Piccolo: 0.8B-27B decision models co-evolved with data via recursive self-improvement, tops Decision Index
agentscope-ai · hf · 2026-11-19
The Piccolo team released Piccolo, a family of multimodal decision models (0.8B–27B) that co-evolves capabilities and training data via Agentic Data Engineering and recursive self-improvement (RSI). The models output probabilities over candidate answers for text/visual judgments, agent harnesses, data workflows, and games.
Results
- Piccolo-27B tops the Decision Index Vision Full leaderboard at 75.59; Piccolo-9B and 4B score 68.35 and 67.46
- All three rank first on the Private benchmark in their size classes
Method: research agents analyze model behavior, build training data, run experiments, and refine subsequent proposals; semantic slices and a shared research memory connect actions across the family, with data selection and feedback-driven refinement at the core of recipes spanning hundreds of thousands of examples.
Three counter-intuitive findings
- Backbone-recipe co-adaptation beats model recency: older backbones can surpass newer ones under identical recipes
- Targeted data curation lets 4B–9B models rival 27B baselines on specialized workflows
- Continuous loss reduction can decouple from, and even degrade, discrete decision accuracy
Weights and tools from the Data-Juicer and AgentScope ecosystems are released.
More from Research
- Google's GUARD paper: frozen base model, self-editing agent harness gains +8.89 points — omarsar0 · 2026-11-19
- Kezek splits mono two-person conversations into per-speaker tracks for full-duplex dialogue training data — manifestro · 2026-11-19
- Elie Bakouch: current AI research benchmarks overfit and can't track RSI — eliebakouch · 2026-11-19
- Hutter Prize text compression jumps ~10% by overfitting small transformers for arithmetic coding — ricklamers · 2026-11-19
- HybridApexW optimizer: 2x faster than torch.compile AdamW+Muon baseline — omendUUde · 2026-11-19
- Sandia Labs' VASt model predicts high-entropy alloy properties from tens of samples — CurieuxExplorer · 2026-11-19