VEPO-Qwen2.5-VL-7B

VEPO-Qwen2.5-VL-7B

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

Overview

The trained VEPO policy uses Qwen2.5-VL-7B-Instruct as its backbone. It performs multi-turn temporal search with the get_frames tool and produces a final answer after observing video evidence. This release contains the policy checkpoint from RL iteration 139.

VEPO combines a three-level trajectory reward, dynamic group filtering, and positive evidence-sufficiency gains assigned to the corresponding search-action token spans. Wrong answers receive 0.0, unsupported correct answers receive 0.4, and evidence-supported correct answers receive 1.0.

Use this checkpoint with the temporal-search environment accompanying the paper.

Download and use

hf download Echoch/VEPO-Qwen2.5-VL-7B --local-dir /path/to/models/vepo-policy

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

Source-model and dataset terms continue to apply. No new license is assigned by this documentation update.