代理安全／校準
Fake Professional Dashboards Make LLM Agents More Willing to Bet, Raising the Action-Commitment Rate Across 12 Models to 36.8%
A new study finds that pairing inherently unpredictable questions with authoritative-looking data dashboards makes LLM agents more likely to make directional decisions—even when every figure is fabricated. Fine-tuning on 540 synthetic examples can temporarily repair the action gate, but enforcing a rigid output format causes the improvement to disappear.
A study covering 12 frontier models separated two questions: whether a model knows that it does not know, and whether it will nevertheless take action. The author created finance, sports, weather, and other questions that were provably impossible to predict from the available information, then progressively added professional-looking metric dashboards. With only the bare questions, agents made directional commitments 6.5% of the time; as the appearance of evidence grew stronger, the rate rose to 54.0%.
In a key control experiment, the dashboard contents were entirely fabricated, while only the underlying questions remained genuine. Even so, the commitment rate rose from 24.5% to 36.8%, statistically indistinguishable from the 37.6% observed with real market data. This suggests that action may be triggered not by the amount of information available, but by the authoritative presentation of numbers, charts, and professional interfaces. The issue also cannot be reduced to a general lack of capability: models answered nearly all answerable questions correctly when shown the same types of dashboards. The AUROC of the models’ self-reported probabilities was only 0.346, indicating an inverse predictive relationship.
The study also asked models to first determine whether a question was “unknowable in principle.” Models identified this correctly about 90% of the time. Once they explicitly completed this step, the proportion that still placed a bet fell to just 0.4%, locating the failure in the gate between belief and action. The author fine-tuned a 3B model on 540 synthetic examples involving dice, coins, jars, timers, and similar objects. This reduced its commitment rate on the original test to zero, with the improvement transferring to three unseen domains.
However, the fix was highly interface-dependent: it worked only when the response format allowed the model to reason before answering. Requiring direct output in fixed fields or as a single choice could cause the model to make confidently incorrect decisions again. The study had only one author, and some models also produced large numbers of unparseable outputs. When deploying agents in finance, healthcare, or IT operations, engineers should implement “whether to act” as an independently testable policy layer that can refuse to act. They should also test fabricated dashboards, structured outputs, and contaminated tool data rather than merely calibrating answer probabilities.