barkup-bench

Researchers at barkup-bench ran seven pre-registered studies with over 13,000 model runs to determine the most reliable way for LLM agents to edit structured data. They found that giving every node a …

A new study from the barkup-bench series reveals that LLM agents suffer from session drift when serializing tree state only once, with per-step success dropping from 98.8% to 83.8% for Claude Sonnet 4…

A benchmark comparing whole-tree HTML rewrite against granular JSON tools for LLM-driven editing found that a defect in the benchmark harness had hidden small-model multi-turn tool reliability. After …