Core Insights - The article discusses the development of REDSearcher, a scalable and cost-efficient framework for long-horizon search agents, which aims to enhance the capabilities of AI agents in deep search tasks [5][28] - It emphasizes the importance of "deep search" in AI agents, which requires maintaining goals, verifying information, and dynamically adjusting strategies, akin to human experts [2][28] Group 1: Challenges in Training AI Agents - The training of autonomous agents faces three major bottlenecks: data scarcity, capability gap, and environmental absence [4] - High-difficulty long-range question-answering tasks heavily rely on manual annotation, which is costly, necessitating an automated synthesis of complex questions [4] - Pre-trained models, while rich in knowledge, lack the ability to interact with real environments for long-range tasks, requiring a low-cost mid-training phase to bridge this gap [4] Group 2: Complexity in Deep Search Tasks - The article defines "sufficiently difficult" deep search questions, focusing on structural complexity rather than just the number of reasoning steps [7] - Topological complexity is measured using the concept of TreeWidth, which characterizes the structural difficulty of complex tasks where information branches and loops [8] - Information dispersion is introduced to prevent shortcuts in searches, indicating that a higher degree of dispersion requires more interactions with the environment to gather sufficient information [10] Group 3: Automated Synthesis of Complex Questions - The REDSearcher team employs a graph-to-text process to synthesize high-difficulty deep search questions, ensuring they are high-quality, solvable, and have unique answers [12] - The framework incorporates structured information and web browsing to cover different search environments, enhancing the complexity of generated questions [13] Group 4: Mid-Training Framework - REDSearcher utilizes a two-stage mid-training framework to enhance the agent's atomic and combinatorial capabilities, transitioning from language modeling to agent functionality [16] - This approach addresses issues like goal drift and repetitive searches that arise from a lack of multi-turn interaction training [16] Group 5: Multi-Modal Expansion - The framework extends from text graphs to multi-modal graphs, enhancing topological structure and incorporating tool usage in question construction [17][18] - Visual attributes are anchored to force the model to recognize images before associating knowledge, ensuring that visual search becomes a necessary part of reasoning [19] Group 6: Continuous Evolution Post-Training - Post-training involves a dual-phase enhancement of SFT and Agentic RL, leading to improved efficiency and performance as the model learns to reduce ineffective calls while increasing accuracy [22] - The model demonstrates a positive feedback loop where training leads to smarter information retrieval strategies [22] Group 7: Experimental Results - REDSearcher achieves state-of-the-art performance in various deep search benchmarks, outperforming several advanced closed-source models like GPT-5 and Gemini-2.5 [26] - The model's ability to filter noise and maintain consistency in multi-turn interactions is highlighted as a key strength [24][25] Group 8: Conclusion - REDSearcher represents a systematic design approach to deep search tasks, providing a reproducible and low-cost training path for AI systems to transition from static knowledge queries to autonomous exploration and information integration in open environments [28]