Reproduce “Beyond Modality Preference: Controlling Multimodal Arbitration under Modality Conflict in MLLMs” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zhaoying Pan, Shenyu Lu, Jun Hui Wong, Junyi Chai, Xiaoqian Wang

The authors identify attention heads and steer them at inference by increasing attention to a task-designated modality while retaining image and text inputs. On Qwen3-VL-8B MC², image-primary steering reaches 79.50% image-reference accuracy and 5.25% text-reference accuracy, compared with 59.95% and 27.60% for MPS. The priority is supplied externally; the method does not determine which source is trustworthy.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.