Method lets streaming models retract a guess when audio contradicts it

single source· 1 articles · confidence: high · first seen 2026-09-13 20:00 UTC

What this means for you

Nothing to ship yet — this is a paper, with no released code, weights or licence described. The reusable part is the failure mode: if you serve a streaming audio-video model, late contradictions are the case your evals likely miss.

A paper reports a fix for streaming audio-video models that commit to a visual guess before the sound arrives. The authors call this premature cross-modal commitment: the wrong reading enters memory and later reasoning repeats it after audio contradicts it. Their method, Omni-Streaming Thinking, keeps audio and visual evidence separate and marks each claim pending until the segment covering it has arrived. On a frozen Qwen3-Omni-30B-A3B-Instruct backbone it beats the strongest open baselines by more than 10% relative on five streaming and audio-visual benchmarks, the authors report. On their new diagnostic set it reports d-prime 2.95 against at most 1.38, with no evaluation date given.

Key facts

  • ·Paper arXiv 2609.15128 was published on 13 September 2026. source
  • ·The method, Omni-Streaming Thinking, uses a frozen Qwen3-Omni-30B-A3B-Instruct backbone with lightweight adaptation. source
  • ·It reports more than 10% relative average improvement over the strongest open baselines on five streaming and audio-visual benchmarks. source
  • ·Each claim is marked pending and linked to a future verification interval, checked against the evidence from the specified modality at the end of that interval. source
  • ·The new OST-DiagBench holds video fixed and edits audio to test agreement, absence, contradiction, coexistence and subtitle-speech conflict. source
  • ·On OST-DiagBench it reports d-prime of 2.95 against at most 1.38 for open baselines, while reducing vision-induced auditory hallucinations. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire