VEPO-ERM-Qwen3-VL-4B

VEPO-ERM-Qwen3-VL-4B

This resource accompanies VEPO: Learning to Credit Visual Evidence Seeking for Agentic Video Understanding.

Overview

This Qwen3-VL-4B-Instruct checkpoint serves as the frozen Evidence Reward Model (ERM) in VEPO. Given a question and visual observations accumulated from searched video windows, it assesses whether their joint content provides sufficient evidence to answer the question.

The training client computes an evidence-sufficiency score from the model's Yes/No token probabilities. Full VEPO jointly evaluates accumulated search observations, using four frames per search window and excluding the initial overview. The final score determines the trajectory reward tier; positive score increases provide local credit to search actions. The ERM remains frozen during policy optimization.

Set ERM_MODEL_PATH to the downloaded directory before starting the ERM server.

Download and use

hf download Echoch/VEPO-ERM-Qwen3-VL-4B --local-dir /path/to/models/vepo-erm

The paper's code release provides the Slime training overlay, temporal-search environment, and ERM serving launcher. Prepare a compatible Slime runtime and the required video files separately.

Repository contents

The repository contains Safetensors model weights, model configuration, tokenizer and chat template, and image/video processor configuration. Training optimizer, scheduler, and trainer state are excluded.

Related resources

Usage terms

The existing other license declaration is retained. Source-model and dataset terms continue to apply. No new license is assigned by this documentation update.