Research

A benchmark for whether text world models keep track of state, rules and side effects.

Video game speedrunning as a test of whether agents can beat strategies humans already know.

Human design workflows in Figma, captured as training data for creative design tasks.

A controlled benchmark that measures hallucination against reference world models.

A language model that stands in for an environment while an agent trains.

Teaching diffusion language models to fill in prompts, not only responses.

A two-agent benchmark for tip-of-the-tongue search across several turns.

54 kinds of reward exploit in code RL, and how well models catch them.

One definition of hallucination: inaccurate world modeling.

Environments that generate their own tasks, world dynamics and rewards.

Reducing score range bias when language models act as judges.

Testing whether agents keep track of facts spread across workplace tools.

A benchmark for finding the step where an agent went wrong.

573 tip-of-the-tongue questions that need search, reasoning and tool use.

A multi-modal benchmark at the frontier of human knowledge.

A small, explainable judge model for grading LLM interactions.

An open-source model for detecting hallucinations in retrieval-augmented generation.

Open-book financial question answering about publicly traded companies.

100 prompts for finding critical safety risks in language models.