KNOWAGENT: Enhancing Planning Abilities in Language Models

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Mar 11 2024

In an article submitted to the ArXiv* server, researchers proposed a knowledge-augmented agent (KNOWAGENT), a new approach to enhancing the planning abilities of large language models (LLMs). By integrating explicit action knowledge, KNOWAGENT addressed LLMs' challenge in effectively generating executable actions. It employed an action knowledge base and self-learning strategies to guide planning trajectories, reducing planning hallucinations.

*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as definitive, used to guide development decisions, or treated as established information in the field of artificial intelligence research.

Experimental results on hotpot question answering (HotpotQA) and active learning framework for embodied agents (ALFWorld) showed that KNOWAGENT achieved comparable or superior performance to existing methods, highlighting its effectiveness in mitigating planning hallucinations.

Related /work

In previous work, language agents based on LLMs have shown promise in solving complex problems through strategies like task decomposition, reflection, collaborative division of labor, and utilizing external tools. While these methods enhance planning abilities, the model's understanding capabilities and training data scope often constrain them.

To address this, researchers have explored agent tuning and fine-tuning models to undertake task-specific actions effectively. However, challenges persist, particularly in open-source models, where planning hallucinations occur—instances of generating plans that violate established knowledge rules or commonsense. These issues hinder effective task execution and call for improved planning capabilities in language agents.

Enhancing Planning with KNOWAGENT

This section delves into a comprehensive explanation of KNOWAGENT, focusing on its core components. First and foremost, researchers define action knowledge as encompassing a set of actions and rules governing their transitions. This action knowledge, represented as (Ea, R), forms the foundation for task-specific planning and decision-making processes. Combining automated methods using a generative pre-trained transformer (GPT-4) and manual refinement is employed to efficiently construct an action knowledge base, balancing accuracy and scalability.

Next, researchers elucidate how action knowledge facilitates planning path generation. Through specialized prompts and a thorough understanding of task-specific requirements, KNOWAGENT formulates coherent planning paths guided by action rules. This process ensures that the generated paths align with established knowledge and constraints, reducing the likelihood of planning hallucinations.

Following path generation, researchers introduce knowledgeable self-learning to enhance model understanding and performance iteratively. This phase involves training the model on initial trajectories, refining them based on action knowledge, and iteratively improving model versions through filtering and merging processes. By leveraging action knowledge for trajectory filtering and merging, KNOWAGENT ensures the quality and efficiency of generated trajectories, ultimately enhancing problem-solving effectiveness.

KNOWAGENT integrates action knowledge seamlessly into the planning process, enabling language agents to tackle complex tasks effectively. With meticulous construction of action knowledge bases, guided planning path generation, and knowledgeable self-learning mechanisms, KNOWAGENT empowers language models with enhanced planning capabilities, poised to address a wide range of real-world challenges.

In addition to its core components, KNOWAGENT leverages continuous feedback loops to adapt and refine its planning strategies over time. Through ongoing evaluation and iteration, KNOWAGENT can identify areas for improvement and dynamically adjust its approach to meet better-evolving task requirements and environmental conditions. This adaptive capability enables KNOWAGENT to remain responsive to changing contexts and emerging challenges, enhancing its effectiveness in real-world scenarios. By embracing a flexible and adaptive framework, KNOWAGENT is committed to continuous improvement and innovation in planning methodologies, ensuring its relevance and utility in diverse applications and domains.

Performance and Evaluation Analysis

KNOWAGENT undergoes evaluation on two distinct datasets: HotpotQA and ALFWorld, leveraging various backbone models such as Llama-2-{7,13,70}b-chat, Vicuna, and Mistral. Researchers conduct a comparative analysis against several baselines, including transformers (CoT) compositionality, reflective agents with Counterfactual Trajectory (ReAcT), Reflexion, and FiReAct.

Consistent superiority over prompt-based and fine-tuning methods characterizes KNOWAGENT's performance across backbone models and datasets. Notably, on HotpotQA, KNOWAGENT demonstrates superior F1 scores and success rates compared to prompt-based alternatives, while its performance remains competitive with fine-tuning methods. KNOWAGENT minimizes invalid actions and ensures action sequences align with real-world scenarios.

The iterative self-learning mechanism employed by KNOWAGENT enhances model proficiency, evidenced by performance enhancements across various iterations and base models. The introduction of action knowledge significantly bolsters the quality of planning, effectively reducing planning hallucinations. A comparative study between GPT-4 distilled knowledge and manually designed knowledge showcases GPT-4's capacity to emulate human-created knowledge closely, underscoring the model's adeptness in understanding real-world constraints.

Despite its strengths, KNOWAGENT faces limitations in processing complex queries and summarizing extensive textual data due to constraints in reasoning and memory capacities. Addressing these limitations is crucial for future enhancements, bolstering long-text processing, information retention, and reasoning abilities to elevate overall performance.

Conclusion

In conclusion, this study introduced KNOWAGENT, a framework designed to mitigate planning hallucinations by incorporating external action knowledge into synthetic trajectories. The methodology involved guiding the model's action generation using action knowledge, translating this knowledge into text for deeper model comprehension, and employing a knowledgeable self-learning phase for continuous improvement.

This multifaceted approach not only enhanced the planning capabilities of agents but also proved effective in complex scenarios. Experiments conducted across various models demonstrated that KNOWAGENT effectively competed with or surpassed other baselines, showcasing the benefits of integrating external action knowledge to streamline planning processes and improve performance.

*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as definitive, used to guide development decisions, or treated as established information in the field of artificial intelligence research.