Probing Outcome-Level Resemblance and Mechanism-Level Alignment in LLM Risk Decisions: Evidence from the St. Petersburg Game
Researchers tested 28 LLMs on the St. Petersburg paradox, a classical economics puzzle where rational expected value diverges sharply from human behavior. The study reveals a critical gap in AI alignment: models that produce cautious-looking outputs may not actually replicate human decision-making logic. By systematically varying game parameters, prompt framing, and comparing base models to instruction-tuned variants, the work exposes whether LLM risk aversion stems from genuine mechanism alignment or surface-level mimicry. This matters for deployment in high-stakes domains where appearing aligned masks fundamentally different reasoning.
Modelwire context
ExplainerThe study's sharpest contribution is methodological: by varying game parameters and comparing base models against instruction-tuned variants, the researchers can isolate whether cautious outputs are a product of training-induced mimicry rather than anything resembling internalized risk reasoning. That distinction is what most behavioral benchmarks quietly skip.
This connects directly to the Bitcoin audit paper from June 1 ('Auditing Asset-Specific Preferences in Financial Large Language Models'), which found that LLM portfolio recommendations shift based on framing rather than fundamentals. Both studies are probing the same underlying problem: surface behavior in financial and risk contexts can look aligned while the internal driver is something else entirely. The eating disorder safety paper from the same date adds a clinical angle, showing that the gap between perceived and actual alignment is not limited to economic reasoning. Together, these three papers form a quiet but consistent thread in recent coverage: evaluation frameworks that measure outputs without interrogating mechanisms are systematically underestimating deployment risk.
Watch whether any of the 28 models tested here show consistent mechanism alignment across both the St. Petersburg framing and the asset-preference framing from the Bitcoin audit work. If the same instruction-tuned models that mimic human risk aversion here also show framing-driven asset bias there, that would suggest the mimicry problem is structural across financial reasoning tasks, not domain-specific.
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsSt. Petersburg Game · LLMs
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.