🔍 LMM-Searcher: AI That Can Search 100 Steps Without Forgetting

Ever asked an AI to research something complex, only to watch it forget everything it found 10 minutes ago?

The longer AI searches, the more images and text pile up in its memory — until it chokes. Most AI agents start hallucinating or going in circles after just 10-20 search steps.

Researchers have built LMM-Searcher, a framework that solves this with an elegantly simple trick: instead of cramming every image into the AI's brain, it stores them externally and pulls them up only when needed.

Think of it like a detective with a filing cabinet instead of trying to hold every piece of evidence in both hands.

🎯 What it achieves:

• Searches across 100+ turns without losing context or accuracy
• Images stored externally with lightweight text references — no memory explosion
• On-demand visual retrieval: loads images only when reasoning requires them
• State-of-the-art among open-source models on MM-BrowseComp and MMSearch-Plus benchmarks
• Built on Qwen3-VL, trained on 12,000 synthesized search trajectories

This matters because real-world research tasks — comparing products, planning trips, investigating topics — require dozens of steps across multiple sources. Current AI agents hit a wall fast. LMM-Searcher pushes that wall 10x further.

The code will be open-sourced, meaning every AI assistant could eventually benefit from this approach.