Recursive AI refers to the prospect of AI taking an increasingly complete role in building and improving itself, often described as the crown jewel of AI for AI. While recursive self‑development is already practical for small models, bounded tasks, and fixed time budgets, creating a release‑ready, frontier‑competitive model remains far more demanding.
This work investigates how little human involvement is sufficient for an agent to develop such a model. Human input is concentrated into a high‑density, low‑frequency interface: experts encode objectives, stage scaffolds, permission boundaries, and operating procedures as reusable research skills, while the agent instantiates these priors, selects experiments, diagnoses outcomes, and revises the training strategy.
In the challenging domain of industrial coding, the agent evolves data and coordinates supervised fine‑tuning (SFT), on‑policy self‑distillation, and reinforcement learning with verifiable rewards, ultimately producing iCoder, a 27B‑parameter model for RTL design and GPU kernel optimization.
Across seven benchmarks, iCoder leads RTLLM, outperforming GPT‑5.5 and Claude‑Opus‑4.8; it ranks second on CVDP and KernelBench L2, exceeding GPT‑5.5 by 16 points; and it ties Claude‑Opus‑4.8 for the best TritonBench result. Exploratory case studies further demonstrate iCoder’s competitive iterative RTL and GPU‑kernel optimization using substantially fewer tokens.
These findings chart an engineering path toward recursive self‑improvement, where humans distill the principles of model building, agents operationalize them through evidence‑driven experimentation, and each generation of AI becomes a more capable architect of the next.
Review