hierarchical planning

Existing language model agents struggle to efficiently execute complex instructions in long-horizon tasks due to insufficient planning capabilities. This work proposes a planner-centric multi-agent framework comprising a planner, an executor, and a memory manager. Through computational resource allocation analysis, we demonstrate that the planning component predominantly governs overall performance. Leveraging this insight, we apply reinforcement learning exclusively to the planner, incorporating trajectory-level rewards and a vision-language model-based evaluation mechanism to enable asymmetric computation allocation. The resulting approach achieves significant performance gains across diverse benchmarks—including web navigation, operating system control, and tool usage—thereby validating the efficacy and strong generalization of prioritizing high-level planning.

Existing multi-agent frameworks face two key bottlenecks in complex planning tasks: unverifiable plans and high variability in instance complexity. This paper proposes Converge, the first multi-agent framework featuring a constraint-guided, iterative verification paradigm built upon a novel tri-agent architecture—Constraint, Verify, and Select. It introduces an instance-complexity-aware reasoning algorithm selection mechanism that dynamically adapts inference strategies. By integrating formal constraint modeling with advanced reasoning enhancements—including Best-of-N, Tree-of-Thought, and REBASE—the framework establishes a closed-loop verification process and enables self-adaptive scheduling of reasoning algorithms. Converge achieves state-of-the-art performance across four challenging benchmarks: NATURAL PLAN (+8%), OlympiadBench (+4%), DocFinQA (+7%), and GPQA (+1%). Empirical results demonstrate substantial improvements in plan verifiability, robustness to input variation, and generalization across diverse complex reasoning tasks.

This work addresses the fragmentation in current AI agent research stemming from the absence of a systematic architectural framework and unified evaluation standards. To bridge this gap, the paper proposes a comprehensive taxonomy encompassing components, orchestration, and deployment, offering a structured analysis of single- and multi-agent architectures, coordination mechanisms, and application scenarios. It integrates core modules—including large language models, memory systems, world models, planners, tool routers, and critic components—and synthesizes key techniques such as chain-of-thought reasoning, self-reflection, hierarchical planning, and multimodal perception. Building on this foundation, the study consolidates evaluation methodologies—spanning task suites, human preference alignment, and success rates under constraints—elucidates the sources of evaluation complexity, advocates for reproducible benchmarking practices, and highlights critical open challenges in verification, memory management, interpretability, and robustness.

To address the challenge of balancing concurrent action conflicts and plan executability in multi-agent task planning, this paper proposes a two-stage LLM-PDDL collaborative framework. First, a large language model (LLM) performs commonsense-driven goal decomposition to generate mutually exclusive, parallelizable sub-goals. Second, a classical PDDL planner (e.g., FF or Fast Downward) independently synthesizes formally verifiable single-agent plans for each agent. This work is the first to integrate the LLM’s high-level goal abstraction capability with the formal correctness guarantees of classical planning. Empirical results demonstrate 100% action executability, significantly reduced planning time, and plan step counts that outperform single-agent baselines while approaching human expert performance—thereby unifying efficiency, feasibility, and coordination quality in multi-agent planning.

This work addresses reliability challenges in agent systems—such as hallucinated actions, infeasible plans, and fragile coordination—often stemming from the absence of explicit architectural constraints. To mitigate these issues, the authors propose DALIA, a declarative, model-agnostic agent architecture layer that enforces a clear separation between discovery, planning, and execution phases. DALIA leverages declarative capability specifications, a task discovery protocol, a federated agent registry, and deterministic task graphs to confine agent behavior within a verifiable action space. Notably, it introduces the first declarative architecture layer integrated with the Model-Controller-Planner (MCP) ecosystem, enabling reproducible and verifiable task-oriented workflows in heterogeneous environments. This approach significantly reduces reliance on speculative reasoning and unstructured coordination, thereby enhancing system reliability and cross-environment consistency.

Existing multi-agent collaborative systems are hindered by static workflows, sequential scheduling, and heterogeneous interfaces, leading to high complexity and poor scalability. This work proposes Agent-as-Tool, a unified paradigm that abstracts both agents and tools into a standardized, learnable action space, and introduces ParaManager—a lightweight coordinator enabling state-aware parallel subtask decomposition, delegation, and asynchronous execution. By unifying communication protocols and incorporating explicit state feedback, the framework facilitates efficient multi-agent collaboration. A two-stage training strategy—combining supervised fine-tuning with a recovery mechanism and reinforcement learning—optimizes task success rate, protocol compliance, response diversity, and reasoning efficiency. Experiments demonstrate that ParaManager achieves strong performance across multiple benchmarks and exhibits robust generalization to unseen agent pools.

In enterprise-scale complex tasks, autonomous agents face two key challenges: difficulty in coordinating multiple tools and weak handling of heterogeneous, multi-source data. Monolithic agent architectures yield unstable execution trajectories, while locally deployed open-weight models suffer from limited context windows, causing them to be overwhelmed by large tool outputs. To address these issues, we propose RP-ReAct—a novel multi-agent framework that decouples strategic planning (performed by a large reasoning model–driven Reasoner-Planner Agent) from tool execution (handled by a ReAct-style Proxy-Execution Agent). We further introduce a context preservation mechanism and external storage for managing voluminous tool outputs. Evaluated on the multi-domain ToolQA benchmark using six open-source reasoning models, RP-ReAct consistently outperforms existing methods, demonstrating superior generalization, robustness, execution stability, and adaptability across varying model scales.

This work addresses the limited planning generalization of large language model (LLM) agents in unseen scenarios by proposing a dynamic policy learning framework that integrates generalized planning with hierarchical task decomposition. The approach automatically extracts and reuses parameterized policy components from successful executions to construct a composable policy library. Central to the method are hierarchical component learning (HCL-GP), semantic-driven policy retrieval, and a dynamic reuse mechanism that enables cross-task knowledge transfer. Evaluated on the AppWorld benchmark, the proposed method achieves task success rates of 98.2% on standard tasks and 97.8% on challenging ones—representing a 15.8 percentage point improvement over static composition. Notably, it elevates the success rate of open-source LLM agents from near zero to 62.5%, substantially enhancing their task generalization capabilities.

This work addresses the challenges of low coordination efficiency, lack of auditability, and insufficient policy compliance in multi-agent systems for complex tasks by proposing a unified orchestration framework that integrates planning, policy execution, state management, and observability mechanisms. The framework introduces two novel protocols—the Model Context Protocol and Agent2Agent Protocol—to enable standardized, extensible communication among agents and between agents and external tools. By supporting policy governance and end-to-end traceability, the proposed architecture provides a practical, transparent, and accountable blueprint for deploying large-scale multi-agent systems within enterprise AI ecosystems.