Recursive AI—the idea that AI systems should increasingly build and improve other AI systems—has long been a goal for the field. A new preprint on arXiv, titled iCoder-27B: Recursive AI-Led Development of Frontier Industrial Coding Model, takes a concrete step in that direction by applying the approach to a large-scale coding model.
The paper presents iCoder-27B, a 27-billion-parameter model designed for industrial coding tasks. Rather than relying entirely on human engineers, the development process is led by AI, which plays a central role in constructing and refining the model. This positions the work as an early example of recursive AI moving beyond small experiments toward more substantial, practical systems.
The authors acknowledge that recursive self-development has so far become practical only in limited contexts—small models, bounded tasks, and other constrained scenarios. iCoder-27B appears to push against those limits, though the abstract leaves open questions about how fully the recursive loop was closed and what specific industrial coding benchmarks were used. As a preprint, the claims have not yet undergone peer review.