Benchmark Radar Day 91: Static Chart Scores Miss What the Chart Hides

Day 91 of Benchmark Radar. The scan fetched 1,310 items, published 466, and marked 163 as recommended.

Author: Koutian Wu; GitHub: ktwu01

Benchmark Radar is live. Open the dashboard, star the repository, or read the paper. The Hugging Face dataset provides the data.

ChartAct tests dynamic chart understanding through real browser interactions. Models and graphical user interface agents must hover, click, zoom, drag, scroll, and select controls to answer questions, rather than interpret only static chart images. What it measures is whether a model can obtain information hidden behind interactive controls. chartact-benchmark/ChartAct.

LongHaul-Bench evaluates a self-improving agent across roughly 1,000 sequential episodes while preserving its evolving knowledge state and enforcing a fixed experience-memory budget. It compares this agent with a frozen control across five reproducibly generated industrial environments. The question is whether accumulated experience improves or degrades reliability over time. LongHaul-Bench: A Reproducible Benchmark for Long-Horizon Reliability of Self-Improving LLM Agents Under Industrial Edge Constraints.

Why this matters

A static visual-question benchmark cannot measure behavior that only appears through interaction. And for agents that keep learning after deployment, one session’s accuracy cannot show whether experience compounds into reliability or into drift.

Scan coverage

This is a keyword-filtered discovery feed, not a representative survey of AI research. The archived briefing used 136 of 466 captured evidence records. These findings summarize its cited releases; discovery dates do not necessarily equal original publication dates.

Data and briefing: 2026-10-25 archived snapshot.