Agent-R²: 4B Productivity Agent Beats Qwen3-235B on Real-Work Benchmarks via Real-Source Data Synthesis
Summary
Agent-R² demonstrates that a 4B-param agent trained on data synthesized from actual web-corpus work products outperforms Qwen3-235B-Thinking on two professional-task benchmarks, making a concrete case that training-data grounding matters more than parameter count for long-horizon productivity work.
Shared on Bluesky by 1 AI expert
Originally reported by paper
Read the original article →Original headline: Agent-R²: 4B Productivity Agent Beats Qwen3-235B on Real-Work Benchmarks via Real-Source Data Synthesis