See the worlds.
Think ahead. Act.

A language-first video learner that converts frames into structured image-and-language descriptors for few-shot video tasks.

Diagnosed the action-knowledge gap in video-language models and patched temporal action dynamics into frozen foundation models.

Introduced Primal Visual Description, a symbolic bridge from precise vector perception to language reasoning.

Dynamically merges—and virtually restores—visual tokens to accelerate VLM inference without permanently discarding information.

An early multi-persona reasoning framework that turns one LLM into a team of specialized, self-collaborating reasoners.

Introduced perception-aware policy optimization, reducing multimodal perception errors while learning to reason.

Grounds natural-language perspective descriptions into camera poses and imagined target views for multi-view spatial reasoning.

Introduces a large affordance knowledge base and 14K tasks to test grounded creative tool repurposing in agents.

Advances creative physical intelligence through visually grounded exploration and affordance-grounded alignment.

Built a hierarchical mobile agent that improves through persistent tips, shortcuts, and past interaction experience.

Enables computer-use agents to discover, verify, and compose professional GUI skills through environment exploration.

Internalizes multimodal policies into conversational agents so interaction skills become native model behavior.

A comprehensive benchmark for collaboration and competition across diverse, interactive multi-agent scenarios.