Robust-TO tackles video reasoning blind trust

Video reasoning models suffer from a silent failure mode where they blindly trust degraded visual inputs. Robust-TO demonstrates that structuring video understanding as an agentic tool orchestration problem is a far more robust path than relying on ever-larger monolithic models.

• –**Blind Trust Mitigation:** Instead of treating all video frames equally, it uses a per-frame reliability-relevance score to filter out corruptions like glare, motion blur, and occlusion.
• –**Three-Tiered Synthesis:** Heterogeneous tools (e.g., action models, OCR) return evidence with calibrated reliability scores, which a three-tier synthesis process weights dynamically during reasoning.
• –**GRPO Optimization:** A specialized confidence-cost reward optimizes the policy via Group Relative Policy Optimization, balancing reasoning accuracy, evidence reliability, and computational efficiency.
• –**Superior Benchmark Performance:** Achieving 56.4% accuracy on clean inputs and 54.3% under corruption, it beats Gemini-2.5-Pro while adding less than 5% latency overhead.