The interesting claim in NeoHorse-1 isn’t “recursive self-improvement” — that phrase is doing a lot of marketing work across the field right now. It’s the mechanism underneath: the routing layer, the thing that decides which model in a heterogeneous pool handles each turn, becomes the training signal itself.
- 🎯 Routing metadata is the label: predicted capability demand and selected service tier for each turn get recorded and converted into training examples — the harness’s own decisions supervise the next round.
- 🔁 The loop closes on evaluation: capability-guided allocation turns eval feedback into the next training mixture, so what the system learns to do shapes what it learns from next.
- 📚 A curriculum, not a data dump: routing signals organize SFT into a three-stage curriculum and extend into routing-guided on-policy distillation, with a teacher grading student rollouts along the same progression.
- 📊 Small models close the gap: post-training lifts a 4B model’s macro-average from 58.94 to 64.87 across eleven agent, tool-use, and coding benchmarks, narrowing the distance to the 9B base.
- ⚠️ RSI is still aspirational: the authors call it an initial prototype and a path toward harness-mediated self-improvement — one iteration, not a runaway loop.
What lands for me is the reframe of routing from a cost knob into a data-generation engine. Most of us treat the router as the thing that saves money at inference time; NeoHorse treats every routing decision as a labeled training signal. Whether that survives past one iteration — before a model starts routing in ways that reinforce its own blind spots — is the question the HF paper page leaves open. My bet: the failure mode of self-routing curricula won’t be raw capability, it’ll be diversity collapse.