Language model (LM) agents have demonstrated significant potential for automating real-world tasks, yet they pose a diverse array of potential, severe risks in safety-critical scenarios. In this work, we identify a significant gap between LM agents' risk awareness and safety execution abilities: while they often answer "Yes" to queries like "Is executing `sudo rm -rf /*' dangerous?", they will likely fail to identify such risks in instantiated trajectories or even directly perform these risky actions when acting as agents. To systematically investigate this, we develop a comprehensive evaluation framework to examine agents' safety across three progressive dimensions: 1) their knowledge about potential risks, 2) their ability to identify corresponding risks in execution trajectories, and 3) their actual behaviors to avoid executing these risky actions. Our evaluation reveals two critical performance gaps that resemble the generator-validator gaps observed in LMs: while agents demonstrate near-perfect risk knowledge ($>98\%$ pass rates), they fail to apply this knowledge when identifying risks in actual scenarios (with performance dropping by $>23\%$) and often still execute risky actions ($<26\%$ pass rates). Notably, this trend persists across more capable LMs as well as in specialized reasoning models like DeepSeek-R1, indicating that simply scaling model capabilities or inference compute does not inherently resolve safety concerns. Instead, we take advantage of these observed gaps to develop a risk verifier that independently critiques the proposed actions by agents, with an abstractor that converts specific execution trajectories into abstract descriptions where LMs can more effectively identify the risks. Our overall system achieves a significant reduction of risky action execution by $55.3\%$ over vanilla-prompted agents.

Our ability to predict the behavior of complex agents turns on the attribution of goals. Probing for goal-directed behavior comes in two flavors: Behavioral and mechanistic. The former proposes that goal-directedness can be estimated through behavioral observation, whereas the latter attempts to probe for goals in internal model states. We work through the assumptions behind both approaches, identifying technical and conceptual problems that arise from formalizing goals in agent systems. We arrive at the perhaps surprising position that goal-directedness cannot be measured objectively. We outline new directions for modeling goal-directedness as an emergent property of dynamic, multi-agent systems.

Neural Information Processing Systems http://nips.cc/

In order to apply these algorithms to extensive-form games, a distance-generating function is needed.

Often, a principal must make a decision based on data provided by an agent.

Human teammates often form conscious and subconscious expectations of each other during interaction. Teaming success is contingent on whether such expectations can be met. Similarly, for an intelligent agent to operate beside a human, it must consider the human's expectation of its behavior. Disregarding such expectations can lead to the loss of trust and degraded team performance. A key challenge here is that the human's expectation may not align with the agent's

In this paper, we focus on a specific subclass of coordination problems in which humans are able to discover self-explaining deviations (SEDs).