Evidence Ledger Boosts Multimodal Reasoning by Up to 15%

Adaptive Multi-modal Agentic Reasoning replaces the full dialogue transcript with a compact, query-driven evidence ledger that the agent re-injects on demand. OmniSeek scored 70.4% on the MMOU benchmark versus the base Qwen3-Omni-Instruct, a 16.3% absolute gain, and 44.2% on LVOmni (+8.4%).

• Evidence ledger swaps the full transcript for a query-driven list of observations
• MMOU: OmniSeek hits 70.4%, a +16.3% gain over base Qwen3-Omni-Instruct
• LVOmni: 44.2%, up +8.4% over the base model
• Gains shown on long-form video and rely on a two-stage RL pipeline