Active Zero: Self-Evolving Vision-Language Models through Active Environment Exploration

2026-01-28 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, long

Summary

Active-Zero is a novel framework designed to enhance vision-language models (VLMs) through active environment exploration, moving beyond passive interaction with static image collections. It employs three co-evolving agents: a Searcher that retrieves relevant images from open-world repositories, a Questioner that synthesizes calibrated reasoning tasks, and a Solver refined through accuracy rewards. This closed-loop system enables self-scaffolding auto-curricula, allowing the model to autonomously construct its learning trajectory. Evaluated on Qwen2.5-VL-7B-Instruct across 12 benchmarks, Active-Zero achieved a 53.97 average accuracy on reasoning tasks, marking a 5.7% improvement, and 59.77 on general understanding, a 3.9% improvement, consistently outperforming existing self-play baselines. The framework also showed robust performance on Qwen2.5-VL-3B-Instruct, with Iter2 consistently yielding the strongest results.

Key takeaway

For research scientists developing self-evolving vision-language models, Active-Zero demonstrates that actively curating visual data from open-world environments is crucial for scalable improvement. You should consider implementing a multi-agent framework that dynamically selects training data based on the model's evolving capabilities, rather than relying on static datasets, to achieve superior performance in both reasoning and general visual understanding.

Key insights

Active environment exploration with co-evolving agents significantly improves VLM reasoning and general understanding.

Principles

• Active data curation is superior to passive interaction.
• Self-scaffolding auto-curricula drive autonomous VLM improvement.
• Iterative optimization of specialized agents enhances multimodal reasoning.

Method

Active-Zero uses a tri-agent framework (Searcher, Questioner, Solver) in an iterative self-play cycle. The Searcher retrieves images, the Questioner synthesizes tasks, and the Solver improves via reinforcement learning.

In practice

• Implement a Searcher to dynamically curate training data.
• Design a Questioner to generate multi-step reasoning tasks.
• Utilize Group Relative Policy Optimization (GRPO) for Solver training.

Topics

• Self-Evolving VLMs
• Active Environment Exploration
• Tri-Agent Architecture
• Auto-Curriculum Learning
• VLM Self-Play

Best for: Computer Vision Engineer, Research Scientist, AI Researcher, AI Scientist, Deep Learning Engineer

Related on AIssential

• D-VLA: A High-Concurrency Distributed Asynchronous Reinforcement Learning Framework for Vision-Language-Action Models — D-VLA decouples simulation and optimization to enable high-concurrency distributed RL for large Vision-Language-Action models.
• EvoVid: Temporal-Centric Self-Evolution for Video Large Language Models — EvoVid enables Video-LLMs to self-evolve temporal reasoning from raw video, eliminating human annotation dependency.
• From Structure to Synergy: A Survey of Vision-Language Perception Paradigm Evolution in Multimodal Large Language Models — The survey unifies vision-language perception in MLLMs, proposing a five-stage evolution taxonomy and future directions for AGI.
• Visual-OPSD: Cross-Modal On-Policy Self-Distillation for Efficient Unified Multimodal Reasoning — The generation pathway of visual thoughts encodes reasoning that can be distilled for efficient, text-only multimodal models.
• Proact-VL: A Proactive VideoLLM for Real-Time AI Companions — Proact-VL enables human-like AI companions for real-time video by balancing proactive timing with controlled, low-latency content generation.
• Active Exploring like a Pigeon: Reinforcing Spatial Reasoning via Agentic Vision-Language Models — An agentic VLM pipeline uses dynamic cognitive maps and Spatial Assertion Codes for robust spatial reasoning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.