Retrieval-augmented generation (RAG) connects large language models (LLMs) to external knowledge, but single-round retrieval is often insufficient for complex multi-hop questions. To improve search on complex tasks, most existing works combine multi-round retrieval with reasoning through end-to-end training. While these approaches improve problem-solving performance, they still face challenges in task reasoning and model training. First, due to the lack of an explicit subgoaling mechanism, most existing RAG methods may produce ambiguous retrieval execution paths, leading to the degradation of model performance. Second, a few existing RAG methods have started to incorporate planning and execution into the model training process. However, they typically rely on a single final reward, causing hierarchical credit entanglement between planning and execution and preventing the model from clearly separating plan errors from retrieval errors. To address these limitations, we propose APEX-Searcher, a coordinated planner–executor framework that optimizes a recursive planner with RL to accommodate multiple valid, dependency-aware decompositions while training the executor with SFT under clearer subtask-level supervision. Extensive experiments show consistent gains in both multi-hop RAG and task planning across benchmarks.

Large language models (LLMs) enabled dialogue systems have become one of the central modes in human-machine interaction, which bring about vast amounts of conversation logs and increasing demand for dialogue generation. The dialogue’s life-cycle spans from Prelude through Interlocution to Epilogue, encompassing rich dialogue elements. Despite large volumes of dialogue-related studies, there is a lack of systematic investigation into the dialogue stages to frame benchmark construction that covers comprehensive dialogue elements. This hinders the precise modeling, generation and assessment of LLMs-based dialogue systems. To bridge this gap, in this paper, we introduce a new research task—Dialogue Element MOdeling, including Element Awareness and Dialogue Agent Interaction, and propose a novel benchmark, DEMO, designed for a comprehensive dialogue modeling and assessment. On this basis, we further build the DEMO agent with the adept ability to model dialogue elements via imitation learning. Extensive experiments on DEMO indicate that current representative LLMs still have considerable potential for enhancement, and our DEMO agent performs well in both dialogue element modeling and out-of-domain tasks.