Contributed to visual coding agent & harnesses and AI-for-AI.

Contributed to visual coding agent.

Contributed to general tool-use capacity.

A live workflow-agent benchmark with refreshable demand signals and verifiable execution traces; 105 tasks across 22 categories, 13 frontier models, top model passes only 66.7%.

BlenderAgent converts an input image into structured 3D scene code, executes it in Blender, and evaluates the rendered reconstruction using geometric, spatial, and appearance metrics.

An on-policy data-evolution framework for visual-native multimodal deep-search agents, using agent rollouts to evolve training data and improve search behavior.

A phone-agent research line spanning open-model training, real-phone execution, scalable mock-app environments, and verifier-backed rollouts.

Co-scales phone-agent training data and executable environments to improve efficient GUI-agent training and evaluation.

A game-generation benchmark for coding agents, with 140 Godot tasks and interactive verification of playable end-to-end game artifacts.

An open canvas-native harness where creative agents plan, generate, revise, and organize long-horizon multimodal projects in an inspectable shared workspace.

A mixed-action phone-agent harness and benchmark that routes across CLI, GUI, and MCP tools with trace-backed verification.

An artistic editing agent system that plans multi-step retouching commands and coordinates expert models for execution-quality image refinement.

Maps visual perception into code-like structured summaries, enabling code-agent-style reasoning and verification for video understanding.

A multi-turn segmentation agent trained with expert-curated trajectories, end-to-end outcome verification, and process rewards for adaptive iterative refinement.