Reproduce “OmniReasoning: Pushing the Limits of Audio-Visual Joint Reasoning” 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

Junming Lin, Yuxuan Wang, Zhenxin Lei, Yuxin Liu, Ruixun Liu, Yinsong Yan, Ling Wang, Minghao Han and 6 more

The authors introduce OmniReasoningBench, a benchmark for audio-visual reasoning, and MFSD, a training method that assigns token-level credit using audio, visual, and joint clue signals. The benchmark tests reasoning over and beyond video, while MFSD measures interaction between clue types. OmniReasoning-30B-A3B reaches 42.5% accuracy, above RLSD and GRPO under matched training budgets, but below Gemini-3.7-Flash’s 57.6%.

Summary by alphaXiv.

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