Evaluating the Goal-Directedness of Large Language Models
Keywords: LLMs, agents, goal-directedness, safety
TL;DR: Do LLMs fully use their capabilities to achieve goals?
Abstract: LLM-based agents may transform AI and society in the near future. Along with opportunities for automation and increased productivity come novel safety and ethics concerns. This means both researchers and regulators need good ways to keep track of progress and properties of LLM-based agents. A key feature of agentic behaviour is goal-directedness, which has so far received limited attention in the context of AI agents. In this work we define the concept of goal-directedness for LLM agents, and develop a framework for evaluating it empirically on tasks involving information gathering, information processing, and execution. Results on state-of-the-art LLM agents indicate a lack of goal-directedness, meaning models often fail to fully deploy capabilities that they evidently have. This raises the question of how we can elicit the full capabilities of LLM-based agents, as well as what policies should be in place for future more goal-directed systems.
Primary Area: alignment, fairness, safety, privacy, and societal considerations
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 10743
Loading