OmniHarness: Harnessing Generalizable Visual Generation via Symbolic Policy Learning
OmniHarness turns successful visual-generation runs into reusable symbolic policies, then adapts them to new tasks without changing model weights.
The paper says the system abstracts verified executions into policies that capture shared steps and when they apply, stripping out task-specific inputs. It adds intermediate checks so failures can be corrected during a run, not only after completion. OmniHarness also creates its own practice tasks near its limits before downstream tasks are given, using feedback to refine the policy library. In tests across six benchmarks, three MLLM backbones, and three visual agent frameworks, it reports a 95.0% resolve rate on ComfyBench Creative tasks, 27.5 points above the strongest baseline. Source: HF Daily Papers' note.
The paper says the system abstracts verified executions into policies that capture shared steps and when they apply, stripping out task-specific inputs. It adds intermediate checks so failures can be corrected during a run, not only after completion. OmniHarness also creates its own practice tasks near its limits before downstream tasks are given, using feedback to refine the policy library. In tests across six benchmarks, three MLLM backbones, and three visual agent frameworks, it reports a 95.0% resolve rate on ComfyBench Creative tasks, 27.5 points above the strongest baseline. Source: HF Daily Papers' note.
score 4