Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

LTLCrit: A Temporal Logic-based LLM Critic for Safe and Efficient Embodied Agents

Created by
  • Haebom

Author

Anand Gokhale, Vaibhav Srivastava, Francesco Bullo

Outline

In this paper, we propose a modular actor-critic architecture consisting of an LLM actor and a Linear Temporal Logic (LTLCrit)-based LLM critic to overcome the limitations of large-scale language models (LLMs) that suffer from low safety and efficiency due to error accumulation in long-term planning tasks. The LLM actor selects high-level actions through natural language observations, and LTLCrit analyzes the entire path to propose novel LTL constraints that prevent unsafe or inefficient future actions. The architecture supports both fixed manually specified safety constraints and adaptive learning soft constraints that improve long-term efficiency, and is model-independent. We demonstrate the safe and generalizable decision-making capability of the LLM with logic-based mutual supervision by achieving 100% completion rate and improved efficiency compared to the existing LLM planner on the Minecraft diamond mining benchmark.

Takeaways, Limitations

•
Takeaways:
◦
Presenting a new modular architecture to enhance LLM's long-term planning capabilities.
◦
A new approach to ensuring the safety and efficiency of LLM by leveraging LTL.
◦
Model-independent architecture that can be applied to various LLM-based planners.
◦
Performance verified with 100% completion rate and improved efficiency in Minecraft Benchmark.
◦
Presenting the possibility of safe and generalizable decision-making in LLM through logic-based mutual supervision.
•
Limitations:
◦
Additional experiments and verification of the generalization performance of the proposed architecture are needed.
◦
Further research is needed on applicability and scalability to various environments and tasks.
◦
Need to improve the computational complexity and efficiency of LTLCrit.
◦
Consideration should be given to the limitations and complexities of expressing constraints via LTL.
👍
Made with Slashpage