What it is
A recursive self-improvement framework for multimodal agents that organizes tasks by shared capabilities, builds simulators from them, and evolves reusable agent modules. Instead of grouping tasks by modality or surface form, it reorganizes them by shared capability needs, then co-evolves simulated worlds and reusable harness modules along a capability tree.
Key results
- +34 mean points on 28 training benchmarks
- +26 mean points on 18 unseen benchmarks
- Gains include 2 real-world embodied tasks
Shared capability structure can drive simulator creation and modular agent evolution that transfers across heterogeneous multimodal tasks.