AI

OmniReasoning: Pushing the Limits of Audio-Visual Joint Reasoning

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 06/10/2026, 10:18:24 UTC.

Briefings

A new benchmark and training pipeline named OmniReasoning enable audio-visual joint reasoning by requiring both modalities for tasks. The system includes OmniQA for evidence-grounded data generation and Modality-Factored Self-Distillation for credit assignment. Results show OmniReasoning-30B-A3B improved OmniVideoBench by 50.0% and OmniReasoningBench by 42.5% compared to the base Qwen3 model.

Key points

• The OmniReasoning benchmark requires both audio and visual evidence to answer questions, unlike previous benchmarks that treat modalities independently.
• The model OmniReasoning-30B-A3B achieved 50.0% accuracy on OmniVideoBench and 42.5% on OmniReasoningBench, outperforming the base Qwen3 model by 12.8 and 9.3 percentage points respectively.

Context

This research addresses the gap where existing benchmarks and training methods treat audio, vision, and language modalities independently, leaving joint reasoning capability poorly evaluated. The findings provide a solid step for facilitating future research in omni-modal joint reasoning.

Limitations

The evidence covers only the development of the OmniReasoning benchmark, data engine, and learning method. Specific evaluation results on other general or long-video benchmarks are mentioned as substantial gains but no specific metrics are provided for those tasks.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 19/10/2026, 19:42:50 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeifn4jdfkbapgpyzayhuheim4wtfbctgubkux2e75wk4ow6tkeru3i