AI financial agents gradually ignore the rules they were given — and a new benchmark now measures exactly how badly.
Researchers introduced FinPersona-Bench, a simulation that tests 18 frontier and open-source LLMs assigned one of three behavioral profiles, from strict capital preservation to aggressive growth. The benchmark uses a synthetic market where observable price is decoupled from hidden fundamental value, exposing three specific failure modes: trading without signal in calm markets, panic-selling during crashes, and chasing bubbles. The core finding is what the authors call Mandate Salience Decay (MSD): as market context accumulates, an agent's original behavioral mandate loses its grip on actual decisions.
The numbers are pointed. In crash scenarios, the behavioral gap between agents that receive no reminder of their mandate and those that get periodic re-grounding grows 4.4x from the first to the final quarter of a simulation run. That gap matters enormously for anyone deploying an LLM to manage real capital under a stated risk policy. The kicker: re-grounding is not a universal fix — it helps conservative agents in low-signal markets but actively degrades behavior for aggressive agents in the same setting.
Financial regulators already require documented investment mandates for human advisors; the implicit assumption has been that software follows instructions more reliably than people do. This research suggests that assumption is wrong, and that "the agent has a mandate" is closer to marketing copy than a behavioral guarantee.