VISOR: Agentic Visual Retrieval-Augmented Generation via Iterative Search and Over-horizon Reasoning

2026-04-10

AI Summary

This work addresses the challenges of sparse visual evidence and search drift over long horizons in existing vision-augmented retrieval-based generation systems for complex multi-step reasoning. To this end, the authors propose VISOR, a unified single-agent framework that constructs a structured evidence space to support cross-page progressive reasoning. VISOR incorporates a visual action evaluation and correction mechanism and introduces a dynamic trajectory sliding window with intent injection to effectively mitigate context overload and search drift. The framework is trained using Group Relative Policy Optimization–based reinforcement learning, augmented with state masking and credit assignment to accommodate dynamic context reconstruction. Experimental results demonstrate that VISOR achieves state-of-the-art performance on multiple benchmarks—including ViDoSeek, SlideVQA, and MMLongBench—significantly improving both accuracy and efficiency in long-horizon visual reasoning tasks.

Technology Category

• Computer Vision: Visual Reasoning & Symbolic Representations
• Search and Optimization: Learning to Search
• Knowledge Representation and Reasoning: Common-Sense Reasoning

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Visual Retrieval-Augmented Generation (VRAG) empowers Vision-Language Models to retrieve and reason over visually rich documents. To tackle complex queries requiring multi-step reasoning, agentic VRAG systems interleave reasoning with iterative retrieval.. However, existing agentic VRAG faces two critical bottlenecks. (1) Visual Evidence Sparsity: key evidence is scattered across pages yet processed in isolation, hindering cross-page reasoning; moreover, fine-grained intra-image evidence often requires precise visual actions, whose misuse degrades retrieval quality; (2) Search Drift in Long Horizons: the accumulation of visual tokens across retrieved pages dilutes context and causes cognitive overload, leading agents to deviate from their search objective. To address these challenges, we propose VISOR (Visual Retrieval-Augmented Generation via Iterative Search and Over-horizon Reasoning), a unified single-agent framework. VISOR features a structured Evidence Space for progressive cross-page reasoning, coupled with a Visual Action Evaluation and Correction mechanism to manage visual actions. Additionally, we introduce a Dynamic Trajectory with Sliding Window and Intent Injection to mitigate search drift. They anchor the evidence space while discarding earlier raw interactions, preventing context from being overwhelmed by visual tokens. We train VISOR using a Group Relative Policy Optimization-based Reinforcement Learning (GRPO-based RL) pipeline with state masking and credit assignment tailored for dynamic context reconstruction. Extensive experiments on ViDoSeek, SlideVQA, and MMLongBench demonstrate that VISOR achieves state-of-the-art performance with superior efficiency for long-horizon visual reasoning tasks.

Problem

• Visual Retrieval-Augmented Generation
• Evidence Sparsity
• Search Drift
• Cross-page Reasoning
• Long-horizon Visual Reasoning

Innovation

• Visual Retrieval-Augmented Generation
• Agentic Reasoning
• Cross-page Evidence Integration
• Search Drift Mitigation
• Reinforcement Learning for Vision-Language