Piccolo: Co-Evolving Decision Models and Data through Recursive Self-Improvement
Abstract
We present Piccolo, a family of multimodal decision models (0.8B to 27B) that coevolves model capabilities and training data via Agentic Data Engineering and recursive self-improvement (RSI). The models predict probabilities over candidate answers for general-purpose text and visual judgments, agent harnesses, data workflows, and interactive games. Piccolo-27B tops the Decision Index Vision Full leaderboard with a score of 75.59, while Piccolo-9B and Piccolo-4B achieve 68.35 and 67.46, respectively. All three rank first on the Private benchmark in their respective model-size categories. Research agents analyze model behavior, construct training data, conduct experiments, and use the results to refine subsequent proposals. Semantic slices and a shared research memory connect these actions across the model family's development. With training recipes comprising hundreds of thousands of examples, data selection and feedback-driven refinement are central to the self-improvement process. Our large-scale RSI exploration yields three counter-intuitive insights for decision-model development: (1) Backbone-recipe co-adaptation outweighs raw model recency: earlier backbones can surpass newer generations on visual tasks under identical recipes; (2) Targeted data curation enables 4B-9B models to rival 27B baselines on specialized workflows; and (3) Continuous loss reduction can decouple from and even degrade discrete decision accuracy. Together with the released weights and tools from the Data-Juicer and AgentScope ecosystems, Piccolo provides an actionable blueprint for building frontier decision models and scaling data-driven RSI.
Figures & tables
| Component | Fixed within a comparison | Varied between comparisons |
|---|---|---|
| Data | Pool version, group IDs, admission rules | Selection, mixtures, synthesis, new sources |
| Training | Parent weights, targets, optimizer, schedule | Parent, objective, learning rate, merging |
| Evaluation | Questions, scoring, inference settings | Input arrangement, diagnostic cohorts |
| Research | Target, comparator, devices and time | Analysis tools, priorities, follow-ups |
| Model | Full | Private | Public |
|---|---|---|---|
| Piccolo-27B | 75.59 | 70.97 | 80.22 |
| JDI-Decision V1 | 75.05 | 70.52 | 79.58 |
| WeMM-JEV-27B | 73.81 | 69.70 | 77.93 |
| Valen 1.5 27B | 73.29 | 63.68 | 82.90 |
| Sieve 31B v3 | 72.07 | 68.57 | 75.57 |
| StartLux v1.1 27B | 71.61 | 69.39 | 73.83 |
| Model / text share | Text | Vision |
|---|---|---|
| Text-oriented parent | 68.53 | 73.16 |
| Image–text parent | 67.89 | 77.29 |
| Merge: 25% text | 67.94 | 76.92 |
| Merge: 50% text | 68.15 | 75.70 |
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
| Primary use | Source collections |
|---|---|
| Language understanding, retrieval, and dialogue (10) | Banking77 Casanueva et al. [2020] , CLINC150 Larson et al. [2019] , ESCI Reddy et al. [2022] , GoEmotions Demszky et al. [2020] , iSarcasmEval Abu Farha et al. [2022] , MASSIVE FitzGerald et al. [2022] , MNLI Williams et al. [2018] , SGD Rastogi et al. [2020] , SNLI Bowman et al. [2015] , WANLI Liu et al. [2022] |
| Knowledge and reasoning (10) | AQuA-RAT Ling et al. [2017] , ARC Clark et al. [2018] , BoolQ Clark et al. [2019] , CommonsenseQA Talmor et al. [2019] , Cosmos QA Huang et al. [2019] , FinQA Chen et al. [2021] , HellaSwag Zellers et al. [2019] , MuSiQue Train Trivedi et al. [2022] , QASC Khot et al. [2020] , WinoGrande Sakaguchi et al. [2019] |
