🔒Открыть пост
и посмотреть медиа
и посмотреть медиа
🤖 AI Models Are So Smart They Can't Pretend to Be Bad Students
Researchers at Drexel University tried to make language models act like students with different levels of algebra knowledge. But simply telling them to "answer like a struggling student" wasn't enough to make them play the role convincingly.
✏️ Gemini 3.1 Flash Lite, Claude Haiku 4.5, and GPT-5.4-mini were given 379 algebra problems and assigned five different student personas. Yet they failed at the task: whether the model was an "A student" or a "poor student," its accuracy stayed between 96.8% and 100%.
The issue is that the model already knows the correct answer and struggles to suppress its own capabilities. So the authors split the task into two stages: first, a separate algorithm models what a student knows and where they are likely to make mistakes; then, an LLM explains the student's answer. This produced more plausible performance differences: the "near-expert" student scored 85.2% accuracy, the average student 57.8%, and the struggling student 44.1%.
👨 The authors acknowledge that they have not yet compared their virtual "students" with real ones. So for now, this is more a method for teaching AI to make believable mistakes than an accurate model of how humans learn.
@hiaimediaen
