OmniReasoning: Pushing the Limits of Audio-Visual Joint Reasoning

Name: OmniReasoning: Pushing the Limits of Audio-Visual Joint Reasoning
Uploaded: 2026-10-04T23:05:57.455329+00:00
Duration: 581 s
Description: The research introduced OmniReasoningBench to test genuine audio-visual joint reasoning, OmniQA to create evidence-grounded training data with timestamped clues, and ModalityFactored Self-Distillation (MFSD) as a learning method. This approach explicitly models cross-modal interactions between audio and visual evidence, leading to significant performance gains on complex reasoning tasks.

The research introduced OmniReasoningBench to test genuine audio-visual joint reasoning, OmniQA to create evidence-grounded training data with timestamped clues, and ModalityFactored Self-Distillation (MFSD) as a learning method. This approach explicitly models cross-modal interactions between audio and visual evidence, leading to significant performance gains on complex reasoning tasks.

Open the full episode