DATA

Multi-step tasks across journalism, trade, and education scored by expert-designed rubrics.

Extends GDPval+ by burying input files in a stack of realistic noise like superseded drafts, outdated guidance, adjacent reports. Tests whether a model can find the signal before it starts writing.

An RL environment for customer conversations. PathPrint marks the decision points in every conversation an AI agent handles across customer service scenarios, then merges them into a map of where models succeed, stall, or fail.

Seven agentic coding environments for evaluating frontier language model ability to resolve bugs and feature requests