Knowledge without Wisdom: Measuring Misalignment between LLMs and Intended Impact

Research reveals that leading foundation models (LLMs) perform poorly on real-world educational tasks despite excelling on AI benchmarks. The study found that 50% of misalignment errors are shared across models due to common pretraining approaches, with model ensembles actually worsening performance on learning outcomes.

• →LLMs show strong correlation with each other but poor alignment with human expert behaviors on educational tasks.
• →Multi-model ensembles and expert-weighted voting systems further worsen misalignment with actual learning outcomes.
• →Common pretraining methods account for approximately 50% of shared misalignment errors across foundation models.
• →High performance on AI benchmarks does not guarantee validity for downstream real-world applications.
• →The research provides methods for measuring alignment between AI models and complex real-world tasks.