Data World & MRTPO: Synthesizing 1,037 Causal Data Worlds Boosts Qwen3.6-35B-A3B by +20.9 Points on DABstep-Hard via RL-Zero

Background and the Problem

Data agents write code to analyze heterogeneous files and extract actionable insights, but scaling their RL training is stymied by data privacy laws, expensive environment curation, and the impossibility of verifying open-ended analytical conclusions when the underlying ground-truth data generation mechanism is unknown.

Architecture and How It Works

Two complementary frameworks—Data World (arXiv:2610.21656) and MRTPO (arXiv:2610.21439)—unlock scalable environment synthesis and token-level multi-reward RL:

1. World-First Data Environment Synthesis (Data World): Specifies the explicit causal and mathematical world behind the files before rendering the heterogeneous files themselves, pairing independent recoverability audits with deterministic rubrics to generate 1,037 disjoint training worlds (2,135 tasks) and the DataWorld-Bench benchmark;
2. Reward-Specific Token-Level Credit Assignment (MRTPO): In multi-reward RL (such as Text-to-SQL and tool calling), replaces uniform sequence-wide advantage broadcasting by measuring the per-token Jensen-Shannon distance when replaying responses with and without each verifier's feedback.

Benchmarks and Measured Results

Empirical results across data analysis and tool-use benchmarks demonstrate transformative gains:

• +20.90 Points on DABstep-Hard via Pure RL-Zero: Training Qwen3.6-35B-A3B with pure RL-zero on Data World (zero teacher distillation) yields DataWorld-35B-A3B, improving over the base model by +20.90 points on DABstep-Hard, +14.28 on DAComp, +12.13 on JobBench, and +7.06 on DataWorld-Bench;
• Faster Convergence on Text-to-SQL & Tool Calling: MRTPO consistently outperforms GDPO across text-to-SQL, multi-constraint instruction following, and tool-calling tasks.

Getting Started for Developers

Both frameworks are open-sourced at https://github.com/Zhanli-Li/Data-World and https://github.com/zhaihaotian/MRTPO:

• Synthesize Privacy-Safe Enterprise Data Sandboxes: Use Data World to generate verifiable multi-file analytics environments grounded in known causal graphs;
• Pinpoint Multi-Reward Credit: Integrate MRTPO into RLVR pipelines where format verifiers and execution verifiers operate simultaneously.