| Preferences and quality judgments (5) | Bekko Selected hotchpotch [2026] , HelpSteer2 Wang et al. [2024] , PRM800K Lightman et al. [2023] , Prometheus Kim et al. [2023] , RAGTruth Niu et al. [2024] |
| Actions and workflows (4) | JevForge Mind2Web AndeyTait [2026] , Deng et al. [2023] , Procedural Typed Decisions tasksource [2026a] , Sileo et al. [2026] , ToolJudge Local, When2Call NVIDIA [2025] |
| Converted and synthesized decision collections (14) | Eikos Decisions Vicentino [2026] , Helmo Synthetic helmo [2026] , JEV Decisions Clean50K samatv256 [2026] , JEV Distill v3 SargeDev [2026] , MinoJev Data zeredy879 [2026] , Open JEV ZefanCai [2026b] , Open JEV v1.1 ZefanCai [2026a] , Plumb Decisions crh225 [2026] , Synthetic Typed Decisions tasksource [2026b] , TaskSource JEV tasksource [2026c] , Sileo [2024] , Typed Decisions LocalLLaMA [2026] , Typed Decisions JA hiroki-abe-58 [2026] , LocalLLaMA [2026] , Typed Decisions PT/ES telepatia-ai [2026] , LocalLLaMA [2026] , Typed Decisions Synth Nazeem [2026] |
| Axis | Values | What it describes |
|---|---|---|
| Domain | 334 | Subject area, with 26 parent groups. |
| Capability | 144 | Required skill, with 14 parent groups. |
| Task function | 20 | The requested judgment, such as routing or tool-call validity. |
| Input quality | 3 | Usable input, missing information, or conflicting rules. |
| Verification method | 4 | Programmatic, semantic-model, hybrid, or manual checking. |
| Decision relation | 4 | Direct decision, evaluative judgment, decision support, or standalone solution. |
| Model | Full | Private | Public | Difference |
|---|---|---|---|---|
| Piccolo-27B | 75.59 | 70.97 | 80.22 | 4.63 |
| Piccolo-9B | 68.35 | 65.04 | 71.67 | 3.32 |
| Piccolo-4B | 67.46 | 63.49 | 71.42 | 3.96 |
| Model | HarnessJevBench | DataJevBench | JevBench |
|---|---|---|---|
| Piccolo-27B | 84.41 | 82.60 | 92.21 |
| JDI-Decision-V1 | 82.71 | 84.40 | 90.48 |
| WeMM-JEV-27B | 82.03 | 82.40 | 92.21 |
| RBDash-Decision v0.2 | 72.37 | 79.60 | 88.31 |
| Valen-Decision 1.5 27B | 79.15 | 80.20 | 91.34 |
| Perplexity Decider v1.3 (27B) | 82.03 | 83.00 | 90.91 |
| Model | CV-Bench | BLINK | RealWorld QA | CharXiv | Infographic VQA | Mind2Web |
|---|---|---|---|---|---|---|
| Piccolo-27B | 83.41 | 82.76 | 72.53 | 80.95 | 98.91 | 90.96 |
| JDI-Decision V1 | 81.63 | 78.74 | 74.75 | 81.95 | 98.36 | 88.90 |
| WeMM-JEV-27B | 81.38 | 74.22 | 74.20 | 81.70 | 98.36 | 85.89 |
| RBDash v0.2 | 77.82 | 75.35 | 73.36 | 79.20 | 97.81 | 87.67 |
| Valen 1.5 (27B) | 78.14 | 85.82 | 70.87 | 80.95 | 97.81 | 93.29 |
| H2O Lightning 31B v1.3 | 80.82 | 57.79 | 66.98 | 86.97 | 97.27 | 86.98 |
| Task family | Decision focus | Pass rate |
|---|---|---|
| State tracking | Track changes to a described state | 63.43% |
| Argument validation | Check tool-call argument constraints | 100.00% |
| Budget allocation | Allocate limited resources under constraints | 15.38% |
| Project | Reported scope | Data and resources |
|---|---|---|
| Piccolo | Archived 27B exploration | 647 attempts; approximately 6,300 PPU-ZW810E accelerator-hours. |
| Mammoth-1 / Elephant-1 Kirkby et al. [2026] | Released 27B / 9B SFT recipes | 6.21M examples, one epoch; 72 GB300 / 128 H100 accelerators; 9,859 / 5,546 updates. |
| StartLux-Decision Zhuang et al. [2026] | Full SFT and adapter stages | Stage 1: 0.8M examples for 27B, 2.7M for 0.8B–9B; Stage 2: 1.7M records. |
| Drex Max Dhankhar et al. [2026] | SFT, low-rank RLAF, vision SFT | 1.15M, 180K, and 95K records; 9,000, 1,200, and 1,500 updates, respectively. |
Language Model Self-Improvement
Language model self-improvement studies how models use their own outputs, feedback, and experience to improve with less external supervision.
See all 195 papers