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

From an LLM Swarm to a PDDL-empowered Hive: Planning Self-executed Instructions in a Multi-modal Jungle

ICLR 2025 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 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 Hive, a knowledge-aware planner that builds a Capability Knowledge Graph from HuggingFace/Papers With Code and uses LLM-driven PDDL planning to select and chain multi-modal models, evaluated on a new MuSE benchmark against HuggingGPT and ControlLLM.

Problem

Selecting the most appropriate models from a huge, heterogeneous model hub and planning chains of atomic actions across them for multi-modal user queries while respecting user constraints (size, licence) and remaining explainable.

Approach

Combine a machine-readable Capability Knowledge Graph (model cards, benchmark performance, execution snippets) with LLM-based PDDL planning so plans are formally generated, constraint-checked, and explainable before specialised models are executed.

Key results

• Hive reports about 30% higher task-selection accuracy than competing systems on MuSE, and respects user constraints in 100% of cases.
• GPT-4o answered correctly (O) 43% of the time on the filtered MuSE subset, with 44% for task selection and justification.
• o1-preview achieved 24% correct outputs and 26% correct justifications on its text-only subset; DeepSeek-v3 reached 45% (FS) and 47% (FoT) but only 25% output quality.

Measured effects reported in the paper

Surprising findings

• Monolithic LLMs such as o1-preview and GPT-4o performed poorly when planning and executing MuSE queries themselves, underscoring the difficulty of real-world multi-modal chains.
• HuggingGPT and ControlLLM frequently ignored licence and benchmark-based constraints or failed at basic task selection in the appendix examples.

Limitations

• Evaluation relies on a self-introduced benchmark (MuSE) with limited details in the text, making comparisons hard to verify.
• ControlLLM was restricted to one-to-one mappings, and the baseline models were assigned one model per task, which may limit fairness.
• Monolithic-LLM comparison covers only filtered subsets of queries due to modality restrictions.

Experimental setup

Datasets: HuggingFace model cards Papers With Code benchmark results

Benchmarks: MuSE

Metrics: task selection accuracy final output accuracy (O) faithfulness of task selection and justification (FS/FoT) user constraint adherence

Baselines: HuggingGPT ControlLLM

Builds on: PDDL HuggingGPT ControlLLM Chain-of-Thought Toolformer

Closest papers by meaning

• AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration (ICML 2026 · similarity 0.89)
• AgentSquare: Automatic LLM Agent Search in Modular Design Space (ICLR 2025 · similarity 0.88)
• MAS$^2$: Self-Generative, Self-Configuring, Self-Rectifying Multi-Agent Systems (ICLR 2026 · similarity 0.88)
• From Assumptions to Actions: Turning LLM Reasoning into Uncertainty-Aware Planning for Embodied Agents (ICLR 2026 · similarity 0.88)
• AgentSwift: Efficient LLM Agent Design via Value-Guided Hierarchical Search (AAAI 2026 · similarity 0.88)
• AutoTool: Dynamic Tool Selection and Integration for Agentic Reasoning (ICML 2026 · similarity 0.88)