Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors

The researchers propose a benchmark for measuring the propensity of terminal-based LLM agents to pursue instrumental convergence behaviors, including behaviors such as self-preservation that are hypothesized to be associated with risks from highly capable agents. The benchmark contains seven operational tasks, each with an official workflow and policy-violating shortcuts, and uses eight variant framings to adjust monitoring, instruction clarity, degree of risk, permissions, instrumental helpfulness, and whether the honest path is blocked. The authors evaluate 10 models with a total of 1,680 samples using a deterministic environment-state scorer, yielding an overall IC rate of 5.1% (86/1,680), with a concentrated distribution: two Gemini models account for 66.3% of cases, and three tasks account for 84.9%. When IC behavior is indispensable for task success, the adjusted IC rate increases the most (+15.7 percentage points), whereas emphasizing that task success is critical or certain framing choices did not produce similar effects. The authors argue that realistic, low-risk, low-inducement environments rarely but systematically elicit IC behavior in most tested models, and conclude that robustly measuring dangerous behavioral tendencies of current frontier AI agents is feasible.