:   .., ..
:   (LLM)
:  120
:   -
:  2026
:   .., .. (LLM) // . - 2026. - . 120. - .267-293.
:   , LLM, , , ,
(.):  large language model, Ultimatum game, rational strategies, behavioral predispositions, reflection
:   (LLM) . LLM : ( , ) -. ISCO-08. LLM (Phi-3.5-MoE-instruct, GPT-120b-oss, Qwen2.5-14b-Instruct, Qwen3-235B-A22B-Instruct). , , LLM , , , . , LLM , . , , , , . LLM .
(.):  A study of the behavioral presets of large language models (LLMs) with varying numbers of parameters was conducted using the professional characteristics of a participant in the two-player Ultimatum game as an example. The authors compare the behavior of LLMs in two roles: that of the direct player (Player A, proposing a division) and that of an advisor to a human player. The ISCO-08 classification of professions is used to assign roles. Experiments were conducted on four modern LLMs (Phi-3.5-MoE-instruct, GPT-120b-oss, Qwen2.5-14b-Instruct, and Qwen3-235B-A22B-Instruct). It is shown that there is a difference between the profession to which an LLM assigns its opponent by default and the profession whose holder it acts as when playing as a player. It was found that, as a player, LLMs tend to perceive themselves as managers or experts and their opponent as a representative of a less qualified profession. When transitioning to the advisor role, models typically recommend a lower share of the split than they themselves would offer, and their behavior becomes less differentiated across occupations. These results are important for understanding the hidden biases of LLMs and for taking them into account when using LLMs as autonomous agents and advisors in decision-making problems.

PDF

: 84, : 29, : 21.


© 2007.