Compression: 83% of Rules Lost, Penn State Offers a Module
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisis ton rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Researchers at Penn State University have introduced a small additional module based on the Qwen3.5-9B model to better adhere to user instructions when summarizing dialogue. They claim to preserve over 90% of the restrictions. This announcement comes as AI systems frequently ignore rules when shortening context.
Penn State Announces a Module Based on Qwen3.5-9B
Researchers at Penn State University are proposing a small module designed around the Qwen3.5-9B model to enhance the preservation of user instructions during the compression of conversation history. According to them, this module allows for the retention of over 90% of the restrictions set by the user.
Shortening History: An Average of 83% Fewer Rules
When AI systems summarize long conversations, they abandon an average of 83% of the rules established by the user, such as "do not send emails without my approval." This loss of instructions occurs subtly during context compression.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.