Many agent-improvement loops treat each benchmark as a separate problem, even when different tasks depend on the same underlying skill. The central insight here is to organize learning around those shared capabilities: build simplified worlds to practice them, then test whether the resulting modules transfer to original tasks.
Key Findings
- Three roles divide the loop: Creator builds verifiable simulated worlds, Explorer collects execution trajectories and refines modules, and Evolver maps tasks into a hierarchical capability tree. This makes the improvement process reusable across tasks with different modalities and surface forms.
- Harness modules are selectively inherited along a fixed-topology tree, while a router chooses a compatible configuration using public task information. The design aims to accumulate validated changes without turning every task into a bespoke agent.
- The paper reports average gains of 34 points across 28 training benchmarks and 26 points across 18 unseen benchmarks. Iterative improvement continues through four rounds, and evaluations include two embodied tasks in simulation and on physical robots.
Method and Evidence
Capability-Driven Real2Sim2Real uses simplified, verifiable simulations for learning and evaluates transfer on original tasks. The study holds the backbone and capability-tree structure fixed during recursive improvement, helping isolate Harness-level iteration. The embodied results are encouraging but evaluate complete Harness–skill configurations: task-specific skill preparation is part of the setup, so the experiments do not isolate the contribution of module evolution alone.
Who Should Read It
Great fit for researchers designing multimodal agents, reusable agent architectures, and benchmark-based self-improvement loops. Look elsewhere for evidence that an agent can improve its model weights autonomously: this work focuses primarily on Harness-level changes, and its transfer claims are bounded by the selected benchmarks, fixed backbone, and evaluation protocols.