MirroS
MirroS is a research organization building Physical Recursive Self-Improvement (Physical RSI): intelligent systems that turn encounters with the unknown into verified, reusable capabilities. Its approach mirrors reality into an evolving agentic world model that can be executed, tested, revised, and improved, with the world model and agent co-evolving through interaction and feedback.
At a Glance
- AI and multimodal-model researchers
- World-model and video-generation developers
- Robotics and embodied-intelligence teams
- Researchers studying physical reasoning, spatial reasoning, and continual learning
- +1 more
AI Tools by MirroS
(1)
HarnessEval-W
Visual World Model Evaluation
Discussions
No discussions yet
Be the first to start a discussion about MirroS
Latest News
AgentGarten: Code Worlds for Evolving Agents; technical report and neural-renderer training/inference code released
S-Space: Exploring Spatial Workspace in Multimodal Models
Representing the Physical World through Structured Language / Code-as-World release
HarnessEval: The Era of Harness for Benchmarking / HarnessEval-W release
Products & Services
A code-world environment paired with a real-time neural renderer: code controls world state and rules, the renderer produces visual observations, and agents interact, review experience, write playbooks, and improve across rounds.
A research system/concept for an internal spatial workspace in multimodal models. It studies how continuous 3D object locations and abstract concepts are encoded, persist across contexts, and can be causally manipulated.
An executable representation of physical worlds discovered through an agentic propose-execute-render-verify loop, turning multimodal observations into grounded supervision for physical reasoning.
An open benchmark and executable agentic evaluation system for interactive visual/world models. It dynamically plans evaluations, selects specialized skills and sub-agents, gathers visual evidence, and produces inspectable reasoning traces and scores.
Market Position
MirroS positions itself around Physical RSI and executable, verifiable world models rather than only pixel prediction. Its differentiation is the combination of structured-language/code representations, agentic interaction and improvement loops, and evaluation backed by inspectable evidence trees; HarnessEval-W explicitly contrasts this with fixed-rubric world-model benchmarks and the lab's research contrasts programmable code worlds with visually rich but less directly editable video-generation models.
Founding Story
MirroS's stated starting vision is to move beyond intelligence optimized within a closed world. Its thesis argues that every out-of-distribution surprise should reveal an environment or skill gap, be converted into a reproducible learning experience, and drive improvement of both the world model and the acting system.
Business Model
Revenue Model
MirroS presents itself as a research organization and open research/code lab. The reviewed official pages describe open technical reports, project pages, repositories, benchmarks, and public datasets rather than a commercial pricing or subscription model.
Target Markets
- AI and multimodal-model researchers
- World-model and video-generation developers
- Robotics and embodied-intelligence teams
- Researchers studying physical reasoning, spatial reasoning, and continual learning
- Benchmarking and evaluation practitioners
- Physical reasoning and prediction
- World-model training and evaluation
- Robotics and embodied intelligence
- Video/world-model generation with controllable state and interactions
- Navigation, object manipulation, tool use, and multi-agent interaction
- Continual learning from experience and recursive self-improvement