LLMs show surface-level risk alignment, study finds

A new research paper explores whether Large Language Models (LLMs) truly align with human decision-making mechanisms when faced with risk, using the St. Petersburg game as a testbed. While many LLMs produce human-like finite bids in the original game, this outcome-level resemblance often hides differing underlying reasoning processes. Controlled variants of the game reveal that LLMs frequently shift to conditionally rational behavior rather than maintaining human-consistent mechanisms, even after instruction tuning. AI

IMPACT Highlights the need for deeper evaluation of LLM decision-making beyond surface-level outcomes to ensure true alignment.

RANK_REASON Academic paper analyzing LLM behavior on a specific task. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• LLMs
• St. Petersburg game

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →