Research map / Multi-agent systems and games / LLM agents

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

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes JADE, which jointly trains a planner and seven executor roles as shared-parameter personas of one Qwen2.5-7B-Instruct backbone via PPO with outcome-based global rewards for dynamic agentic RAG.

Problem

Dynamic agentic RAG systems optimize the planner while treating executors as frozen black-box tools, causing a strategic-operational mismatch where sophisticated plans fail to execute and performance can even drop.

Approach

Model the planner and executors as a cooperative multi-agent team sharing one LLM backbone, with a unified experience buffer of heterogeneous role transitions, a shared terminal F1-minus-cost reward, and local format penalties, so planning and execution co-adapt end to end via PPO.

Key results

• JADE reaches an average F1 of 53.86 across seven QA benchmarks, +8.29 over the best baseline MMOA-RAG (45.57).
• JADE scores 53.86 vs 41.22 for MAO-ARAG, which freezes executors, a gap of 12.64 points.
• With frozen executors, GPT-4o achieves 56.82 average on NQ and HotpotQA, while JADE's trained 7B executors reach 58.24 on the same two datasets.

Surprising findings

• Adding the Document Selector module degrades performance on the frozen backbone (41.74 to 41.13) but helps after joint training (57.10 to 58.24).
• Aggressive joint cost penalties cause the policy to collapse into a single-turn retrieve-generate loop equivalent to Vanilla RAG.

Limitations

• Ablation of the DS module and executor-backbone comparison use only NQ and HotpotQA.
• Relies on a fixed Wikipedia E5 retriever that is not trained.
• Reward hyperparameters (alpha, beta) are sensitive and can cause collapse.

Experimental setup

Datasets: Natural Questions PopQA AmbigQA HotpotQA 2WikiMultiHopQA Musique Bamboogle

Benchmarks: Natural Questions PopQA AmbigQA HotpotQA 2WikiMultiHopQA Musique Bamboogle

Metrics: F1

Baselines: LLM w/o RAG Vanilla RAG RRR BGM MMOA-RAG Adaptive RAG Search-R1 MAO-ARAG

Builds on: MMOA-RAG MAO-ARAG Search-R1 PPO MAPPO

Closest papers by meaning

• FML-bench: A Controlled Study of AI Research Agent Strategies from the Perspective of Search Dynamics (NeurIPS 2026 · similarity 0.88)
• C-3PO: Compact Plug-and-Play Proxy Optimization to Achieve Human-like Retrieval-Augmented Generation (ICML 2025 · similarity 0.88)
• EvoMAS: Evolutionary Generation of Multi-Agent Systems (ICML 2026 · similarity 0.87)
• AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems (NeurIPS 2025 · similarity 0.87)
• MARS: Modular Agent with Reflective Search for Automated AI Research (ICML 2026 · similarity 0.87)
• SWE-Search: Enhancing Software Agents with Monte Carlo Tree Search and Iterative Refinement (ICLR 2025 · similarity 0.87)