Summary KnowAgent enhances Large Language Model (LLM) agents' planning capabilities by augmenting them with external action knowledge. It targets researchers and developers building sophisticated LLM agents, enabling more robust decision-making and improved execution in complex tasks through a knowledge-augmented approach.

How It Works The core methodology involves creating a task-specific action knowledge base, converting it to text for LLM comprehension, and then generating action trajectories. A subsequent self-learning phase iteratively refines the agent's knowledge and application by processing these trajectories, thereby strengthening its planning abilities and adaptability.

Quick Start & Requirements Installation requires cloning the repository and running pip install -r requirements.txt. Key dependencies include ALFWorld and FastChat, with datasets (HotpotQA, ALFWorld) needing placement in specified directories. Planning scripts (run_alfworld.sh, run_hotpotqa.sh) in Path_Generation support training/testing with configurable LLMs (e.g., llama-2-13b). Self-learning involves data reformatting and training scripts in Self-Learning, utilizing deepspeed for distributed training.

Highlighted Details

• Paper accepted to NAACL 2025 Findings.
• Received Best Paper Award at KnowledgeNLP workshop, ACL 2024.
• Released synthesized trajectories from Llama-{7,13,70}b-chat models.

Maintenance & Community The project commits to long-term maintenance, including bug fixes and issue resolution. It integrates with FastChat (for training and LangChain integration) and BOLAA (for inference).

Licensing & Compatibility KnowAgent is licensed under Apache 2.0, which is permissive for commercial use and closed-source linking.

Limitations & Caveats The README does not explicitly detail limitations, unsupported platforms, or known bugs. The project appears focused on specific benchmark datasets (HotpotQA, ALFWorld) and LLM architectures.