The authors built ProbeBench, a suite of 200 tasks synthesized from InsightBench, with four matched interface conditions: numeric output, self-generated charts, a mixture, and an oracle.
Teaching data agents to use charts inside their reasoning loop improves analysis
ProbeBench exposes a gap between chart-reading ability and chart use, while ChartMind learns when to select visual or textual evidence, yielding consistent gains.
Chinese Tech
Xinzhe Wang · Xudong Weng · Zang Li · Shaolei Zhang
Tencent · Renmin University of China
Research Digest··2 min read
The authors probe how data agents use charts during reasoning, find a persistent gap between chart-reading capability and actual chart use, and propose ChartMind, a method that distills chart-selection behavior into a reusable, frozen experience package.
§