Piccolo-27B Tops Decision Index Vision Leaderboard at 75.59
TL;DR
- Piccolo-27B hit 75.59 on the Decision Index Vision Full leaderboard, topping the overall ranking; Piccolo-9B and 4B scored 68.35 and 67.46.
- The 0.8B-to-27B decision-model family used research agents to propose data, run experiments, and refine proposals across roughly 600 experimental runs.
- Training consumed about 6,300 PPU-ZW810E accelerator-hours, and the authors release weights alongside the Data-Juicer and AgentScope ecosystems.
Piccolo-27B posted 75.59 on the Decision Index Vision Full leaderboard, top of the overall ranking, according to the paper from the team behind the Data-Juicer and AgentScope stacks. Smaller siblings Piccolo-9B and Piccolo-4B scored 68.35 and 67.46, with all three ranking first on the Private benchmark in their respective size categories.
The claim the authors lead with is not the number but the training loop. "By binding fine-grained semantic slices, training dynamics, and weight-space trajectories into a persistent research memory, our agent orchestrator transforms brute-force hyperparameter search into cumulative scientific discovery," the paper writes. Research agents propose data, run experiments, and read the results back into their next proposals across a family that spans 0.8B to 27B parameters.
Three findings inside the paper cut against the usual recipe. Earlier backbones can beat newer generations on visual tasks under identical recipes. Targeted data curation at 4B and 9B rivals untargeted 27B baselines on specialized workflows. And continuous loss reduction can decouple from, and sometimes degrade, discrete decision accuracy; the authors note that "Dynamic data curation is more compute-efficient than prolonged static training."
The budget behind it: about 600 experimental runs and roughly 6,300 PPU-ZW810E accelerator-hours. The paper landed the same day as WorldRuler, another 2026-11-19 multimodal release on our tracker, inside a 90-day window that has carried 122 multimodal stories through our feed. The authors release weights alongside the Data-Juicer and AgentScope ecosystems.
Originally reported by huggingface.co
Read the original article →Original headline: Alaya Lab's Piccolo-27B Tops Decision Index Vision at 75.59 After 600 Recursive Self-Improvement Runs