BRZ911/Latent-VC-9B

Latent-VC-9B: Enhanced Video Reasoning with Latent Visual Cache

Latent-VC-9B is a 9 billion parameter multimodal model developed by BRZ911, specifically engineered for grounded long-form video reasoning. Built upon the Qwen3.5-9B-Base architecture, its core innovation is the Latent Visual Cache (Latent-VC), a recurrent latent memory integrated into the decoder. This cache actively mitigates "Visual Anchoring Decay," a common challenge in long-form video analysis where models lose grounding to visual evidence over time.

Key Capabilities & Innovations

• Recurrent Latent Visual Cache: A novel architectural component that maintains a compact latent visual memory, ensuring continuous visual grounding throughout extended video sequences.
• Two-Stage Training: Undergoes Supervised Fine-Tuning (SFT) with contrastive cache alignment, followed by GRPO (Generative Reinforcement Learning with Policy Optimization) using vision-grounded rewards and latent grounding supervision.
• Mitigates Visual Anchoring Decay: Addresses the problem of models losing track of visual evidence in long videos, leading to more consistent and accurate reasoning.
• Outperforms Baselines: Consistently surpasses strong CoT (Chain-of-Thought) and SFT+GRPO baselines across six diverse video benchmarks, particularly excelling in tasks requiring deep grounding and long-video understanding.
• Efficient Responses: Achieves higher accuracy with substantially shorter response lengths compared to other models.

Ideal Use Cases

• Long-form video analysis: Excels in scenarios where understanding extended video content is crucial.
• Grounded video question answering: Provides more accurate and visually-grounded answers to complex questions about video content.
• Video event comprehension: Suitable for tasks requiring detailed understanding of events and actions over time within videos.
• Research and development in multimodal AI: Offers a robust foundation for exploring advanced video reasoning techniques.