Research map / Vision-language and multimodal / Multimodal LLMs

Agent0-VL: Exploring Self-Evolving Agent for Tool-Integrated Vision-Language Reasoning

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Agent0-VL is a self-evolving LVLM that unifies a tool-using Solver and a tool-grounded Verifier in one model, trained with SFT cold start and GRPO-based self-rewarding RL to improve multimodal reasoning.

Problem

Vision-language models rely on human-annotated or external reward signals, and text-only self-evaluation fails to verify complex visual reasoning steps, causing evaluation hallucinations.

Approach

Integrate tool use into both reasoning and self-evaluation: a single LVLM alternates between a Solver (multi-turn tool-integrated reasoning) and a Verifier (tool-grounded critique producing score, confidence, and critique), with confidence-gated self-repair and process rewards optimized via GRPO in a Self-Evolving Reasoning Cycle.

Key results

• Agent0-VL-7B averages 65.6 across seven benchmarks versus 58.3 for Qwen2.5-VL-7B, a reported 12.5% improvement over the base model.
• Agent0-VL-8B averages 74.6, surpassing Qwen3-VL-8B (70.3) and Qwen3-VL-8B-TIR (71.3).
• Used as a process reward model for Best-of-8, Agent0-VL raises Qwen2.5-VL-7B overall accuracy from 58.3 to 62.8 (average gain of 7.3% across models).

Surprising findings

• Removing SERC and keeping only SFT causes the largest ablation drop (about 8.7% average), while removing self-repair costs only about 2.5%.

Limitations

• Pretraining with external-reward RL is needed as warm-up before the purely self-rewarding phase.
• Reported gains are inconsistent in places (e.g., the 12.5% headline versus 65.6 vs 58.3 averages; the abstract's claim of zero external reward conflicts with the external-reward warm-up).
• Training data relies on GPT-5 and Qwen2.5-VL-72B teacher trajectories, so the pipeline is not fully self-generated.

Experimental setup

Datasets: Geometry3K GeoQA Mulberry MM-Eureka Retool MathVerse MathVista WeMath ChartQA arXivQA ThinkLiteVL LLaVA-OV-Image

Benchmarks: MathVerse MathVision MathVista WeMath HallusionBench ChartQA MMMU

Metrics: accuracy average accuracy Best-of-8 accuracy

Baselines: Qwen2.5-VL-7B-TIR Qwen3-VL-8B-TIR ThinkLite-VL-7B Vision-Zero EvoLMM VisPlay

Builds on: GRPO Vision-Zero TTRL

Closest papers by meaning

• Agent-X: Evaluating Deep Multimodal Reasoning in Vision-Centric Agentic Tasks (ICLR 2026 · similarity 0.91)
• Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.90)
• SenseSearch: Empowering Vision-Language Models with High-Resolution Agentic Search-Reasoning via Reinforcement Learning (CVPR 2026 · similarity 0.90)
• SSL4RL: Revisiting Self-supervised Learning as Intrinsic Reward for Visual-Language Reasoning (ICML 2026 · similarity 0.90)
• VICO: Visual Environments Co-Evolving for Vision-Language Model Reasoning (NeurIPS 2026 · similarity 0.90)
• VisPlay: Self-Evolving Vision-Language Models (CVPR 2026 · similarity 0.90)