Visual and Spatial Intelligence for Multimodal Agents

We study how multimodal agents represent, reason about, and act in visual and spatial environments. IR3D-Bench explores scene reconstruction through Blender tool use, while Struct-to-Reason improves video understanding with structured summaries. DynamicVerse builds physically grounded 4D scene data from real videos, and Thinking in Dynamics studies dynamic perception and reasoning. RobotWorld examines how these capabilities support robot use across tasks and embodiments.

Agent Foundation Models for Coding and Tool Use

Seed2.0 and Seed1.8 combine multimodal understanding, coding, and tool use for complex real-world workflows. UI-TARS-2 focuses on GUI-centered interaction, augmented with terminal and MCP tools.

Recursive Self-improvement (RSI) for Agent Evaluation and Data Synthesis

Claw-Eval (Live) and On-policy Data Evolution study real-world workflow evaluation and agent data self-improvement: verifiable live tasks in Claw-Eval, and rollout-driven training-data evolution in On-policy Data Evolution.

Computer and Phone-use Agents: Training, Harness, Environments

PhoneBuddy, HyMobileAgent, PhoneWorld, and PhoneHarness advance phone-use agents through open-model training, data-environment co-scaling, scalable interactive environments, and mixed CLI/GUI/MCP execution. PhoneSafety and PhonePrivacy complement this line with capability-aware safety and privacy evaluation.