---
library_name: transformers
license: apache-2.0
base_model: Qwen/Qwen3-VL-8B-Instruct
pipeline_tag: image-text-to-text
tags:
- vision-language-model
- qwen3-vl
- conversational
---

# ReCAP-8B-Instruct

ReCAP-8B-Instruct is a vision-language model fine-tuned from
[Qwen/Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct), designed to enable robust CAPTCHA solving within native GUI agents while preserving general GUI interaction capabilities.

This model is introduced in “[CAPTCHA Solving for Native GUI Agents: Automated Reasoning-Action Data Generation and Self-Corrective Training](https://arxiv.org/abs/2603.23559)”, accepted at ICML 2026.

---

## 🚀 Overview

ReCAP-8B-Instruct extends a general-purpose GUI agent with CAPTCHA-solving ability by learning from structured reasoning-action trajectories.

It operates end-to-end:

- Input: raw screenshots
- Output: reasoning + executable GUI actions (click, type, drag)

---

## ✨ Key Features

- Unified agent: Handles both CAPTCHA and general GUI tasks
- Reasoning-action modeling: Learns both decisions and execution
- Self-correction: Improves robustness by learning from failures
- Efficient interaction: Generates multiple actions per step

---

## 🧠 Capabilities

Supports diverse CAPTCHA types:

- Text / OCR
- Icon selection & matching
- Image grid reasoning
- Slider / drag tasks
- Multi-step interaction challenges

Core skills:

- Visual understanding
- Spatial reasoning
- Continuous control
- Multi-step planning

---

## 📊 Performance

- ~78.6% success rate on synthetic CAPTCHA benchmark
- 14.50 percentage-point average improvement over Qwen3-VL-8B-Instruct across 26 real-world CAPTCHA types in zero-shot evaluation
- Strong improvements on interaction-heavy tasks (e.g., slider, image grid)
- Maintains competitive performance on general GUI benchmarks

---

## ⚠️ Ethical Considerations

This model is released for research purposes only.
It is intended to study and improve the robustness of human-verification systems, not to bypass them.