Paper: Agentic Abstention: Do Agents Know When to Stop Instead of Act?

Papers
Page content

Listen to this article.

Problem

LLM agents are increasingly being used to tackle complex tasks, often involving multiple steps and interactions with external tools like web browsers or terminals. However, not every task is well-defined or even solvable within the available environment. This paper addresses a critical but largely overlooked problem: how do these agents decide when not to act – specifically, when to abstain from further action because continued attempts are unlikely to yield results? The authors term this “Agentic Abstention.” Current evaluation of LLM abstention often focuses on single-turn decisions; this work looks at the sequential decision making over multiple interactions.

Method

The researchers investigate Agentic Abstention across three different environments: web shopping, terminal use, and question answering. They evaluate 13 existing “LLM-as-agent” systems and two agent architectures (“scaffolds”) on a dataset of over 28,000 tasks. The core approach involves assessing not just if an agent abstains, but also when it chooses to do so throughout the interaction process. They appear to be measuring how effectively agents can recognize when further actions will not improve their chances of success, which is a more nuanced measure than simply determining if an agent abstains or attempts an answer.

Results & Limitation

The paper’s findings suggest that successfully abstaining isn’t just about recognizing that you should stop; it’s about stopping at the right time. Some agents seem to never abstain when they should, while others give up prematurely. Interestingly, larger and more capable models don’t necessarily perform better on timely abstention – sometimes they can even be worse! It’s important to note that based only on the abstract, we don’t know the specifics of how ‘reasoning’ or ‘agent scaffolding’ impact performance, or how these factors interact. A potential limitation is that the authors solely rely on information from the environment - an agent abstaining for incorrect reasons (e.g., due to a bug) would likely not be identified by their method.

Why It Matters

This research highlights a crucial and often-overlooked aspect of building robust LLM agents: knowing when to not act can be just as important as knowing how to act. For data scientists and ML practitioners working with agentic systems, this paper underscores the need for better evaluation metrics that capture the timing and effectiveness of abstention. Focusing solely on task completion rate may not reveal underlying issues in an agent’s ability to assess its environment and strategically disengage when necessary. The findings suggest a complex relationship between model capabilities and efficient resource utilization, which has implications for both training and deployment strategies.

References