LMM-Searcher: Long-horizon Agentic Multimodal Search (April 2026)

Free

RUC: file-based visual context management + progressive on-demand image loading — scales to 100-turn search horizons, SOTA on MM-BrowseComp and MMSearch-Plus

FreeFree tier

Inputs: text, imageOutputs: text

Type

Open Source

Visit Website GitHub (40 stars)

About LMM-Searcher: Long-horizon Agentic Multimodal Search (April 2026)

LMM-Searcher is a novel long-horizon multimodal deep search framework introduced in an April 2026 arXiv paper. It tackles the challenges of managing heterogeneous visual information and high token costs over extended search horizons by employing a file-based visual representation mechanism. Visual assets are offloaded to an external file system and mapped to lightweight textual identifiers (UIDs), reducing context overhead while preserving multimodal information. The agent is equipped with a fetch-image tool for progressive, on-demand visual loading. The paper also presents a data synthesis pipeline that generates complex cross-modal multi-hop queries, distilling 12K high-quality trajectories to fine-tune Qwen3-VL-Thinking-30A3B into a specialized agent. LMM-Searcher achieves state-of-the-art performance among open-source models on MM-BrowseComp and MMSearch-Plus, scales to 100-turn search horizons, and demonstrates strong generalizability across different base models. The code is scheduled for open-source release.

Key Features

File-based visual context management with lightweight textual identifiers (UIDs)

Progressive on-demand image loading via fetch-image tool

Scales to 100-turn search horizons

State-of-the-art on MM-BrowseComp and MMSearch-Plus among open-source models

Data synthesis pipeline for generating complex cross-modal multi-hop queries

12K distilled trajectories for fine-tuning Qwen3-VL-Thinking-30A3B

Strong generalizability across different base models

Open-source release of code

Pros & Cons

Pros

Mitigates context explosion by offloading visual assets to an external file system
Preserves full multimodal information for future access via UIDs
Progressive on-demand visual loading reduces unnecessary token usage
Demonstrates strong performance on long-horizon benchmarks (100 turns)
Achieves state-of-the-art results among open-source models on MM-BrowseComp and MMSearch-Plus
Generalizable across different base models, not limited to the fine-tuned variant
Open-source availability allows community adoption and improvement

Cons

Code not yet released at the time of the paper (scheduled for future release)
Requires external file system for visual asset storage, adding infrastructure dependency
Reported performance is based on fine-tuning a specific base model (Qwen3-VL-Thinking-30A3B); results may vary with other base models
Currently an academic research contribution; no ready-to-use product or API available

Best For

Complex multimodal search tasks requiring iterative collection of textual and visual evidenceLong-horizon agentic tasks like multi-step research or deep browsingScenarios with high token costs from multimodal inputs where context explosion is a concernApplications needing scalable visual context management across many turns

FAQ

What is LMM-Searcher?

LMM-Searcher is a long-horizon multimodal deep search framework proposed in an April 2026 arXiv paper. It uses a file-based visual representation to manage context and a fetch-image tool for on-demand image loading, enabling scalable agentic search over 100 turns.

How does LMM-Searcher handle long search horizons?

It offloads visual assets to an external file system and maps them to lightweight textual identifiers (UIDs), reducing context overhead. The agent can then fetch images on demand using a fetchtool, preserving multimodal information without context explosion.

What benchmarks does LMM-Searcher excel on?

LMM-Searcher achieves state-of-the-art performance among open-source models on MM-BrowseComp and MMSearch-Plus, which are challenging long-horizon multimodal benchmarks.

Is LMM-Searcher open source?

Yes, the authors plan to release the code as open source. The paper states: 'Our code will be released in [GitHub URL].'

What base model is used for LMM-Searcher?

The fine-tuned version of LMM-Searcher is built on Qwen3-VL-Thinking-30A3B, using 12K distilled trajectories from a novel data synthesis pipeline. The framework also demonstrates generalizability to other base models.