Piccolo: Co-Evolving Decision Models and Data through Recursive Self-Improvement

arxiv.org1 min read

Researchers introduced Piccolo, a family of multimodal decision models ranging from 0.8B to 27B parameters that coevolve capabilities and training data through agentic data engineering and recursive self-improvement. The models achieve top results on the Decision Index Vision Full leaderboard and Private benchmark. Their large-scale exploration yielded counter-intuitive insights regarding backbone-recipe co-adaptation, targeted data curation, and the decoupling of continuous loss reduction from discrete decision accuracy. Open article →