MirroS

Research

Intelligence beyond the known world.

AgentGarten teaser showing code worlds rendered in diverse appearances and tasks, and agents that act on rendered views, write skills, and develop new strategies

AgentGarten:

Code Worlds for Evolving Agents

AgentGarten pairs code-based environments with a real-time neural renderer: code keeps the state and rules, the renderer shows the world, and agents improve from their own experience by writing playbooks across rounds of practice.

Linear spatial representation in S-Space, with horizontal, vertical, and distance directions encoding an object's coordinates

S-Space:

Exploring Spatial Workspace in Multimodal Models

S-Space is an internal spatial workspace in multimodal models that encodes continuous 3D object locations. By analyzing this mechanism, we gain insights into spatial reasoning.

Code-as-World teaser showing executable world representations and their applications to physical intelligence

Code as Worlds:

Agentic Discovery of Executable World Representations for Physical Reasoning

Code-as-World represents physical worlds as executable code and discovers them through an agentic propose–execute–render–verify loop, turning multimodal observations into quantitatively grounded supervision for physical reasoning.

HarnessEval-W teaser showing its data construction and agentic evaluation pipeline

HarnessEval-W:

Agentifying the Evaluation of Visual Worlds

HarnessEval-W turns world-model evaluation into an agentic workflow: specialized agents inspect each case and return a score backed by a clear, verifiable evidence trail.