Charts used as internal steps improve agent data analysis
Original title: Charts as Epistemic Actions for Experience-Driven Agentic Data Analysis
A 200-task probe shows agents read self-made charts as well as numbers but often skip plotting.
The authors test whether charts can act as internal steps inside a data agent's reasoning loop rather than only as outputs. A 200-task suite under four interface conditions shows agents score as well from self-generated charts as from numbers, yet plotting access alone does not produce consistent use. Charts help temporal, pattern, and anomaly questions, while text helps precise reading and conditional checks. A trajectory-distilled method, without parameter updates, raises accuracy and visual-evidence quality on three held-out benchmarks with four LLM backbones.
- Experience covers both which operations to verify and when to use charts.
- A trajectory judge and validation gate admit only helpful experience.
- Backbones that never drew charts begin using them.
- Model parameters are not updated.
From the original article
1 min read at arxiv.orgData agents built on ReAct-style loops now execute complex analytical workflows over structured data, yet their intermediate reasoning still runs almost entirely through text. Human analysts instead move between numbers, tables, and charts, reading visual evidence to expose trends and anomalies that text summaries drop.
By Xinzhe Wang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.