OmniHarness: Harnessing Generalizable Visual Generation via Symbolic Policy Learning

OmniHarness. Harnessing Generalizable Visual Generation via Symbolic Policy Learning. By Xu Xu and seven others, from Beihang University, the Chinese University of Hong Kong, and one other institution. Published on the thirteenth of September. Its code repository is public and real: about twenty source files.

Unified multimodal large language models, and multi-agent systems built on them, have pushed visual generation forward. Three limitations remain. First, existing methods often distill experience that is specific to one task, and it generalizes poorly. Second, reflection is usually deferred until a task is complete, so a mistake early on is only noticed at the end. Third, knowledge is typically acquired only in response to what a downstream task demands. The system learns when asked, and not before.

OmniHarness is a framework for generalizable visual generation through symbolic policy learning. Verified executions are abstracted into symbolic policies for families of visual generation tasks. A policy captures the shared procedure, together with the conditions under which it applies, and removes the inputs specific to any one instance. It is a recipe rather than a record of one meal. The harness then instantiates, adapts and composes these policies for new tasks.

Each of the three limitations gets a direct answer. Against narrow experience, the policies are abstracted to cover a task family. Against late reflection, intermediate verification guides refinement and failure recovery during execution, not after it. Against reactive learning, self-directed inquiry has OmniHarness generate and run its own practice tasks near the limits of its capability, before any downstream objective is specified. Throughout, execution feedback keeps refining the policies while the model parameters stay fixed.

Experiments span six benchmarks, three MLLM backbones, and three visual agent frameworks, and show strong performance along with continual expansion of capability. On ComfyBench's Creative tasks, OmniHarness reaches a ninety-five percent resolve rate, twenty-seven point five percentage points above the strongest baseline. Frozen snapshots of the policies also improve existing visual agent systems through plug-and-play reuse. The policies, in other words, are portable. They carry what was learned into systems that never did the learning.

Learning here lives in the policies, not in the weights. That is what allows practice to happen before any task arrives, and what lets a frozen snapshot be lifted out of one system and dropped into another.