The Trinity:Capability-Driven Recursive Self-Improvement for Multimodal Agents

The Trinity. Capability-Driven Recursive Self-Improvement for Multimodal Agents. By Weidong Tang and nineteen others, from InfRec Cardinal AI Lab. Published on the twenty-sixth of October.

Multimodal intelligence can be viewed as human-guided co-evolution between worlds and agents, through interaction. Scaling that process runs into two bottlenecks. Constructing heterogeneous, verifiable environments is costly. And iterating on the Harness by hand cannot keep pace as tasks grow in diversity and complexity. Both bottlenecks grow with the tasks themselves. If every new kind of task calls for its own environment and its own hand-tuned Harness, the effort rises with every task added.

The key insight is that tasks with different modalities and surface forms can share the same underlying capability requirements. If that holds, tasks no longer have to be handled one surface form at a time. A capability shared by many tasks can be practised once, in a world built for it, and the improvement carried back to all of them. The Trinity builds on that. It is a recursive self-improvement framework that couples world generation with Harness evolution, through three roles: Creator, Explorer and Evolver. Each takes one part of the loop. Evolver organises the tasks, Creator builds the worlds, and Explorer evolves the Harness.

Evolver works from verified execution trajectories. It shifts the basis of task organization from surface categories to shared requirements, reorganizing the original task tree into a hierarchical capability tree. Creator carries out what the authors call Capability-Driven Real2Sim2Real. For heterogeneous tasks at the same node, it abstracts their common requirements into simplified, verifiable simulated worlds where the capability can be learned, while evaluation on the original tasks measures transfer.

Explorer moves along the capability tree, progressively evolving reusable capability modules through selective inheritance. The result is a Harness tree, whose validated updates accumulate across iterations within a fixed topology. Because the resulting Harnesses specialize in different capabilities, a Capability-Aware Harness Router uses public task information to select compatible module configurations from the tree.

The evolved Harnesses improve mean performance by thirty-four points across twenty-eight training benchmarks. On eighteen unseen benchmarks, including two real-world embodied tasks, the improvement is twenty-six points. The second figure is the one that speaks to transfer. Those benchmarks are unseen, so gains there suggest that what was learned along the capability tree carries beyond the tasks it was built from.

The organizing move is to stop sorting tasks by how they look. Once tasks are grouped by the capability they demand, one simplified world can train for many of them, and one evolved module can be inherited wherever that capability recurs.