curriculum learning using the router's predicted capability tiers (C0–C3); and apply router-guided On-Policy Distillation, with the next data mix steered by a "capability gap map." On evaluation, the 4B model improved from 58.94 to 64.87 (+5.93) and the 9B from 65.60 to 69.04 (+3.44), with the 4B nearing the Qwen3.5-9B base; gains concentrate on tasks with clear workflows and verifiable feedback, while long-horizon tasks remain hard.

The author cautions this is not full RSI—only one evaluate-select-update round is validated, and cross-generation gains remain unproven—but its significance is framing RSI as a reproducible engineering problem: a continuously running Harness that leaves structured records becomes the cheapest and hardest-to-copy training data for next-generation agent models.

Original post →

More from Models

Models channel →