LongVideoAgent Qwen2.5-7B Checkpoint

This repository provides the specific Qwen2.5-7B-based language model checkpoint utilized within the LongVideoAgent multi-agent framework. LongVideoAgent is designed for complex question answering over long video content, decomposing reasoning into specialized roles.

Key Capabilities & Architecture

• Specialized for Long-Video QA: This 7.6 billion parameter model is fine-tuned to support the LongVideoAgent's approach to understanding and answering questions about extended videos.
• Multi-Agent Integration: It functions as a core component within a framework that includes a MasterAgent for planning, a GroundingAgent for subtitle-based temporal grounding, and a VisionAgent for extracting local visual evidence.
• Performance Improvement: The model demonstrates a notable performance gain, improving LongTVQA+ accuracy by 6.66 percentage points compared to the base Qwen2.5-7B-Instruct model, achieving 64.00% accuracy.

Intended Use Cases

• Research: Ideal for academic research into long-video question answering and multi-agent reasoning systems.
• Reproducibility: Designed for reproducing experiments conducted with the LongVideoAgent framework.
• Agentic Reasoning Studies: Useful for studying how agentic approaches can be applied to complex video understanding tasks.

Note: This checkpoint is not a standalone general-purpose video model. For full functionality, it requires integration with the official LongVideoAgent codebase and evaluation pipeline, available on their project page and GitHub repository.