Omni-Streaming Thinking
The paper targets a streaming failure where early visual guesses keep steering the model after audio proves them wrong.
OST keeps evidence, forecasts, and claims structured, with each claim left pending until a later verification window. It stores audio and visual evidence separately, then checks claims against the relevant modality before allowing them to drive an answer. When contradiction appears, the method weakens the claim and its dependent states before updating with the new evidence. The authors report more than 10% average relative gains over open baselines across five benchmarks, plus stronger results on their OST-DiagBench test for audio-visual conflicts. HF Daily Papers' note
OST keeps evidence, forecasts, and claims structured, with each claim left pending until a later verification window. It stores audio and visual evidence separately, then checks claims against the relevant modality before allowing them to drive an answer. When contradiction appears, the method weakens the claim and its dependent states before updating with the new evidence. The authors report more than 10% average relative gains over open baselines across five benchmarks, plus stronger results on their OST-DiagBench test for audio-visual conflicts. HF Daily Papers' note
score 4