Riemann-Bench: Evaluating Moonshot Mathematical Reasoning
A private set of research-level problems designed to evaluate long-form theorem work beyond contest-style shortcuts.
Corecraft: Training Generalizable Agents in High-Fidelity Worlds
We show how dense professional simulations can improve tool use, planning, and transfer to unseen workplace tasks.
The Hierarchy of Agentic Capabilities
An empirical framework for measuring whether agents can plan, adapt, stay grounded, and recover from realistic ambiguity.
Multimodal Reward Models for Interleaved Text and Image
A benchmark for reward models that must judge image editing, interleaved generation, and multimodal reasoning together.
Rubric Reliability for Expert Human Preference Data
A study of how expert disagreement changes when rubrics ask evaluators to weigh correctness, usefulness, and judgment.
Measuring Factual Memory Under Adversarial Context
A controlled evaluation of whether models preserve facts when prompts include misleading dates, names, and citations.