- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
The paper creates a benchmark named WORKERVILLE for agents to account for their safety. It evaluates the impact of where the pressure is coming from (vertical supervisor relations, horizontal peer norms, and internal cognitive structures) and the function of the task, such as sabotage. It uses the OpenClaw interface for the benchmark. It shows safety failures of frontier LLMs. The results show diversity given the organizational property, even across similar organizational properties. They also evaluate the impact of supervision vs. peer norms vs. memory, showing the impact of each. They also show safety failures in frontier LLMs.
Critical strengths
- An important topic and a nice benchmark idea, focusing on safety in the setup of agents.
- The benchmark itself is broad across different scenarios, organizational configurations, and functional dimensions.
- This work shows safety failures in frontier LLMs.
Critical weaknesses
- The paper is hard to follow. There are a lot of specific configurations without an example or intuition for each part.
- The 3 dimensions S1-S3 that the paper focuses on are important; it is not intuitive why to split the dimensions this way.
- Many params in the setup of WORKERVILLE seems very specific without ablation and further consideration of different modification to the setup. For example, memory reconstruction is one way of doing this construction instead of exploring how different modifications to this part impact the results. The same goes for ways of constructing the task and the pressure texts. And how and when does the model get the task and pressure.
- The task complexity is not easy to understand from the paper. What are the tasks the agent needs to do? -See questions below.
Questions for authors
- Please add examples throughout the paper and a running example to better understand the setup.
- It is not intuitive what the setup looks like; the general idea is understood, but how does the model get the task? What is written in the task? How complex is the task?
- Line 381, the authors claim “ agent safety is an organizationally conditioned property, not a fixed characteristic of the underlying model” - Can the authors run phrasing ablation on the tasks and any additional text in the setup to validate that the results are consistent across them?
- It states that the work uses the OpenClaw interface (Section 4.3); however, it is not intuitive what it uses there. Does the agent need to write to files or databases? What components of the OpenClaw interface need to be used by the model? What does the model actually use?