Omni-Decision: Evidence-Ledger Planning for Omni-Modal Agents
The paper argues that planning, not perception, is the main failure point for omni-modal agents.
Omni-Decision replaces a growing conversation history with an evidence ledger tracking missing, confirmed, and conflicting evidence. A critic filters noisy observations before they reach the planner, keeping the working context compact. The authors report 81.4% accuracy on OmniGAIA at about 43% of Gemini-3.1-Pro’s cost per question, plus 65.0% on WorldSense long-video understanding. HF Daily Papers' note
Omni-Decision replaces a growing conversation history with an evidence ledger tracking missing, confirmed, and conflicting evidence. A critic filters noisy observations before they reach the planner, keeping the working context compact. The authors report 81.4% accuracy on OmniGAIA at about 43% of Gemini-3.1-Pro’s cost per question, plus 65.0% on WorldSense long-video understanding. HF Daily Papers' note
score 5