#Nextquick Une simple requête à ChatGPT, c’est déjà des centaines de milliards de calculs

The models’ ROUGE-L and BERTScore scores consistently maintain stability, irrespective of the politeness level of the prompts, which infers that the models can correctly summarize the article content in the summarization tasks.

However, the models manifest substantial variation in length correlated to the politeness level.

A progressive reduction in the generation length is evident as the politeness level descends from high to lower scales. Conversely, a surge is noted in the length of the outputs of GPT-3.5 and Llama2-70B under the exceedingly impolite prompts. The propensity exhibited by the models to generate more extended output in polite contexts. Polite and formal language is predominantly used in scenarios demanding descriptive instructions or instructional literature, often associated with longer text. Conversely, antagonistic and fervent discourse involves impolite language, which is also associated with extended lengths. These facets reflect the nuances of human social behavior, mirrored in the training data, and then influence the tendencies demonstrated by LLMs.

However, GPT-4 did not echo this trend of increased output length in the presence of highly impolite prompts. It is conjectured that GPT-4, being a superior model, might prioritize the task itself and effectively control the tendency to “argue” at a low politeness level.