The Stochastic Student Knowledge Graph (SSKG) addresses the tendency of LLMs to overestimate student mastery levels during simulations. By grounding simulations in a curriculum knowledge graph, it overcomes the performance bias seen in models like GPT-5.4-mini and Claude Haiku 4.5.
HOW THIS AFFECTS YOU
●
builderIf you are building synthetic tutoring data, use KG-grounding to avoid inflated performance metrics.
●
researcherThis exposes a systematic bias where LLMs default to high-capability responses regardless of assigned personas.