Long‑horizon multimodal agents rely on specialized capabilities such as perception, retrieval, reasoning, verification, and execution. Existing designs usually activate a fixed set of capabilities or follow a predefined workflow, which incurs high computational cost and cannot adapt to stage‑dependent capability demands. This work studies the combinatorial capability allocation problem: at each interaction step the system must select a cost‑sensitive subset of capabilities. The value of a capability depends on the already selected subset, and previous allocations affect the states encountered later, making the decision process highly interdependent. We introduce CoCA (Combinatorial Capability Allocation), an on‑policy learning framework. On states visited by the student policy, a stronger teacher compares the marginal net values $\Delta v_i$ of candidate capabilities conditioned on the current subset. A conditional utility model then converts these comparisons into an autoregressive capability‑subset policy, avoiding explicit enumeration of all subsets. To mitigate distribution mismatch between teacher and student, CoCA employs dual‑level on‑policy distillation: one level aligns across environment states, the other aligns within partial subsets encountered during set construction. Finally, trajectory‑level reinforcement learning refines the distilled policy toward higher task success, lower activation cost, and greater allocation stability. At inference time, allocation is performed solely by the lightweight student policy without teacher queries or online updates. Experiments on long‑horizon multimodal environments and controlled capability‑demand shifts demonstrate that CoCA consistently outperforms state‑of‑the‑art baselines.
Review