Title: PhotoCraft: Enhancing Deep Image Search via Agentic Reasoning and Hierarchical Self-Evolving Memory

Abstract:

Conducting deep image search demands complex, multi-step reasoning that leverages rich contextual signals, including temporal data, geographical locations, and event connections. Yet, current Large Language Model (LLM)-based agents typically operate as stateless, reactive systems. They lack the persistent memory necessary to sustain long-term context or transfer prior experiences between different tasks, a deficiency that frequently results in execution drift and isolated experiences.

To overcome these challenges, we introduce PhotoCraft, a training-free, hierarchical memory framework designed specifically for photo-search agents. Drawing inspiration from human cognitive processes, PhotoCraft integrates working, episodic, and semantic memories into Multimodal Large Language Models (MLLMs). These memory components are dynamically activated during the reasoning phase to ensure logical consistency and facilitate knowledge transfer throughout the multi-step reasoning and answer generation processes.

Our extensive evaluations on the DISBench benchmark reveal that PhotoCraft consistently enhances context-aware retrieval performance across various MLLM backbones. The system delivers performance improvements of up to 18.5% and successfully addresses critical bottlenecks associated with memoryless deep image search, thereby providing a viable and practical route toward developing reliable and generalizable multimodal search agents.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC