The Trinity: Recursive Self-Improvement Built Around Capabilities

Multimodal agents face a scaling problem that is not solved simply by adding more tasks. Building heterogeneous environments with reliable verification is expensive, while manually redesigning an agent Harness becomes increasingly difficult as tasks diversify. The Trinity addresses this problem by organizing improvement around capabilities shared across tasks.

Replacing task categories with a capability tree

Tasks that look unrelated on the surface may require the same underlying skills: planning, state tracking, visual grounding, tool use, or multi-step execution. The Trinity uses verified execution trajectories to identify such common requirements and reorganizes the original task tree into a hierarchical capability tree. The objective is therefore not only to solve an individual benchmark, but to accumulate modules that can be reused across different tasks and modalities.

The framework assigns three roles to the improvement process:

• Creator abstracts shared requirements and builds simplified, verifiable simulated worlds.
• Explorer searches those worlds for useful strategies and module combinations.
• Evolver selects validated updates and inherits them through later iterations.

Capability-Driven Real2Sim2Real

The central idea is Capability-Driven Real2Sim2Real. Instead of reproducing every real-world task as a separate training environment, Creator extracts the common requirement of heterogeneous tasks located at the same capability node. It then constructs a more controllable simulated world where the capability can be learned and tested repeatedly. Performance is finally measured on the original tasks to examine whether the learned ability transfers across different surface forms.

This approach aims to separate capability learning from the cost and complexity of the original environment. It also offers a way to generate training worlds around reusable requirements, rather than around fixed benchmark labels.

Evolving a tree of Harnesses

Explorer progressively develops reusable capability modules through selective inheritance. The resulting Harness tree contains branches specialized for different capabilities, while validated updates accumulate over iterations within a fixed topology. The paper further introduces a Capability-Aware Harness Router. Using information publicly available in a task description, the router chooses a compatible configuration of modules instead of forcing every task through one universal Harness.

The reported outcome is substantial: The Trinity is described as gaining an average of 26 points across 18 unseen benchmarks, including real-world robotics settings. This suggests that capability-centered task organization may improve reuse when an agent must operate across different modalities and environments.

The evaluation should nevertheless be read carefully. The supplied material indicates that Section 4.1 performs five preparation trials for each benchmark and Harness pair, converts the resulting trajectories into skills, and then scores the Train and Unseen settings. Consequently, the reported gain combines cross-task transfer with adaptation to the target benchmark. The material does not provide a no-preparation comparison for the 18 unseen benchmarks, nor does it report the token or compute cost of those trials.

The broader contribution of The Trinity is its closed-loop view of agent improvement: execution traces reveal shared capabilities, Creator turns those capabilities into verifiable worlds, and Explorer and Evolver accumulate reusable Harness updates. For multimodal and embodied agents, this capability-oriented structure could become a practical route toward scaling environments and modular self-improvement, provided that future evaluations isolate transfer from target-specific preparation.

Source: Hugging Face Daily Papers