Agreement among large language models does not guarantee correct student diagnosis
Agreement Is Not Validity: Cross-Model LLM Consensus in Diagnosing Student Failure Modes in K-12 Math Tutoring Dialogue
Computation and Language
Summary
When students talk with math tutors, their conversations can show where they struggle. The authors studied how well large language models (LLMs) can spot five common student mistakes by analyzing these talks. They found that while different LLMs tend to agree with each other a lot, their agreement with human judgments is only moderate. This means that just because many LLMs agree doesn't mean their understanding is actually accurate or valid.
What this means in practice
- •For learning analytics teams: Use caution when relying on multiple LLMs to label student dialogue, as agreement does not guarantee correct identification of student difficulties.
- •For educational software developers: Incorporate independent validation steps instead of trusting LLM consensus alone for diagnosing student problem-solving errors in tutoring systems.
Authors
Clayton Cohn, Joyce Fonteles, Kirk Vanacore, Gianni Mazza, Candida Crawford, Tom Hooper, Gautam Biswas, Rene Kizilcec
Abstract
In K-12 mathematics tutoring, student-tutor dialogue provides rich evidence of learners' problem-solving processes and sources of difficulty. Learning analytics research increasingly relies on large language models (LLMs) to extract such information from dialogue for a variety of downstream tasks, including knowledge tracing, behavioral modeling, and diagnosis of student reasoning errors. However, the validity of these model-generated interpretations remains insufficiently understood. In this exploratory study, we examine the validity of LLM classifications of five student failure modes in mathematics tutoring dialogue using an operational diagnostic codebook: uncertainty, misattribution, operator selection, conceptual gap, and procedural slip. Across models, human-LLM agreement was moderate (kappa = .524-.597), while cross-model agreement was substantially higher (kappa = .755-.781; alpha = .769). These findings show that cross-model agreement can create a misleading appearance of correctness, challenging the assumption that consensus among LLMs constitutes evidence of valid learner interpretation. For learning analytics, the implication is clear: scalable labeling is useful only if the inferred constructs are valid, and model consensus cannot substitute for independent evidence of that validity.