Echoch/VEPO-ERM-Qwen3-VL-4B
Echoch/VEPO-ERM-Qwen3-VL-4B is a 4 billion parameter vision-language model based on Qwen3-VL, developed by Echoch. It functions as an Evidence Reward Model (ERM) within the VEPO framework, designed to assess the sufficiency of visual evidence from video observations to answer a given question. This model is specifically optimized for agentic video understanding tasks, providing an evidence-sufficiency score to credit search actions in video analysis.
Loading preview...
Overview
Echoch/VEPO-ERM-Qwen3-VL-4B is a 4 billion parameter vision-language model built upon the Qwen3-VL architecture. It serves as a frozen Evidence Reward Model (ERM) within the VEPO (Visual Evidence Seeking for Agentic Video Understanding) framework. Its primary function is to evaluate whether a combination of a question and accumulated visual observations from searched video windows provides sufficient evidence to answer that question.
Key Capabilities
- Evidence Sufficiency Assessment: Computes an evidence-sufficiency score based on the model's Yes/No token probabilities, indicating how well visual information supports a given query.
- Agentic Video Understanding: Designed to provide local credit to search actions in video analysis by determining trajectory reward tiers based on positive score increases.
- Frozen ERM: Operates as a stable, frozen component during policy optimization within the VEPO system, ensuring consistent evidence evaluation.
Good for
- Integrating into VEPO-based systems: Essential for researchers and developers working with the VEPO framework for agentic video understanding.
- Evaluating visual evidence: Useful for tasks requiring an objective assessment of whether visual data from videos adequately addresses a specific question.
- Research in video-based AI agents: Provides a core component for developing and training AI agents that learn to seek and utilize visual evidence from video content.