Selected publications. For the full list, see Google Scholar.
Are agents ready for real enterprise work? A large, realistic stateful environment for long-horizon planning and tool execution under real enterprise constraints, now tracked on the Artificial Analysis leaderboard.
Whether enterprise agents need learned world models, and why reading the live configuration is critical for inferring how business rules cascade through a system.
Terminal agents with direct API access match or outperform MCP and GUI agents: the same accuracy as web agents at 5–6× lower cost.
GroundCUA, a large human-verified dataset of desktop UI elements, and the GroundNext models that reach state-of-the-art grounding with a fraction of the training data.
Maps vision features to a weighted average of the LLM's own text embeddings, so visual tokens stay in a space the language model can read.
Renders SVG rollouts, compares them to the input image, and uses that visual feedback as the RL reward.
Generates editable SVG code with a code-pretrained language model, writing compact primitives instead of traced paths; introduces SVG-Stack (2.1M SVGs) and SVG-Bench.
BigDocs-7.5M, an open, license-checked dataset of 7.5 million documents across 30 tasks, and BigDocs-Bench, ten structured-output tasks such as screenshot-to-HTML.
How well can models understand and build websites? A real-world, multilingual benchmark collected with 117 expert annotators.
A state-of-the-art chart reasoning model, with a new dataset and a visual reinforcement fine-tuning recipe.
Specify tasks with vision or language prompts, ground them in the domain's dynamics, and learn the behaviors in imagination, with strong multi-task generalization.
A model-based agent that discovers skills in its world model and adapts them by evaluating many in parallel, in imagination.
Linearizes nonlinear dynamics in a latent space so long-range prediction can be parallelized, while accounting for the agent's actions.
Unsupervised model-based pre-training plus a hybrid Dyna-MPC planner for fast adaptation to downstream tasks.
Makes large pretrained models consistent under input transformations, for better sample efficiency and robustness.
Uses surprise from mismatches in touch feedback to drive exploration in hard, sparse-reward robot tasks.
Brings multi-label inductive biases to single-label datasets using iterated learning, inspired by language emergence.
Position-conditioned and latent-conditioned policy learning that lets world-model agents solve object-positioning tasks.
What it takes for zero-shot LLM agents to perform novel tasks on HTML-based interfaces, and where they fall short.