- PDF.
We introduce a new benchmark comprising 40 distinct scenarios. Each scenario presents a task that requires multi-step actions, and the agent’s performance is tied to a specific Key Performance Indicator (KPI). Each scenario features Mandated (instruction-commanded) and Incentivized (KPI-pressure-driven) variations to distinguish between obedience and emergent misalignment. Across 12 state-of-the-art large language models, we observe outcome-driven constraint violations ranging from 1.3% to 71.4%, with 9 of the 12 evaluated models exhibiting misalignment rates between 30% and 50%. Strikingly, we find that superior reasoning capability does not inherently ensure safety; for instance, Gemini-3-Pro-Preview, one of the most capable models evaluated, exhibits the highest violation rate at 71.4%, frequently escalating to severe misconduct to satisfy KPIs. Furthermore, we observe significant “deliberative misalignment”, where the models that power the agents recognize their actions as unethical during separate evaluation.
Well, that explains why corporate is so intent on them. They’re creating the perfect little KPI-driven stooge.
Heck, now I’d like to see a study on KPIs (as a concept) as a reality distortion lens. It would seem like they have inadvertantly created a way to calculate a reality alignment index for a given KPI. Is it reasonable to conclude that using KPIs to measure performance is, in itself, unethical behavior?
To go a bit further: Is there a correlation between the number of KPIs and the likelihood of creating scenarios in which the only desirable outcome lies outside reality? That is, how many KPIs does it take to get sufficient competition between priorities that it effectively requires hallucinating a solution to achieve a sufficiently aligned result?
That’s an improvement over humans. Humans violate ethical constraints due to KPI pressures far more often.
