Data World: Scaling Data Environments from Scratch
We introduce Data World, a multi-agent engine that synthesizes environments, tasks, and rubrics from explicit mechanisms.
ProofPaper ↗
Key points
- Data agents use code to process heterogeneous files and derive analytical insights.
- Scaling their training faces costly environment curation, privacy and intellectual-property restrictions, and difficulty verifying open-ended findings without known generative truth.
- Human preference-based curation for question realism selects 100 tasks across 50 held-out worlds and 16 domains for DataWorld-Bench.
- RL-zero training on the synthesized tasks produces DataWorld-35B-A3B from Qwen3.6-35B-A3B without trajectory distillation.
Sources (2)
- [1]Data World: Scaling Data Environments from ScratcharXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 10:35 AM
We introduce Data World, a multi-agent engine that synthesizes environments, tasks, and rubrics from explicit mechanisms.
Data agents use code to process heterogeneous files and derive analytical insights.
- [2]Data World: Scaling Data Environments from ScratchHugging Face Daily Papers · Oct 15, 12:00 AM · same content
Extractive summary: sentences quoted from the sources.
Before this
- Oct 14, 2026Harness Policy Distillation: Aligning Agent Policies with Runtime Semantics
- Oct 14, 2026RM-OPD: On-Policy Distillation for Generative Reward Models
- Oct 14, 2026BluePixel: Toward Native 4K Agentic Image Editing
- Oct 14, 2026Who's Behind the Text? Robust Model Authorship Attribution
- Oct 14, 2026Beyond Corrected Actions: Training LLM Agents on Textual Critiques
- Oct 13, 2026Run fast, private agents on-device with LFM2.5-24B-A2B
