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

Benchmark Model Rank Results
agents-on-bfclv4NeoHorse-1-9B#2Accuracy: 67.43
code-generation-on-humanevalNeoHorse-1-9B#3Pass@1: 98.17
agents-on-pinchbenchNeoHorse-1-9B#5Accuracy: 82.25
instruction-following-on-ifbenchNeoHorse-1-9B#24Accuracy: 66.33
code-generation-on-livecodebench-v6NeoHorse-1-9B#68Pass@1: 0.6514