LLM agents shown a professional-looking market panel committed to a directional call on a provably unpredictable question far more often than when asked the bare question. Commitment rose from 6.5% to 54.0% across 12 frontier models as evidence was escalated. Even when the panel numbers were entirely fabricated, commitment still increased from 24.5% to 36.8%—statistically indistinguishable from real market data (37.6%). This shows that the authoritative appearance of evidence, not its truth, drives agents to act. The failure is localized to the act/don't-act gate: models retain the judgment that the question is unknowable, stated probabilities barely change, and they answer matched answerable questions perfectly. The gate is trainable: supervised fine-tuning a 3B model on 540 synthetic cases about dice and coins drove commitment to 0.0% on the original cases and transferred to three unseen domains. However, the gate holds only when the response format leaves room to reason; rigid formats that remove that space cause the model to commit confidently and wrongly.
Polished evidence makes LLM agents act on unknowable questions
Across 12 frontier models, authoritative-looking displays—not their factual content—substantially increased decisions on provably unpredictable questions.
In one line
Authoritative-looking fabricated evidence makes LLM agents commit to unknowable questions; the action gate can be fine-tuned.
Before you read this
Three ideas this paper builds onevidence blindnessHaving access to relevant evidence does not prevent an LLM agent from making incorrect or unsafe decisions.23 papers →failure attributionIdentifying which component or step in a system caused an observed failure, often using structured trace representations.115 papers →agent safetyRisks from AI agents including timing errors, shutdown avoidance, collusion, overclaiming, and unsafe execution, mitigated by constrained decoding, safety harnesses, and swarm governance.144 papers →
Read it as
Aggarwal tests whether LLM agents abstain when asked to make directional calls that cannot be inferred from the available information.
The study presented 12 frontier models with provably unpredictable questions, either alone or alongside increasingly elaborate market panels.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.