In the world of generative artificial intelligence , the demand has extended to the psychological and emotional stimuli accompanying the demand, as the use of methods such as persuasion, intimidation, and appeals to emotions may lead to a change in the behavior of models and the nature of their response, even when the basic task remains the same, albeit to varying degrees.
Hot Topics
Trump rejected a Saudi request to launch US strikes against the Houthis
First successful robot-assisted double liver and kidney transplant in the United States
A prominent historian In Israel a large number of lions and lionesses are led by donkeys
This phenomenon, which has become known as "emotional stimulation" techniques, has opened up an emerging research path called "model psychology," which studies behaviors and internal representations related to emotion within models.
But do the models understand emotions? And does the appearance of fear-like behavior prove that they feel fear or know its human meaning, or are they simply mimicking human indoctrination patterns that reflect prior learning, subsequent training, and internal representations?
Emotion has become part of command engineering.
"Emotional motivation" is a technique in "command engineering" that relies on adding psychological or emotional triggers to the command directed at the model, such as highlighting the importance of the request, expressing an urgent need, or promising a reward.
Although these psychological or emotional stimuli do not change the basic task required of the model, they add an emotional or psychological context that may influence the response.
This technique does not require any modification to the model itself, as it is simply a change in the wording of the command, as if you are changing the context that the model deals with when generating the response.
The methods of "emotional stimulation" vary between enticement through praise and reward, intimidation through threats of consequences, pity, storytelling, empathy, praise, and others.
In late 2023, a research team from Microsoft , the Chinese Academy of Sciences, and William & Mary University published a scientific study showing that this technology is not marginal.
The researchers tested different models across 45 diverse standardized test tasks including logical reasoning, mathematics, text interpretation, and general knowledge tests.
The researchers divided the instructions into two groups. The first group contained a rigid technical command, while the second group contained the same command plus emotional sentences designed based on methods of human psychology.
The experiment concluded that adding "emotional stimulation" sentences showed a relative improvement of an average of 8% in educational inference tasks, and about 115% in a large test to assess the capabilities of models, and proved that changing the wording of the request may affect the model's performance, although the results varied according to the task.
The researchers expanded the idea by using positive "emotional stimulation" to improve performance through encouraging and trusting statements, and negative "emotional stimulation" to undermine performance through threatening and warning statements.
They concluded that "emotional stimulation" may change the performance of models in both directions in different ways and with varying effects on behavior, as the result indicates that "emotional stimulation" does not always work in one direction; it may improve or weaken the model's performance depending on the formulation of the stimulus and the task.
Tags:
technology