Data World & MRTPO: Synthesizing 1,037 Causal Data Worlds Boosts Qwen3.6-35B-A3B by +20.9 Points on DABstep-Hard via RL-Zero
Training data analysis agents is bottlenecked by privacy restrictions and unverifiable open-ended insights. Researchers introduce Data World (arXiv:2610.21656), a multi-agent engine that specifies underlying causal world mechanisms before generating heterogeneous files and verifiable rubrics across 1,037 disjoint worlds. Pure RL-zero training on Qwen3.6-35B-A3B yields DataWorld-35B-A3B, jumping +20.90 points on DABstep-Hard and +14.28 on DAComp. Alongside MRTPO (arXiv:2610.21439) for token-level multi-reward credit assignment, these works unlock scalable RL for data and SQL agents.
Key Takeaways
- World-before-files synthesis: Data World constructs 1,037 disjoint causal worlds and 2,135 verified tasks with guaranteed target recoverability
- Massive RL-zero gains on Qwen3.6-35B-A3B: DataWorld-35B-A3B jumps +20.90 points on DABstep-Hard, +14.28 on DAComp, and +12.13 on JobBench without distillation
- Token-level multi-reward credit via MRTPO: uses per-token Jensen-Shannon divergence to route each verifier's reward strictly to its relevant tokens
Linked Models & Agents:
⚡ DataWorld-35B-A3B⚡ Qwen3.6-35B-A3B🤖 Data World🤖 MRTPO🤖 Data Analysis Agents🤖 SQL Agents🌐 Alibaba Qwen