JADE: Bridging the Strategic-Operational Gap in Dynamic Agentic RAG

TL;DR

JADE tackles the strategic-operational mismatch in dynamic agentic RAG by unifying planning and execution under a shared LLM backbone and optimizing them end-to-end with outcome-based rewards. Through a Shared-Parameter MSMDP, a cooperative multi-agent setup, and a Unified Experience Replay with PPO, JADE enables co-adaptation where planners respect executor capabilities and executors align with high-level strategy. Empirically, JADE achieves state-of-the-art average F1 across seven benchmarks, with notable gains in multi-hop tasks and a tunable efficiency-performance trade-off via turn and retrieval penalties. The work shows that joint optimization of modular roles can surpass decoupled approaches and even larger frozen-backbone baselines, offering a scalable, efficient path for sophisticated agentic reasoning. The practical impact lies in delivering a flexible, cost-effective framework that balances reasoning depth and resource use in open-domain QA and related tasks.

Abstract

The evolution of Retrieval-Augmented Generation (RAG) has shifted from static retrieval pipelines to dynamic, agentic workflows where a central planner orchestrates multi-turn reasoning. However, existing paradigms face a critical dichotomy: they either optimize modules jointly within rigid, fixed-graph architectures, or empower dynamic planning while treating executors as frozen, black-box tools. We identify that this \textit{decoupled optimization} creates a ``strategic-operational mismatch,'' where sophisticated planning strategies fail to materialize due to unadapted local executors, often leading to negative performance gains despite increased system complexity. In this paper, we propose \textbf{JADE} (\textbf{J}oint \textbf{A}gentic \textbf{D}ynamic \textbf{E}xecution), a unified framework for the joint optimization of planning and execution within dynamic, multi-turn workflows. By modeling the system as a cooperative multi-agent team unified under a single shared backbone, JADE enables end-to-end learning driven by outcome-based rewards. This approach facilitates \textit{co-adaptation}: the planner learns to operate within the capability boundaries of the executors, while the executors evolve to align with high-level strategic intent. Empirical results demonstrate that JADE transforms disjoint modules into a synergistic system, yielding remarkable performance improvements via joint optimization and enabling a flexible balance between efficiency and effectiveness through dynamic workflow orchestration.

Figures (4)

• Figure 1: Different Paradigms of Agentic RAG
• Figure 2: The overall framework of JADE. The system operates in an iterative loop of planning and execution. (1) During Inference (Left), the process is recursive: for the current unsolved node (i.e., a specific sub-query, see Eq.\ref{['eq:trace_node']}) in the global state $s_t$, the Planner is invoked to orchestrate a dedicated dynamic workflow. This workflow is executed by specialized Executors (e.g., Query Decomposition, Retrieval Agent) to update $s_t$ for the next round. (2) During Training (Right), to achieve joint optimization, every agent involved in the multi-turn trajectory generates transition triplets $\langle o_{t,k}, a_{t,k}, r_{t,k} \rangle$. These transitions are aggregated into a unified Experience Buffer, which is then used to update the parameter-sharing policy model, aligning strategic planning with operational execution.
• Figure 3: Hyperparameter Sensitivity Analysis of the Cost Reward $R_{cost}$. We analyze the training dynamics under varying penalty coefficients. (a) Varying the turn penalty $\alpha$ (with $\beta=0$) effectively constrains the Planner to reduce reasoning steps (Solid Lines) while maintaining F1 Score. (b) Varying the retrieval penalty $\beta$ (with $\alpha=0$) encourages Executors to reduce retrieval calls. (c) Jointly increasing both parameters ($\alpha=\beta$) leads to over-penalization, causing the system to rapidly degenerate into a static single-turn RAG workflow to minimize costs, resulting in pronounced performance degradation.
• Figure 4: Detailed Analysis of Workflow Patterns and Data Selection (DS) Dynamics.(a) and (b) illustrate how the Planner's strategy evolves over training steps across different task complexities. (c) and (d) depict the utilization intensity of the Document Selector, calculated as the ratio of DS calls to total reasoning turns.