GenSearcher/Gen-Searcher-SFT-8B

Overview

Gen-Searcher-SFT-8B is an 8 billion parameter supervised fine-tuned (SFT) model, developed as part of the Gen-Searcher project by Feng et al. This model is specifically designed to function as a multimodal deep research agent for image generation, particularly for scenarios requiring complex real-world knowledge. It represents an intermediate stage, intended for further reinforcement learning (RL) training using the GRPO algorithm with dual reward feedback.

Key Capabilities

• Agentic Search: Capable of searching the web, browsing evidence, and reasoning over multiple sources.
• Visual Reference Search: Integrates the ability to search for visual references to inform image generation.
• Enhanced Accuracy: Aims to enable more accurate and up-to-date image synthesis by grounding generation in real-world information.
• Performance Gains: Achieves significant improvements, with over 15-point gains on the KnowGen and WISE benchmarks, demonstrating strong transferability to various image generators.
• Dataset Development: Utilizes dedicated training datasets, Gen-Searcher-SFT-10k and Gen-Searcher-RL-6k, and introduces a new benchmark, KnowGen, for search-grounded image generation.

Good For

• Research in Agentic AI: Ideal for researchers exploring deep research agents and multimodal reasoning.
• Image Generation with External Knowledge: Suitable for applications requiring image synthesis grounded in up-to-date, real-world information.
• Developing Search-Augmented LLMs: Provides a foundation for models that integrate web search and reasoning into creative tasks.
• Further RL Training: Serves as a robust base model for subsequent reinforcement learning experiments in agentic image generation.